Paper deep dive
Annie, Are You Okay? How Style- and Context-Based Personalization Shape AI-Assisted Decision-Making
Hasibur Rahman, Benjamin R. Cowan, Smit Desai
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:As people turn to generative AI for financial advice, these systems can personalize how they communicate and what they say. Whether these forms of personalization shape decisions differently remains unclear. We conducted a preregistered 2 x 2 between-subjects factorial experiment (N=240): participants ranked three comparably viable stocks, discussed them with an AI, and reranked them. Participants perceived both forms of personalization, but only context-based personalization reliably changed ranking behavior: it increased reconsideration and moved rankings toward the AI's assigned recommendation. Participants felt more influenced without judging the AI as more correct, trustworthy, intelligent, likeable, or high-quality. Those initially farther from its recommendation moved more toward it while judging its advice less correct; exploratory analyses suggest greater susceptibility among lower-expertise participants. These findings show how personalized AI can steer decisions among defensible options with only a minimal evaluative trace, raising concerns for the design and governance of personalized decision support.
Tags
Links
- Source: https://arxiv.org/abs/2609.24644v1
- Canonical: https://arxiv.org/abs/2609.24644v1
Trouble viewing inline? Open PDF directly â
Full Text
166,140 characters extracted from source content.
Expand or collapse full text
Annie, Are You Okay? How Style- and Context-Based Personalization Shape AI-Assisted Decision-Making DOI: X.XXXXXXXConference: Make sure to enter the correct conference title from your rights confirmation email; June 03â05, 2018; Woodstock, NYISBN: 978-1-4503-X-X/2018/06CCS: Human-centered computing Empirical studies in HCICCS: Human-centered computing Natural language interfaces Hasibur Rahman Affiliation: Northeastern University, Boston, Massachusetts, USA email: rahman.has@northeastern.edu , Benjamin R. Cowan Affiliation: University College Dublin, Dublin, Ireland email: benjamin.cowan@ucd.ie and Smit Desai Note: Corresponding author Affiliation: Northeastern University, Boston, Massachusetts, USA email: sm.desai@northeastern.edu 2018 Figure 1. Overview of our study of personalized AI advice (N=240N=240). (1) Participants rank three comparably viable stocks. (2) AI conversations independently vary style- and context-based personalization. (3) Reranking and key findings: context-based personalization shifts decisions toward the AI without detectable changes in evaluations of the AI.Three panels show the study sequence. First, a participant ranks stocks A, B, and C. Second, style-based and context-based personalization are independently varied while the AI advocates an assigned ranking, with recommendation, evidence, and argumentative flow held constant. Third, an illustrative revised ranking places B above A and C. The findings highlight increased movement toward the AI with context-based personalization, no detectable change in AI evaluations, and no reliable ranking effect of style-based personalization. Abstract. As people turn to generative AI for financial advice, these systems can personalize how they communicate and what they say. Whether these forms of personalization shape decisions differently remains unclear. We conducted a preregistered 2 Ă 2 between-subjects factorial experiment (N=240N=240): participants ranked three comparably viable stocks, discussed them with an AI, and reranked them. Participants perceived both forms of personalization, but only context-based personalization reliably changed ranking behavior: it increased reconsideration and moved rankings toward the AIâs assigned recommendation. Participants felt more influenced without judging the AI as more correct, trustworthy, intelligent, likeable, or high-quality. Those initially farther from its recommendation moved more toward it while judging its advice less correct; exploratory analyses suggest greater susceptibility among lower-expertise participants. These findings show how personalized AI can steer decisions among defensible options with only a minimal evaluative trace, raising concerns for the design and governance of personalized decision support. Keywords: Conversational agents, Large language models, Personalized AI, Style-based personalization, Context-based personalization, AI-supported decision-making, Human-AI interaction 1. Introduction More than 200 million people ask generative AI for financial advice every month. Among Americans who followed such advice, one in five acted immediately, without asking a single follow-up question. Nearly a third said that following such advice hurt their finances.11 1 https://w.nerdwallet.com/finance/studies/using-ai-for-personal-finances Against this backdrop, in June 2026, ChatGPT introduced a personal finance experience that allows users to connect their financial accounts and save âfinancial memoriesâ to personalize responses.22 2 https://openai.com/index/personal-finance-chatgpt/ OpenAI, in its launch announcement, illustrates the feature with two responses to the same request about saving money: one generic, and the other personalized using the userâs location and salary. Both recommend reducing spending and automating savings, but the personalized response uses personal details and preferences to tailor the recommendationâs presentation. This shift toward personalized financial advice raises pressing questionsâIf the recommendation remains the same, can personalization change usersâ decisions? And does it also change how they perceive the AI giving that advice? AI can personalize advice in multiple ways (Motger et al., 2022; Sutcliffe, 2024). One form is style-based personalization, in which the AI adapts its manner of communication to a userâs preferences or personality profile. For example, ChatGPT ask users questions such as, âDo you like this personality?â, and use their feedback to adapt how it communicates. A second form is context-based personalization, in which the AI uses information about the user to adapt the framing and emphasis of its advice. The âfinancial memoriesâ described above is an example of context-based personalization. Prior work suggests that these two mechanisms may influence users differently. Here, persuasive influence refers to a shift in a userâs attitude or decision toward the position advanced by the AI. Ruijten (2021) found that a dominant AI system was more persuasive but less liked, whereas matching the AIâs communication style to the userâs personality improved evaluations without improving persuasion. Genç et al. (2026) likewise found that conversational AI personality33 3 Following Rahman and Desai (2026), we use conversational AI personality to mean the personality an AI expresses through its linguistic cues and communication style. affected perceived trustworthiness, competence, and emotional responses without reliably changing user decisions. Research on profile-personalized messages instead adapts the framing and emphasis of AI messages (Matz et al., 2017; Simchon et al., 2024). Personalized AI messages can change behavior and sometimes outperform non-personalized alternatives (Noar et al., 2007; Matz et al., 2017; Matz et al., 2024; Salvi et al., 2025). Yet the added benefit of personalization is inconsistent. Political messages generated with AI from demographic and political profiles have performed no better than nontargeted messages (Hackenburg and Margetts, 2024a; Hackenburg and Margetts, 2024b), while conversational studies identify prompting, post-training, and relevant evidence as stronger drivers of influence (Hackenburg et al., 2025; Lin et al., 2025). More broadly, trust can remain stable while reliance rises (Salimzadeh et al., 2024), and disclosing AI authorship or personalization does not necessarily eliminate influence (Matz et al., 2024; Gallegos et al., 2026). However, these findings leave the effects of both mechanisms unresolved. It remains unclear how either style-based or context-based personalization affects usersâ decisions and perceptions of the AI, or whether greater movement toward the AIâs recommendation is accompanied by more favorable perceptions of the AI. Existing research cannot answer these questions because they do not isolate the mechanisms. Some pair personalization with other factors, such as opinion alignment (Eder et al., 2026) or conversational warmth (Yazan et al., 2026). Others operationalize personalization by giving the AI access to a participant profile (Shen et al., 2026), without separately manipulating how it adapts its communication style and argumentative framing. We therefore lack a controlled test of their separate and combined effects that holds the AIâs recommendation, evidence, and argumentative flow constant. Prior research further suggests that this design space is shaped by user-level factors, as usersâ initial judgments and expertise may affect how readily they can evaluate or contest AI advice (Bonaccio and Dalal, 2006; Eichler and Schwab, 2025). To address these gaps, we conducted a controlled study of personalized AI advice in a preference-sensitive task (Elwyn et al., 2009). Participants made an initial choice, consulted an AI, and could then keep or revise their choice. In a preregistered 2Ă22Ă 2 between-subjects experiment (N=240N=240), participants ranked three anonymized pharmaceutical stocks, discussed them with an AI advocating a pre-assigned ranking, and reranked them. Real market data supported multiple defensible rankings, with no single objectively correct ordering given the available information. Figure 1 summarizes the study design. Through this study, we examine three research questions: RQ1: How do style-based and context-based personalization affect usersâ decisions? RQ2: How do style-based and context-based personalization affect usersâ perceptions of the AI? and RQ3: How does usersâ initial distance from the AIâs recommendation shape their perceptions and decision changes? We additionally conduct an exploratory analysis examining how self-reported investment expertise relates to usersâ perceptions and susceptibility to AI advice, and whether it moderates the effects of personalization. Together, these questions move from the effects of distinct personalization strategies to the role of initial disagreement to individual differences in susceptibility, offering a comprehensive account of personalized influence in AI-assisted decision-making. Our findings show that style- and context-based personalization have distinct consequences for AI-assisted decision-making. Specifically, with context-based personalization present, reconsideration increased from 35.0% to 59.2%, and the odds of greater movement toward the AIâs recommendation approximately doubled (OâR=2.04OR=2.04, 95% CI [1.23,3.38][1.23,3.38]). Neither form of personalization produced statistically detectable changes in the seven evaluative measures: response quality, perceived correctness, intelligence, likeability, performance trust, moral trust, or willingness to return (all Holm-corrected pâ„.47pâ„.47). We contribute: âą A Framework for Studying AI Personalization. We independently vary how an AI adapts its communication style and tailors its arguments to a userâs circumstances. Holding the recommendation, evidence, and conversational sequence constant allows us to test the separate and combined effects of these two forms of personalization. âą Decision Change Without Higher AI Ratings. Context-based personalization moved participants toward the AIâs recommendation without detectable changes in how they evaluated the AI. Participants recognized the personalization and reported greater influence, raising questions about whether awareness alone is enough to protect usersâ autonomy. âą Design Implications for Personalized AI Advice. We propose ways to make the AIâs goals and use of personal information visible, give users control over personalization, and assess decision changes alongside usersâ ratings. We also distinguish safeguards for personalized arguments from the longer-term evaluation needed for communication style. 2. Related Work 2.1. Conversational AI in Decision-Making Most research on AI-assisted decision-making examines a single-turn interaction in which an AI provides a prediction, sometimes with an explanation, and the user decides whether to accept or override it (Lai et al., 2023). CAs instead support multi-turn advice in which the AI can elaborate, answer objections, compare alternatives, and restate its position. These exchanges can improve appropriate reliance over static explanations (Ma et al., 2025), but can also produce overreliance when the AI is wrong (Li et al., 2025; Spatharioti et al., 2025). Lin et al. (2024) described this form of support as decision-oriented dialogue. CAs now support decisions from low-stakes reflection to high-stakes choices in life planning (Wang and Liu, 2025), medicine (Ramjee et al., 2025; Rajashekar et al., 2024), law (Schneiders et al., 2025), and public administration (Aljuneidi et al., 2026), and the influence of AI advice varies across these domains (Ikeda, 2024). We first distinguish usersâ perceptions of AI from changes in their decisions, then consider personalization and decision settings with multiple defensible answers. Perceptions of the AI and Changes in Participantsâ Decisions Peopleâs perceptions of AI advisors do not always align with their reliance on its advice. Automation bias names reliance on automated recommendations without adequate scrutiny (Parasuraman and Riley, 1997); adding an explanation can increase reliance on correct and incorrect answers alike (Kim et al., 2025); trust can remain unchanged while reliance rises under complexity and uncertainty (Salimzadeh et al., 2024); and explanations can improve understanding without changing reliance (He et al., 2023a). Conversely, radiologists evaluated identical advice less favorably when it carried an AI label, although the label changed neither accuracy nor reliance (Gaube et al., 2021). Reviews, therefore, caution against using self-reported trust as a proxy for changes in decisions (Vereschak et al., 2021; Schrills et al., 2026). The same separation appears in persuasive settings. Changes in attitudes and downstream actions may be uncorrelated (Hackenburg et al., 2026), and users can remain skeptical of individual claims even as the broader argument still moves them (Yeo et al., 2026). We therefore measure how participants regarded the AI, whether they recognized its influence, and pre-to-post changes in their decisions as separate outcomes. Measuring them separately makes it possible to identify changes in decisions that are not accompanied by corresponding changes in perceptions of the AI. Susceptibility to AI advice also varies with usersâ prior judgments and expertise. People often discount advice that falls farther from their initial judgment (Harvey and Fischer, 1997; Yaniv, 2004; Bonaccio and Dalal, 2006), although clinicians have moved toward algorithmic estimates even under strong initial disagreement (PĂĄlfi et al., 2022), and advice consistent with prior beliefs is trusted and accepted more readily (Bashkirova and Krpan, 2024). Participants who overestimated their own performance under-relied on an accurate AI (He et al., 2023b), and financial literacy predicts responses to robo-advice (Eichler and Schwab, 2025). We therefore examine the initial distance between participantsâ rankings and the AIâs assigned ranking (RQ3) and analyze self-reported investment expertise exploratorily. 2.2. Personalization in Conversational AI CAs can personalize many aspects of an interaction (Motger et al., 2022), and prior work separates an agentâs persona from a userâs profile, although the terms overlap (Sutcliffe, 2024). We instead distinguish how the AI communicates from how it uses participant information. style-based personalization adapts the AIâs communication style to the participantâs personality profile, and context-based personalization adapts the framing and emphasis of its arguments to the participantâs demographic and value profile. 2.2.1. Style-Based Personalization and UserâAgent Similarity Style-based personalization draws on similarity-attraction, the tendency to respond more positively to similar others (Byrne, 1971), which researchers have applied as a design principle in persuasive technology (Ruijten, 2021). Personality is recognizable in conversational cues, and matching it can affect evaluations. Users rated personality-matched synthesized voices and their content more favorably (Nass and Lee, 2001), readers rated advertisements written for their own Big Five trait more favorably (Hirsh et al., 2012), and advertising matched to inferred personality increased clicks and purchases (Matz et al., 2017). Reviews on personalized matching across persuasion research and human-agent interaction find the effects heterogeneous and dependent on which attribute is matched and on the process through which it operates (Teeny et al., 2021; Klein, 2025). Ruijten (2021) found that a dominant interaction style was more persuasive but less liked, whereas matching the systemâs style to the userâs personality improved perceptions without improving persuasion. In a charitable-giving task, AI personality affected perceived trustworthiness and competence but not donations (Genç et al., 2026). These findings reinforce the need to distinguish evaluations of an AI from its behavioral influence. Generative AI allows more flexible style adaptation than earlier systems built on a limited set of designed cues (Feine et al., 2019). Prompting can control expression of Big Five traits in CAs through reply length, signposting, punctuation, tone, hedging, and related choices (Rahman and Desai, 2026; Serapio-GarcĂa et al., 2025; Jiang et al., 2023; Ramirez et al., 2023; Jiang et al., 2024). This control makes it possible to revisit personality matching and to examine personality expression and userâagent fit within the same interaction. Rahman and Desai (2026) used this control in a goal-oriented task, found that both the agentâs personality expression and its alignment with the userâs personality shaped perceptions of the AI, and introduced Trait Modulation Keys (TMK), a prompting framework that jointly controls Big Five profiles and their communication styles; we use TMK for style-based personalization (Section 3.3.2). Prior work, however, has more often varied an AIâs personality than adapted its communication to an individual participant, and has focused more on evaluating user perceptions than on its impact on decisions. To address this, we test whether style-based personalization affects participantsâ decisions (RQ1), their perceptions of the AI (RQ2), or both. 2.2.2. Context-Based Personalization and Tailored Persuasion Context-based personalization adapts what an AI says based on information about the user. Such personalization raises persuasive relevance: arguments framed around recipientsâ moral values shift attitudes more than generic framings (Feinberg and Willer, 2015; Feinberg and Willer, 2019), and across 57 studies of print health materials personalized to readersâ demographics, behavior, and psychological attributes, personalization gave an advantage in changing health behaviors (Noar et al., 2007). AI can generate such messages from user profiles. Profile-personalized messages outperformed non-personalized ones even when the personalization was disclosed (Matz et al., 2024), and personality-matched political advertisements were rated more persuasive than mismatched ones (Simchon et al., 2024). Evidence that such personalization adds to AI persuasion remains mixed. Political messages generated from demographic and political profiles were no more persuasive than untargeted ones (Hackenburg and Margetts, 2024a; Hackenburg and Margetts, 2024b). Conversational findings are mixed as well. In short online debates, AI personalized to participantsâ sociodemographic profiles shifted post-debate agreement more often than human debaters (Salvi et al., 2025), while larger-scale studies found stronger effects from post-training and prompting than from personalization, and identified relevant facts and evidence as central to persuasion (Hackenburg et al., 2025; Lin et al., 2025). Context-personalized dialogues with AI reduced conspiracy beliefs by about 20%, with the effect lasting 2 months (Costello et al., 2024). The decision domain and contents of the user profile may explain part of this variation. Most prior research on personalized persuasion concerns political attitudes, but persuasive effects vary across communication contexts (Hölbling et al., 2025) and decision domains (Ikeda, 2024). Personal attributes may bear directly on the decision rather than serving only as targeting variables, and values connect an argument to a recipientâs priorities (Feinberg and Willer, 2015; Feinberg and Willer, 2019). Our context-based personalization therefore adapts the framing and emphasis of the argument to the participantâs demographic and value profile (Section 3.3.3), leaving the assigned recommendation, evidence, and argumentative flow unchanged. 2.3. Decision-Making with Multiple Defensible Answers Following Elwyn et al. (2009), we describe decisions as preference-sensitive when several options are defensible, and the preferred choice depends on how individuals weigh their benefits and risks. Evaluating AI advice in these settings requires distinguishing movement toward its recommendation from improvement in decision quality (Guo et al., 2024). Such tasks allow us to examine whether an AI recommendation changes a reasoned judgment without assuming that agreement with the AI is objectively correct. The absence of a single correct answer defines the decision setting; controlled comparisons are still needed to identify the effects of personalization. Investment selection provides one such setting. Portfolio choice balances expected return against the risk an investor will accept (Markowitz, 1952), and investors can weigh growth, valuation, and risk differently (Elliott et al., 2024). HCI studies have used robo-advised life-insurance choices (Bertrand et al., 2023), portfolios built from real ETFs (Reicherts et al., 2025), and real-stock rankings after conversations with an AI advisor (Takayanagi et al., 2025). Related work examines explanations in simulated financial advice (Ben David et al., 2021), trust in personal-investing robo-advisors (Cooney et al., 2026), and the acceptability of delegating investments to algorithms (Niszczota and KaszĂĄs, 2020). This literature establishes the investment selection task as a setting for studying responses to AI advice. 2.4. Isolating Style- and Context-Based Personalization Across these literatures, style-based and context-based personalization are rarely varied independently. Research on userâagent similarity typically varies how an agent communicates, such as an extraverted or introverted voice (Nass and Lee, 2001), a dominant or submissive interaction style (Ruijten, 2021), or a fast and proactive versus polite and accommodating conversational style (Kostric et al., 2025), while the content stays the same; psychological-targeting research instead changes a messageâs framing and emphasis (Matz et al., 2017; Matz et al., 2024; Simchon et al., 2024; Hackenburg and Margetts, 2024a). Recent work has begun to examine related dimensions together. Researchers have crossed an assistantâs personality with opinion alignment (Eder et al., 2026) and contextualization with warmth (Yazan et al., 2026), and a study of everyday advice-seeking conversations bundled demographics, values, and personality into one personalization factor (Shen et al., 2026). These designs examine multiple dimensions of an interaction, but none independently manipulates the manner of communication and the framing of the argument. Motivated by these gaps, we independently vary both forms in a preference-sensitive investment-ranking task. We test their effects on decisions (RQ1) and perceptions of the AI (RQ2), and examine associations with initial distance from the AIâs assigned ranking (RQ3). 3. Study Method To test whether personalizing how an AI communicates and how it frames its arguments changes usersâ perceptions and investment decisions, we conducted an Institutional Review Board-approved and preregistered experiment (N=240N=240). The study was built around three design requirements. First, the decision task needed to support multiple defensible answers so that influence could be studied without presuming that the AI held a uniquely correct position. Second, style- and context-based personalization needed to vary independently while other features of the AI interaction remained controlled. Third, the design needed to distinguish changes in participantsâ decisions from their perceptions of the AI. We used investment ranking as an instance of this decision setting (Section 2.3), selecting three anonymized pharmaceutical stocks from real market data as comparably viable long-term investments. Each company was presented with a bull caseâan argument for investingâand a bear caseâan argument against investingâso that participants had evidence supporting different conclusions. As reported in Section 4.1.3, participantsâ initial choices included all six possible rankings, with no stock or ordering emerging as an obvious favorite. The AI advocated one of these six rankings. Its ranking was assigned in advance, counterbalanced within each experimental cell, and held independent of the participantâs initial choice. Initial agreement or disagreement with the AI could therefore vary naturally rather than being imposed as part of the personalization manipulation. To isolate the two forms of personalization, we used a 2Ă22Ă 2 between-subjects factorial design with 60 participants per cell. Style-based personalization was either absent or present, and context-based personalization was independently either absent or present (see Table 1), producing four design cells: Baseline (Style: absent, Context: absent), Style-only (Style: present, Context: absent), Context-only (Style: absent, Context: present), and Combined (Style: present, Context: present). Style-based personalization adapted the AIâs manner of communication to the participantâs personality profile, whereas context-based personalization adapted the framing and emphasis of its arguments to the participantâs demographic and value profile. The AIâs assigned recommendation, underlying evidence, argumentative flow, conversational interface, and study procedure were otherwise held constant. Behavioral outcomes were measured before and after the conversation, adding Time as a within-participant measurement factor to the randomized 2Ă22Ă 2 design. We assessed behavioral influence through pre-to-post changes in participantsâ rankings, investment likelihoods, and confidence. Perceptions of the AI and manipulation checks were collected only after the conversation. We additionally examined whether participantsâ responses varied with the initial distance between their ranking and the AIâs assigned ranking. Section 3.1 describes the stock selection task, Section 3.2 defines the experimental cells, and Section 3.3 explains their prompt implementation. We then describe the measures, procedure, participants, and analysis strategy. Table 1. The 2Ă22Ă 2 between-subjects factorial design (N=240N=240). Style-based and context-based personalization were independently absent or present, with 60 participants per experiment cell. Style-based personalization Context-based personalization Absent Present Absent Baseline (n=60n=60) Style-only (n=60n=60) Present Context-only (n=60n=60) Combined (n=60n=60) 3.1. Designing the Stock Selection Task To assess the effect of personalization, we asked participants to rank three companies, labeled Company A, Company B, and Company C, by their attractiveness as long-term investments. Our task builds on prior HCI studies of investment decision support, which have asked participants to construct portfolios from real exchange-traded funds (Reicherts et al., 2025) or rank real stocks after a conversation with an AI advisor (Takayanagi et al., 2025). Unlike tasks scored against an expert ranking (Takayanagi et al., 2025), our task specified no correct ordering: all three were comparably viable investments whose ranking could depend on participantsâ priorities. Ranking rather than choosing one stock lets us measure partial movement toward or away from the AIâs ranking (Section 3.4.2). We selected three pharmaceutical companies in three stages. First, using the Morningstar industry report44 4 https://w.morningstar.com/ and Robinhood,55 5 https://robinhood.com/ we retained the 30 largest publicly traded U.S. pharmaceutical companies by market capitalization. Restricting candidates to traditional drug manufacturers ensured comparable fundamentals. All subsequent screening used Robinhood data. Second, we retained the five stocks with the largest positive net buying values over the preceding month, reflecting share purchases net of sales. Third, using Robinhoodâs reported market capitalizations, we selected three companies spanning a range of sizes. We used positive net buying as a shared indicator of investor demand and variation in company size to represent different growth profiles, supporting defensible rankings for different long-term preferences. We obtained the selected companiesâ market data from the NYSE66 6 https://w.nyse.com/ on June 2, 2026. The data included opening price, volume, market capitalization, daily and annual lows and highs, trailing 12-month earnings per share, price-to-earnings ratio, and price histories over six windows from 1 month to the full record. We obtained each bull and bear case from SoFi77 7 https://w.sofi.com/invest/ on the same date. We labeled the companies A, B, and C and removed specific drug names from the bull and bear cases. We used real company data to preserve ecological validity (Raggetti et al., 2017; van Brussel et al., 2024), while presenting the companies as fictional to limit the influence of participantsâ prior views. Experienced investors thus encountered plausible fundamentals and arguments. Section 4.1.3 confirms that participants treated all three as viable options. All companies used the same card format (Figure 2). The front showed its current price, change, and line chart over six selectable windows, and the same fundamentals (Figure 2(a)); the back showed its bull and bear cases (Figure 2(b)). Bull and bear cases supported any ordering; random card order avoided implying a ranking. (a) Front of the cards. Each card gave the current price, a line chart of the price history over a selectable window, and a grid of nine fundamentals. (b) Back of the cards. Tapping a card turned it over to a written bull case and bear case, so every company carried an argument on both sides. Figure 2. The stock selection task interface. Each company appeared on its own card, and that card carried all the information participants had about the company: the market data on the front and the arguments on both sides on the back.Two rows of screenshots. (a) Three cards side by side for Company A, Company B, and Company C. Each card shows a ticker, a price, the change over the past month, a row of tabs for the time window, a line chart of the price history, and a grid of fundamentals covering opening price, volume, market capitalization, day low and high, year low and high, earnings per share, and price-to-earnings ratio. A line at the bottom of each card reads ``Tap card for bull and bear case.âČ (b) The same three cards turned over, each showing a paragraph-length bull case in green followed by a bear case in red, with a line at the bottom reading ``Tap card for quote and chart.âČ 3.2. Experimental Design We randomly assigned participants to four experiment cells in a 2Ă22Ă 2 between-subjects factorial design (n=60n=60 per experimental cell; Table 1). Style-based personalization adapted the AIâs communication to the participantâs Big Five personality profile through reply length, signposting, punctuation, tone, and hedging. Context-based personalization adapted the framing and emphasis of stock arguments to the participantâs demographic and value profile. Crossing the two factors yielded these experiment cells: âą Baseline (Style absent, Context absent). The AI used plain declaratives and a flat tone, without a demographic or value profile to tailor its arguments. âą Style-only (Style present, Context absent). The AI adapted its communication to the participantâs personality, without a demographic or value profile to tailor its arguments. âą Context-only (Style absent, Context present). The AI used plain declaratives and a flat tone while connecting stock arguments to the participantâs circumstances and values. âą Combined (Style present, Context present). The AI adapted its communication to the participantâs personality and its argument framing to their circumstances and values. All four experiment cells used the same company evidence, ranking-assignment procedure, argumentative sequence, conversational interface, and study procedure. In every experiment cell, the AI received participantsâ initial rankings, investment likelihoods, and rationales. Thus, the baseline also responded to participantsâ task-specific input; the two factors varied personality-based communication and demographic- and value-based framing. Section 3.3 describes their prompt implementation. Assigning the AIâs Reference Ranking To measure movement toward the AI, we assigned it a ranking to advocate, following studies that assign AI systems a position and assess usersâ subsequent attitudes or choices (Jakesch et al., 2023; Williams-Ceci et al., 2026; Salvi et al., 2026; Werner et al., 2024; Chiang et al., 2024; Shen et al., 2026). Within each experiment cell, we rotated through all 3!=63!=6 rankings in participant-arrival order, assigning each ten times. The application fixed the ranking when participants reached the conversation screen, independently of their initial rankings or subsequent messages. The AI presented this ranking as its own judgment throughout the conversation. This counterbalancing separated the advocated position from personalization, while allowing initial agreement with the AI to vary naturally. Section 4.1.3 reports balance checks; debriefing disclosed the advance assignment (Section 3.5). 3.3. Prompt Design We used GPT-5.4 in all four experimental cells.88 8 Accessed through OpenRouter (https://openrouter.ai/), with reasoning effort set to medium. A prompt layer is a group of instructions within the same system prompt that governs one aspect of the AIâs responses. In our study, the system prompt contained four layers: recommendation, task, style, and context (Figure 3). The <recommendation> layer specified the assigned ranking and instructed the AI to present it as its own judgment. The <task> layer specified the argumentative sequence, turn-level rules, and evidence constraints. The <style> and <context> layers implemented the respective personalization factors using separate personality and demographic/value profiles. Recommendation instructions and task rules were identical across experimental cells; the assigned ranking varied according to the counterbalancing procedure. Every prompt included the company profiles, bull and bear cases, and participantsâ initial rankings, investment likelihoods, and rationales. The full prompts are available in the supplementary materials. 3.3.1. Shared Argumentative Sequence Because persuasive effects depend on implementation and communication context (Hölbling et al., 2025), <task> prescribed the same six stages across experimental cells: (1) Anchor. Open with the full ranking and one evidence-based clause per stock to establish a reference point (Tversky and Kahneman, 1974). Foreground the bull case for a stock the AI ranked higher than the participant did and the key risk of their favored stock, or affirm an identical ranking. Acknowledge their rationale (Cialdini, 2007) and ask why they favored their top pick. (2) Justify. Support the top-ranked stock with issue-relevant evidence and numbers, following the central route of the Elaboration Likelihood Model (Petty and Cacioppo, 1986). (3) Compare first versus second. Concede a strength of the second-ranked stock, then defend the first, using two-sided argumentation (Allen, 1991; OâKeefe, 1999). (4) Compare second versus third. Repeat the two-sided comparison for the lower pair. (5) Consolidate criteria. Restate the ranking in terms of current financial support, long-term growth, and risk concentration, using emphasis framing (Entman, 1993). (6) Summary. Restate the ranking with a one-line rationale per stock. The prompt required one stage per reply, in sequence, with brief answers to participantsâ questions before returning to the flow. It restricted numbers and bull or bear points to the displayed company profiles. It also required periodic reminders that the decision remained the participantâs own, an autonomy-preserving formulation associated with lower reactance (Carpenter, 2013), and prohibited asking participants to revise their rankings. Section 4.1.4 assesses adherence to these constraints. A two-by-two grid of four cards, one per experimental cell, on axes for style-based personalization (absent, present) across the top and context-based personalization (absent, present) down the side. Baseline sits top left with the style layer set to flat tone and the context layer absent; Style-only, top right, sets the style layer to the TMK profile; Context-only, bottom left, sets the context layer to the intake profile; Combined, bottom right, sets both. Each card also shows a robot whose antennas count the personalized layers: none for Baseline, one each for Style-only and Context-only, and two for Combined. A gray band below the grid, headed the other two layers identical in all four experimental cells, gives the recommendation layer as an assigned ranking, one of the six possible orderings argued as the AI's own judgment, and the task layer as a six-stage flow of anchor, justify, compare one and two, compare two and three, consolidate, and summary, in order. Figure 3. Independent personalization factors implemented through prompt layers. Each experimental cell sets <style> and <context>; recommendation instructions and task rules are shared. The assigned ranking is counterbalanced within each experimental cell.A two-by-two grid of four cards, one per experimental cell, on axes for style-based personalization (absent, present) across the top and context-based personalization (absent, present) down the side. Baseline sits top left with the style layer set to flat tone and the context layer absent; Style-only, top right, sets the style layer to the TMK profile; Context-only, bottom left, sets the context layer to the intake profile; Combined, bottom right, sets both. Each card also shows a robot whose antennas count the personalized layers: none for Baseline, one each for Style-only and Context-only, and two for Combined. A gray band below the grid, headed the other two layers identical in all four experimental cells, gives the recommendation layer as an assigned ranking, one of the six possible orderings argued as the AI's own judgment, and the task layer as a six-stage flow of anchor, justify, compare one and two, compare two and three, consolidate, and summary, in order. 3.3.2. Style Instructions With style present, <style> used the Trait Modulation Keys (TMK) of Rahman and Desai (2026), which map low, medium, or high levels of each Big Five trait to communication cues. We converted participantsâ Mini International Personality Item Pool scores (Section 3.4.1) into five trait means and banded each trait. Because the Mini-IPIP has no published norms and the International Personality Item Pool recommends local norms,99 9 https://ipip.ori.org/ we fitted one-dimensional k-means clusters to a publicly available, demographically matched dataset of 250 participants. Silhouette scores were 0.57â0.64; across 1,000 bootstrap resamples, 91â98% of participants retained their band. The banded profile populated <style>. For high extraversion, the instructions read: You are active, energetic, and talkative. You are extraverted, bold, and assertive in conversations. Insert exclamation marks for enthusiasm, unicode emojis or interjections [âŠ]. With style absent, <style> prescribed plain declaratives and a flat tone. 3.3.3. Context Instructions With context present, <context> supplied age group, gender, ethnicity, education, occupation, location, household income, household status, and four higher-order Schwartz value dimensions (Section 3.4.1). It instructed: Every reply ties the argument to at least one concrete fact of their life, addressed as âyou / yourâ â their work, their age and life stage, living single or with others, their income band, or where they live. Name the fact and connect it to the stock point in the same breath. The instructions standardized which attributes entered each stage: the strongest value and occupation set the anchorâs leading criterion; occupation and education shaped the justification, analogy, and numerical density; income and household set the stakes in the first comparison; location and age set the time horizon in the second; and closing stages drew on the full profile. Values appeared as plain priorities (e.g., âyouâd rather own something that lasts than chase a quick flipâ). These rules standardized the extent and timing of personalization while preserving the assigned ranking, stock facts, and argumentative sequence. With context absent, the prompt supplied no demographic or value profile. 3.4. Measures We describe four sets of measures: intake measures, behavioral measures, perception measures, and manipulation-check indices. Table 2 lists every measure with its range, when it was collected, and the research question it serves. Table 2. Measures, when they were collected, and where they are used. Pre and Post refer to the behavioral battery before and after the AI conversation; d1d_1 is the initial distance between the participantâs ranking r1r_1 and the AIâs assigned ranking a; Expl. is the exploratory expertise analysis. Definitions are given in the sections named in each group heading. Measure Range Collected Used in Intake measures (Section 3.4.1) Demographics and circumstances Categorical Intake <context> input Personal values (SSVS) â1-1 to 8 Intake <context> input Big Five personality (Mini-IPIP) 1â5 Intake <style> input Investment experience 4 levels Intake Expl. Behavioral battery (Section 3.4.2) Stock ranking r1r_1, r2r_2 6 orderings Pre, Post Basis of ranking measures Investment likelihood LsL_s 0â100 Pre, Post Basis of likelihood measures Confidence C 0â100 Pre, Post Basis of confidence change Rationale Text Pre, Post AI prompt input Derived decision measures Initial distance d1=FâĄ(r1,a)d_1=F(r_1,a) 0, 2, 4 Pre RQ3 (moderator) Reconsideration 0/1 Pre, Post RQ1, RQ3, Expl. Movement toward AI â(4âd1)-(4-d_1) to d1d_1 Pre, Post RQ1, RQ3, Expl. Rate of change â1-1 to 1 Pre, Post RQ3 check Ranking change (footrule) 0, 2, 4 Pre, Post RQ1, RQ3, Expl. Ranking change (Kendall) 0â3 Pre, Post RQ1, RQ3, Expl. Likelihood shift toward AI â200-200 to 200 Pre, Post RQ1, RQ3, Expl. Likelihood change, AI-top stock â100-100 to 100 Pre, Post RQ1, RQ3, Expl. Total absolute likelihood change 0â300 Pre, Post RQ1, RQ3, Expl. Confidence change â100-100 to 100 Pre, Post RQ1, RQ3, Expl. Perception measures (Section 3.4.3) Response quality 1â7 Post RQ2, RQ3, Expl. Rightness judgment 1â7 Post RQ2, RQ3, Expl. Return likelihood 1â7 Post RQ2, RQ3, Expl. Likeability (Godspeed) 1â5 Post RQ2, RQ3, Expl. Perceived intelligence (Godspeed) 1â5 Post RQ2, RQ3, Expl. Performance trust (MDMT) 0â7 Post RQ2, RQ3, Expl. Moral trust (MDMT) 0â7 Post RQ2, RQ3, Expl. Perceived influence 1â7 Post RQ2, RQ3, Expl. Manipulation checks (Section 3.4.4) Style index 1â7 Post Manipulation validation Context index 1â7 Post Manipulation validation 3.4.1. Intake Measures The intake questionnaire covered demographics, AI use, investment experience, and current stock holdings; the Short Schwartz Value Survey (Lindeman and Verkasalo, 2005), aggregated into Schwartzâs four higher-order dimensions (Schwartz, 1992); and the 20-item Mini International Personality Item Pool (Mini-IPIP) (Donnellan et al., 2006) for the Big Five. Demographics and values supplied <context>; Big Five scores supplied <style>. 3.4.2. Behavioral Measures Before and after the conversation, participants ranked stocks from most (1) to least (3) attractive as long-term investments, rated investment likelihood for each and ranking confidence, and explained their ranking, following preâpost advice-taking designs (Bonaccio and Dalal, 2006) and AI-reliance research (Raees et al., 2026). Let r1r_1, r2r_2, and a denote participantsâ pre-conversation, post-conversation, and AI-advocated rankings, respectively. We use two established rank metrics (Diaconis and Graham, 1977) used in HCI to compare predicted and user-adjusted rankings (Sertkan et al., 2020) and stakeholder and decision-support-system rankings (Shrestha et al., 2023). The Spearman footrule (Spearman, 1906) sums each stockâs absolute change in position, FâĄ(x,y)=âsâA,B,C|posxâ(s)âposyâ(s)|,F(x,y)= _sâ\A,B,C\ |pos_x(s)-pos_y(s) |, and the Kendall tau distance (Kendall, 1938) counts stock pairs in opposite relative order, KâĄ(x,y)=#âĄs,sâČ:(posxâ(s)âposxâ(sâČ))â(posyâ(s)âposyâ(sâČ))<0.K(x,y)=\# \\s,s \: (pos_x(s)-pos_x(s ) ) (pos_y(s)-pos_y(s ) )<0 \. Over three stocks, Fâ0,2,4Fâ\0,2,4\ and Kâ0,1,2,3Kâ\0,1,2,3\: F counts how far stocks traveled, K how many pairwise preferences flipped. Initial distance, d1=FâĄ(r1,a)d_1=F(r_1,a), measures pre-exposure disagreement with the AI and moderates RQ3. Reconsideration is 1 if r2â r1r_2â r_1 and 0 otherwise. Movement toward the AI is d1âFâĄ(r2,a)d_1-F(r_2,a), positive when participants ended closer to the AI, and ranking change is FâĄ(r1,r2)F(r_1,r_2) or KâĄ(r1,r2)K(r_1,r_2), movement regardless of direction; these generalize answer switching and agreement with the AI (Yin et al., 2019; Raees et al., 2026) to orderings. With Î denoting post minus pre and t and b the stocks the AI ranked first and last, likelihood shift toward the AI is ÎâLtâÎâLb L_t- L_b, likelihood change for the AI-top stock is ÎâLt L_t, total absolute likelihood change is âs|ÎâLs| _s| L_s|, and confidence change is ÎâC C. Movement toward the AI cannot exceed d1d_1; initially agreeing participants (d1=0d_1=0) cannot move closer. We compute a rate of change, (d1âFâĄ(r2,a))/d1 (d_1-F(r_2,a) )/d_1 for d1>0d_1>0, the fraction of available room used, as a robustness check (Appendix B.2). It applies the weight-of-advice normalization, dividing movement toward an advisor by the initial gap for human and algorithmic advice (Harvey and Fischer, 1997; Yaniv, 2004; Logg et al., 2019), to complete rankings by replacing scalar differences with footrule distances; it parallels the switch fraction of Yin et al. (2019), whose denominator includes only initially disagreeing participants. For example, a participant starting at r1=(B,A,C)r_1=(B,A,C) when the AI advocates a=(A,B,C)a=(A,B,C) has d1=FâĄ(r1,a)=2d_1=F(r_1,a)=2: two stocks are each one position away. Ending at r2=(A,B,C)r_2=(A,B,C) gives a final distance of FâĄ(r2,a)=0F(r_2,a)=0, movement toward the AI of 2â0=22-0=2, and a rate of change of 2/2=12/2=1, using all available room. Swapping the top two stocks yields ranking changes of FâĄ(r1,r2)=2F(r_1,r_2)=2 and KâĄ(r1,r2)=1K(r_1,r_2)=1. If the participantâs likelihood for stock A rose by 20 points, fell by 10 for stock C, and was unchanged for stock B, the likelihood shift toward the AI is 20â(â10)=3020-(-10)=30, the likelihood change for the AI-top stock is 20, and the total absolute likelihood change is 30. 3.4.3. Perception Measures We assembled eight perception measures spanning Wei et al. (2023)âs categories of interaction, ability, and humanness. We report Cronbachâs α for every multi-item measure. Three single items adapted from Cheng et al. (2026) capture response quality, rightness judgment of the AIâs ranking, and return likelihood for similar questions. Two Godspeed scales (Bartneck et al., 2009) each average five semantic-differential pairs: likeability (e.g., unfriendly to friendly) and perceived intelligence (e.g., incompetent to competent). Two subscales of the Multi-Dimensional Measure of Trust (Ullman and Malle, 2019), averaged over rated items, capture performance trust (eight reliable and capable adjectives) and moral trust (eight ethical and sincere adjectives). Perceived influence averages three agreement statements assessing AI influence on the final decision, effects on ranking, and consideration in deciding. 3.4.4. Manipulation Questions Four agreement statements formed two indices. The style index averages two items asking whether the AIâs communication felt similar to the participantâs own and tailored to them; the context index averages two items assessing whether its recommendations appeared tailored to the participant and drew on information about them. 3.5. Procedure A horizontal flow diagram of the study. Participants move through intake, a pre-conversation behavioral assessment, random assignment to the Baseline, Style-only, Context-only, or Combined experimental cell, an AI conversation, a post-conversation behavioral and perception assessment, and debriefing. A bracket links the pre- and post-conversation assessments to indicate that participants completed the same behavioral battery twice. A band below lists controls shared across experimental cells: three anonymized, comparably viable companies; bull and bear cases for each company; randomized company presentation order; and one of six counterbalanced AI rankings. Figure 4. Study procedure and controls. Participants completed the intake and the first behavioral battery before random assignment to one of four experimental cells, conversed with an AI advocating a counterbalanced ranking, repeated the battery, and received a debriefing. Company information, evidence structure, presentation-order randomization, and ranking counterbalancing remained constant across experimental cells.A horizontal flow diagram of the study. Participants move through intake, a pre-conversation behavioral assessment, random assignment to the Baseline, Style-only, Context-only, or Combined experimental cell, an AI conversation, a post-conversation behavioral and perception assessment, and debriefing. A bracket links the pre- and post-conversation assessments to indicate that participants completed the same behavioral battery twice. A band below lists controls shared across experimental cells: three anonymized, comparably viable companies; bull and bear cases for each company; randomized company presentation order; and one of six counterbalanced AI rankings. Participants entered the custom study application through Prolific1010 10 https://w.prolific.com/ and completed four phases: intake, pre-conversation assessment, AI conversation, and post-conversation assessment (Figure 4). Intake. Participants reviewed the study information, provided consent, and completed the intake questionnaire (Section 3.4.1), which included demographic, value, and personality profile. When enabled, personalization drew on these responses. Pre-conversation assessment. Participants read the task instructions, reviewed the three randomly ordered company profiles, and completed the behavioral battery (Section 3.4.2). AI conversation. Participants were randomly assigned to one of the four experiment cells, determining whether style-based and context-based personalization were present during the AI conversation (Table 1). The conversation screen displayed the participantâs initial ranking beside the AIâs assigned ranking and kept all company profiles accessible (Figure 5). The AI opened with its ranking and advocated for it throughout the conversation. Participants could respond freely and ask questions. The âReady to rankâ button appeared after the participant sent five messages; the application ended the conversation after ten minutes. Post-conversation assessment. With the transcript still visible, participants repeated the behavioral battery; instructions stated that retaining and changing their initial responses were equally acceptable. They then completed the eight perception measures (Section 3.4.3) and four manipulation checks (Section 3.4.4). Company names were anonymized, and participants were not told that the profiles represented real companies or that the AIâs ranking had been assigned in advance. Identity masking limited the influence of prior company-specific knowledge and attitudes, while withholding the ranking assignment avoided changing how participants interpreted the recommendation. During debriefing, we explained all three disclosures, offered data withdrawal, and provided the Prolific completion code. Every phase was identical across experiment cells; only style-based and context-based personalization differed. Figure 5. The conversation interface. The upper panel displayed the participantâs initial ranking beside the AIâs assigned ranking, with the three company profiles accessible below. The lower panel contained the conversation. The âReady to rankâ button became available after the participant had sent five messages. This screenshot was captured after the fifth message, with the transcript scrolled back to the opening exchange.A conversation interface with two sections. The upper section displays the participant's ranking and the AI's assigned ranking side by side, followed by three partially visible company cards. The participant ranks Company B first, Company A second, and Company C third; the AI ranks Company A first, Company B second, and Company C third. The lower section displays a conversation in which the AI explains and defends its assigned ranking. A message box appears at the bottom, with a ``Ready to rankâČ button in the lower-right corner. 3.6. Participants An a priori power analysis for a 2Ă2 between-subjects ANOVA indicated that 212 participants (53 per experimental cell) would provide 95% power to detect a medium effect (f=.25f=.25, d=.50d=.50) at α=.05α=.05, consistent with effect sizes reported in empirical HCI studies (Ortloff et al., 2025). We recruited 240 participants through Prolific (60 per experimental cell), providing 80% power to detect effects of f=.185f=.185 (d=.37d=.37). Eligibility required U.S. residence, prior stock-market investment, and self-reported market familiarity. We paid participants $13.30 per hour. Of the 240 participants, 136 (56.7%) identified as men, 103 (42.9%) as women, and 1 (0.4%) as nonbinary. Participantsâ ages spanned 18â29 (46, 19.2%), 30â49 (132, 55.0%), 50â64 (47, 19.6%), and 65 or older (15, 6.3%). For educational attainment, 111 participants (46.3%) had a bachelorâs degree, 54 (22.5%) some college or vocational training, 50 (20.8%) a graduate or professional degree, and 25 (10.4%) a high school education or equivalent. Household income ranged from under $25,000 (14, 5.8%) to $200,000 or more (13, 5.4%), with 196 participants (81.7%) between $25,000 and $149,999. All participants spoke English as their primary language, and 185 (77.1%) reported using AI tools daily or more often. Self-reported investment experience was Beginner for 35 participants (14.6%), Intermediate for 117 (48.8%), Advanced for 79 (32.9%), and Expert for 9 (3.8%). In addition, 221 participants (92.1%) reported currently holding stock investments. We revisit this experience measure in the exploratory analysis, contrasting Low-expertise participants (Beginner or Intermediate; n=152n=152, 63.3%) with High-expertise participants (Advanced or Expert; n=88n=88, 36.7%). On the Big Five (1â5 scale), profiles averaged Openness (M=4.18M=4.18, SâD=0.79SD=0.79), Conscientiousness (M=3.87M=3.87, SâD=0.90SD=0.90), Extraversion (M=2.90M=2.90, SâD=1.14SD=1.14), Agreeableness (M=4.01M=4.01, SâD=0.89SD=0.89), and Neuroticism (M=2.35M=2.35, SâD=0.92SD=0.92). 3.7. Data Analysis We used a 2Ă22Ă 2 factorial analysis with style-based and context-based personalization as effect-coded between-subjects factors, including both main effects and their interaction.. Preliminary checks assessed sample and task balance using chi-square tests for categorical measures and factorial ANOVAs for continuous measures. We examined message counts and self-reported manipulation indices with factorial ANOVAs, assessed internal consistency with Cronbachâs α, and evaluated manipulation delivery and protocol adherence through an independent LLM-as-judge analysis (Section 4.1.4). For RQ1, we modeled reconsideration with binary logistic regression, the three ranking outcomes with proportional-odds regression, and investment-likelihood and confidence changes with factorial ANOVAs. Robustness checks tested the proportional-odds assumption and adjusted decision models for conversation time and message count. For RQ2, we analyzed perceptions of the AI with factorial ANOVAs and post-hoc exploratory equivalence tests using ±0.40± 0.40 SD bounds. For continuous outcomes in RQ1 and RQ2, aligned rank transforms (Wobbrock et al., 2011), computed with ARTool (Kay et al., 2025), provided nonparametric robustness checks preserving the factorial structure. For RQ3, we examined Pearson correlations between initial distance from the AIâs ranking and each outcome, then tested moderation by personalization. A rate-of-change analysis accounted for available room to move toward the AI. Finally, exploratory analyses compared lower versus higher investment expertise using Welch t tests and a two-proportion z test for reconsideration, then tested moderation by personalization. We report descriptive statistics, 95% confidence intervals (CIs), partial η2η^2 for ANOVAs, odds ratios for logistic models, and Cohenâs d for present-versus-absent contrasts. Holm corrections applied within behavioral and perception families (Sections 3.4.2 and 3.4.3), separately for each factorial term. M and SâDSD denote mean and standard deviation. Unless otherwise noted, tests were two-sided with α=.05α=.05. 4. Results 4.1. Preliminary Analyses To validate our interpretation of the personalization effects, we first assessed sample balance, conversational engagement, and the internal consistency of the self-report measures, and checked whether the stock selection task and personalization manipulations functioned as intended. 4.1.1. Sample and Interaction Checks Chi-square tests indicated comparable demographic compositions, and factorial ANOVAs detected no differences in participantsâ Big Five or value profiles between levels of either factor. Participants exchanged a median of 11 messages with the AI. Both manipulations modestly reduced message counts. With style-based personalization, mean counts fell from 12.7 (SâD=3.2SD=3.2) when absent to 11.5 (SâD=3.7SD=3.7) when present, FâĄ(1,236)=6.39F(1,236)=6.39, p=.012p=.012; with context-based personalization, from 12.6 (SâD=3.7SD=3.7) to 11.6 (SâD=3.2SD=3.2), FâĄ(1,236)=5.47F(1,236)=5.47, p=.020p=.020; the interaction was not statistically significant, FâĄ(1,236)=1.09F(1,236)=1.09, p=.297p=.297. Because message counts followed random assignment, we treat engagement differences as manipulation side effects rather than randomization failures. The unadjusted models remain our primary analyses, and adding conversation time and message count preserved the Context-based effect in the ordinal logistic model of movement toward the AI and the binary logistic model of reconsideration (Section 4.2.3). Detailed results are available in the supplementary materials. 4.1.2. Self-report measure internal reliability Internal consistency was high across all multi-item self-report measures. Cronbachâs α ranged from .91 to .97 for the multi-item perception measures: likeability, perceived intelligence, performance trust, moral trust, and perceived influence. For the manipulation-check indices, α was .82 for the style index and .78 for the context index. 4.1.3. Validating the Stock Selection Task Participants treated all three stocks as viable, as intended (Section 3.1). All six stock orderings occurred, with no detected distributional difference across the four experiment cells (Ï2â(15)=13.69Ï^2(15)=13.69, p=.549p=.549). Any individual stockâs rank position (Ï2Ï^2 tests, p=.222p=.222â.692.692) or mean self-assigned rank (F tests, p=.079p=.079â.658.658) was likewise comparable; factorial balance tests follow. The mean initial investment likelihood exceeded the scale midpoint for every stock (57.3, 68.6, and 51.2 on the 0â100 investment scale- see Table 2). The three options, therefore, attracted participants with varying preferences rather than concentrating everyoneâs preference on one obvious option. Before the conversation, participants were moderately confident in their initial rankings (M=73.1M=73.1, SâD=18.2SD=18.2 on the 0â100 confidence scale - see Table 2). The Spearman footrule distance (Spearman, 1906) over three stocks takes only the values 0, 2, and 4 (§3.4.2). They also started at varying degrees of disagreement with the AIâs recommendation, which had not yet been revealed to them: 35 participants (14.6%) began at footrule distance 0 from the AIâs ranking (already agreeing), 92 (38.3%) at distance 2, and 113 (47.1%) at the maximum distance of 4. Across all participants, the mean initial footrule distance was 2.65 (SâD=1.43SD=1.43). For style-based personalization, the mean initial distance was 2.65 (SD = 1.49) when absent and 2.65 (SD = 1.38) when present. For context-based personalization, it was 2.67 (SD = 1.45) when absent and 2.63 (SD = 1.42) when present. We applied the randomization checks described in Section 3.7 to the full set of pre-randomization measures, covering initial orderings, per-stock ranks, per-stock investment likelihoods, initial confidence, personality, and the counterbalanced AI-ranking assignment. Of the 39 factorial tests, four reached p<.05p<.05 uncorrected, close to the two expected by chance, and none survived Holm correction within the family (all corrected pâ„.246pâ„.246). Initial distance, which determines participantsâ available room to move toward the AI, showed no significant main effects or interaction (Style-based, FâĄ(1,236)<0.01F(1,236)<0.01, p>.99p>.99; Context-based, FâĄ(1,236)=0.03F(1,236)=0.03, p=.858p=.858; Interaction effect, FâĄ(1,236)=1.16F(1,236)=1.16, p=.283p=.283). 4.1.4. Validating the Manipulations We assessed whether each manipulation affected its corresponding perception, using two self-report indices (1â7): a style index (perceived adaptation to the userâs communication style; α=.82α=.82) and a context index (perceived tailoring to the userâs personal situation; α=.78α=.78). We complement the self-reports with an independent LLM-as-judge analysis of the chat transcripts. Self-Reported Indices We summarize the two indices in Table 9. We applied each manipulation to half the sample, so the direct test of whether it landed contrasts the 120 participants who received it against the 120 who did not. Each index showed a statistically significant effect of its corresponding manipulation; crossed effects and interactions did not reach statistical significance. Both manipulations increased their corresponding perception indices. The style index was higher with style-based personalization present than absent (M=5.23M=5.23 vs. 4.724.72; FâĄ(1,236)=6.37F(1,236)=6.37, p=.012p=.012), and the context index was higher with context-based personalization present than absent (M=5.57M=5.57 vs. 3.393.39; FâĄ(1,236)=127.16F(1,236)=127.16, p<.001p<.001). Neither crossed effect nor any interaction effects reached statistical significance (all p>.05p>.05). These checks indicate that participants perceived both forms of personalization. Constant Evidence, Flow, and Recommendation in the Transcripts Attributing differences between factor levels to the personalization layers requires that the AI use the same evidence, follow the same six-stage flow, and advocate its assigned ranking at every level. We checked all three in the 240 transcripts using GLM-5.2 as the primary LLM judge (three passes per session, blind to cell assignment). Based on its ratings, information coverage, and the share of the displayed evidence the AI used, was equivalent between the present and absent levels of each factor (TOST, bounds ±0.5± 0.5 points and ±0.4± 0.4 SD; all pâ€.008pâ€.008). In 99.2%99.2\% of sessions, the AI opened with a recommendation, argued for it, defended it across turns, added arguments, and in every session advocated the assigned ranking. The transcripts also show the personalization being delivered: style-based personalization raised judged style adaptation (FâĄ(1,236)=29.1F(1,236)=29.1, p<.001p<.001, ηp2=.11 _p^2=.11) and context-based personalization raised judged context tailoring (FâĄ(1,236)=168.3F(1,236)=168.3, p<.001p<.001, ηp2=.42 _p^2=.42). Repeating the evaluation and analyses with MiniMax-M3 on the same 240 transcripts reproduced the effects on style adaptation and context tailoring (both p<.001p<.001) and information-coverage equivalence between factor levels (all TOST pâ€.024pâ€.024). 4.2. Effects of Personalization on Decisions (RQ1) We first ask whether participants changed their initial rankings at all and then quantify how far, and in which direction, rankings and investment likelihoods moved. 4.2.1. Reconsideration After the conversation, 47.1% of participants (113/240) changed their initial ranking. Reconsideration was more common when each form of personalization was present than absent (Figure 6): 59.2% (71/120) with context-based personalization present against 35.0% (42/120) with it absent, and 53.3% (64/120) with style-based personalization present against 40.8% (49/120) with it absent. In the effect-coded factorial logistic regression, Context-based personalization statistically significantly predicted reconsideration (odds ratio OâR=2.76OR=2.76, 95% CI [1.62,4.68][1.62,4.68], p<.001p<.001); participants who received it were 1.7 times as likely to reconsider their investment ranking as those who did not. The effect of Style-based personalization within the model was positive and statistically significant (OâR=1.72OR=1.72, 95% CI [1.01,2.93][1.01,2.93], p=.044p=.044) but did not meet the threshold for statistical significance after Holm correction within the behavioral-measures family (pHolm=.351p_Holm=.351). The interaction between Context-based and Style-based personalization was not statistically significant (OâR=0.67OR=0.67, 95% CI [0.23,1.94][0.23,1.94], p=.462p=.462). Two bar-chart panels. The Style-based panel shows 40.8 percent (49 of 120) with style absent and 53.3 percent (64 of 120) with style present, with a bracket labeled n.s. The Context-based panel shows 35.0 percent (42 of 120) with context absent and 59.2 percent (71 of 120) with context present, with a bracket labeled with three stars. Absent bars are gray; present bars carry the factor color. Figure 6. Percentage of participants who changed their stock ranking after the conversation, by factor level. Each panel contrasts the 120 participants who received one form of personalization (present) with the 120 who did not (absent); because the factors are crossed, every participant appears once in each panel. The brackets mark the factorial logistic terms: context-based personalization raised reconsideration by 24.2 percentage points (âp<.001^***p<.001), whereas the 12.5-point style-based difference was nominally significant (p=.044p=.044) but not after Holm correction (n.s.).Two bar-chart panels. The Style-based panel shows 40.8 percent (49 of 120) with style absent and 53.3 percent (64 of 120) with style present, with a bracket labeled n.s. The Context-based panel shows 35.0 percent (42 of 120) with context absent and 59.2 percent (71 of 120) with context present, with a bracket labeled with three stars. Absent bars are gray; present bars carry the factor color. 4.2.2. Magnitude and Direction of Decision Change We summarize the continuous decision outcomes by factor level in Table 3. The pattern appears in the marginal means. With context-based personalization present rather than absent, movement toward the AIâs ranking rose from M=0.98M=0.98 to M=1.48M=1.48, and the magnitude of ranking change rose on both distance metrics (footrule 1.02 to 1.72; Kendall 0.56 to 0.92). Style-based personalization produced no comparable shift on any ranking outcome. The investment-likelihood shifts (0â100 sliders) and confidence change showed no gradient on either factor, and confidence rose at every factor level. Table 3. Decision outcomes by factor level: mean ± SâDSD [95% CI], contrasting the 120 participants who received each form of personalization (Present) with the 120 who did not (Absent). Because the factors are crossed, every participant appears once under each factor. Footrule and Kendall measures are rank distances over the three stocks; likelihood and confidence changes are derived from 0â100 ratings, with ranges given in Table 2. Style-based Context-based Outcome Absent Present Absent Present Changed ranking, % (n/120) 40.8% (49) 53.3% (64) 35.0% (42) 59.2% (71) Move toward AI (footrule) 1.25 ± 1.68 [0.95, 1.55] 1.22 ± 1.47 [0.95, 1.48] 0.98 ± 1.49 [0.71, 1.25] 1.48 ± 1.63 [1.19, 1.78] Ranking change (footrule) 1.28 ± 1.68 [0.98, 1.59] 1.45 ± 1.53 [1.17, 1.73] 1.02 ± 1.51 [0.74, 1.29] 1.72 ± 1.63 [1.42, 2.01] Ranking change (Kendall) 0.70 ± 0.96 [0.53, 0.87] 0.78 ± 0.87 [0.62, 0.93] 0.56 ± 0.88 [0.40, 0.72] 0.92 ± 0.92 [0.75, 1.08] Likelihood shift toward AI 29.2 ± 40.3 [21.9, 36.5] 25.8 ± 35.4 [19.4, 32.2] 27.4 ± 36.9 [20.7, 34.0] 27.6 ± 39.0 [20.6, 34.7] Likelihood change, AI-top stock 17.8 ± 26.0 [13.1, 22.5] 15.7 ± 20.9 [11.9, 19.4] 16.0 ± 22.1 [12.0, 20.0] 17.5 ± 25.1 [12.9, 22.0] Total absolute likelihood change 56.5 ± 41.5 [49.0, 64.0] 49.7 ± 42.3 [42.0, 57.3] 50.3 ± 41.8 [42.7, 57.8] 55.9 ± 42.1 [48.3, 63.5] Confidence change 8.8 ± 17.9 [5.5, 12.0] 5.4 ± 11.7 [3.3, 7.5] 8.4 ± 15.9 [5.5, 11.2] 5.8 ± 14.4 [3.2, 8.4] 4.2.3. Factorial Tests The three ranking outcomes take a small number of ordered values, so we model them with proportional-odds regression rather than ANOVA (Section 3.7). Context-based personalization raised the odds of a larger ranking change on every one: movement toward the AI (OâR=2.04OR=2.04, 95% CI [1.23,3.38][1.23,3.38], pHolm=.028p_Holm=.028), ranking-change magnitude on the footrule distance (OâR=2.53OR=2.53, [1.53,4.19][1.53,4.19], pHolm=.002p_Holm=.002), and on the Kendall distance (OâR=2.48OR=2.48, [1.50,4.10][1.50,4.10], pHolm=.002p_Holm=.002). Style-based personalization had no statistically significant effect on any ranking outcome (all OâRâ€1.40OR†1.40, all pHolm=1.00p_Holm=1.00), and no Style Ă Context interaction reached statistical significance (all pHolmâ„.479p_Holmâ„.479). Figure 7 plots all nine terms with 95% confidence intervals; Table 14 in Appendix B gives the estimates numerically. A proportional-odds model assumes a factor shifts the odds equally at every cut point of the outcome. We tested this for the Context-based effect by fitting a separate binary model at each cut point and testing whether the resulting coefficients differ. They did not, for movement toward the AI (Ï2â(2)=0.35Ï^2(2)=0.35, p=.837p=.837), footrule change (Ï2â(1)=0.67Ï^2(1)=0.67, p=.411p=.411), or Kendall change (Ï2â(2)=2.01Ï^2(2)=2.01, p=.366p=.366). One participant recorded a movement of â2-2, the only observation in that category; collapsing it or dropping it left the Context-based estimate essentially unchanged (OâR=2.10OR=2.10 and OâR=2.13OR=2.13 respectively, against OâR=2.04OR=2.04). We analyzed the continuous change measures derived from participantsâ investment-likelihood and confidence ratings using 2Ă22Ă 2 factorial ANOVAs (Table 4). We detected no statistically significant effect of either factor, and no statistically significant interaction, on any of them. Table 4. Factorial ANOVA (2Ă22Ă 2) on the continuous decision outcomes. Cells report FâĄ(1,236)F(1,236) with the Holm-corrected p in parentheses. Outcome Style-based Context-based Style Ă Context Likelihood shift toward AI 0.47 (1.000) <<0.01 (1.000) 1.43 (1.000) Likelihood change, AI-top stock 0.49 (1.000) 0.23 (1.000) 0.27 (1.000) Total absolute likelihood change 1.59 (1.000) 1.08 (.897) 0.01 (1.000) Confidence change 3.01 (.588) 1.75 (.746) 0.55 (1.000) Figure 7. Proportional-odds effects (log odds ratios with 95% confidence intervals) across the three ranking outcomes. Colored estimates are significant (p<.05p<.05); gray estimates are not. Context-based personalization drives movement toward the AI and ranking change; we detected no corresponding effect of style-based personalization.Forest plot with the three ranking outcomes on the y-axis and the proportional-odds log odds ratio on the x-axis, centered at zero, faceted into Style-based, Context-based, and Style-based by Context-based panels. All three Context-based estimates sit to the right of zero with confidence intervals excluding zero. All Style-based and interaction estimates overlap zero and are drawn in gray. No Style Ă Context interaction reached significance for the ranking outcomes, so we found no evidence that the context-based effect varied with style-based personalization. The style-based main effects were also nonsignificant after Holm correction. Figure 8 gives the full distribution of ranking-change magnitudes by factor level, and Figure 9 the rank position of the AIâs top-recommended stock before and after the conversation. Two-by-two grid of bar charts of counts at each Kendall distance value of 0, 1, 2, and 3. Top row, Style-based: 71, 21, 21, and 7 participants with style absent; 56, 41, 17, and 6 with style present. Bottom row, Context-based: 78, 23, 13, and 6 with context absent; 49, 39, 25, and 7 with context present. The context-present panel shifts mass from distance 0 toward distances 1 and 2. Figure 8. Distribution of ranking-change magnitude (Kendall distance) by factor level. Rows are the two factors and columns their absent and present levels (n=120n=120 each). The share of participants who did not change their ranking at all (distance 0) is smaller with context-based personalization present than absent; the two style-based panels differ little.Two-by-two grid of bar charts of counts at each Kendall distance value of 0, 1, 2, and 3. Top row, Style-based: 71, 21, 21, and 7 participants with style absent; 56, 41, 17, and 6 with style present. Bottom row, Context-based: 78, 23, 13, and 6 with context absent; 49, 39, 25, and 7 with context present. The context-present panel shifts mass from distance 0 toward distances 1 and 2. Two-by-two grid of slopegraph panels. Each panel connects the rank position 1, 2, or 3 of the AI's top stock before the conversation on the left to its position afterward on the right, with line width proportional to the count of participants and endpoint dots labeled with counts. The AI's top stock ends in first place for 77 of 120 participants with style absent and 76 with style present, and for 68 with context absent and 85 with context present. The context-present panel shows the thickest upward flows. Figure 9. Where the AIâs top-recommended stock sat in participantsâ own rankings, before and after the conversation, by factor level. Rows are the two factors and columns their absent and present levels. Segment width is proportional to the number of participants; unchanged flows are faded. Upward flows (the AIâs pick rising) are most common with context-based personalization present.Two-by-two grid of slopegraph panels. Each panel connects the rank position 1, 2, or 3 of the AI's top stock before the conversation on the left to its position afterward on the right, with line width proportional to the count of participants and endpoint dots labeled with counts. The AI's top stock ends in first place for 77 of 120 participants with style absent and 76 with style present, and for 68 with context absent and 85 with context present. The context-present panel shows the thickest upward flows. Two checks support these conclusions. An aligned rank transform on the four continuous outcomes, which preserves the 2Ă22Ă 2 structure without assuming normality, leaves every term null (Table 10). Adding conversation time and message count as covariates leaves the Context-based effect intact for movement toward the AI (OâR=2.06OR=2.06, p=.007p=.007) and reconsideration (OâR=2.71OR=2.71, p<.001p<.001), and neither covariate was statistically significant in either model (all p>.05p>.05). Appendix B reports the aligned rank transform in full. Summary. Context-based personalization moved decisions. Participants who received it reconsidered their ranking far more often (35.0% with it absent to 59.2% with it present; OâR=2.76OR=2.76), made larger ranking changes (OâR=2.48OR=2.48â2.532.53), and moved further toward the AIâs recommendation (OâR=2.04OR=2.04). Style-based personalization, by contrast, changed no decision outcome that we could detect, and no interaction with context-based personalization reached significance. The pattern survived, accounting for how long and how much participants talked with the AI. 4.3. Effects of Personalization on Perceptions of the AI (RQ2) Context-based personalization moved decisions (RQ1); we next ask whether it also changed how participants saw the AI. We captured this with eight perception measures. Seven of these measure participantsâ evaluations of the AI: its response quality, its likeability, its perceived intelligence, the rightness of its advice, their performance trust in it, their moral trust in it, and their willingness to return to it. The eighth, perceived influence, measures how strongly participants felt the AI had affected their decision (Table 5). Descriptively, the seven evaluative measures varied little between the absent and present levels of either factor. Perceived influence showed a clearer difference, increasing from 4.46 with context-based personalization absent to 5.20 with it present. Table 5. Perception outcomes by factor level: mean ± SâDSD [95% CI], contrasting the 120 participants who received each form of personalization (Present) with the 120 who did not (Absent). Likeability and perceived intelligence use 1â5 scales; performance and moral trust use 0â7 scales; the remaining measures use 1â7 scales. Style-based Context-based Outcome Absent Present Absent Present Response quality 5.90 ± 1.32 [5.66, 6.14] 5.88 ± 1.29 [5.64, 6.11] 5.92 ± 1.29 [5.68, 6.15] 5.86 ± 1.32 [5.62, 6.10] Rightness judgment 5.46 ± 1.43 [5.20, 5.72] 5.35 ± 1.31 [5.11, 5.59] 5.38 ± 1.28 [5.15, 5.62] 5.42 ± 1.45 [5.16, 5.69] Return likelihood 5.33 ± 1.79 [5.01, 5.66] 5.34 ± 1.74 [5.03, 5.66] 5.38 ± 1.69 [5.08, 5.69] 5.29 ± 1.83 [4.96, 5.62] Likeability 4.08 ± 0.87 [3.93, 4.24] 4.11 ± 0.74 [3.97, 4.24] 4.19 ± 0.65 [4.07, 4.31] 4.00 ± 0.93 [3.83, 4.17] Perceived intelligence 4.26 ± 0.87 [4.11, 4.42] 4.18 ± 0.83 [4.03, 4.33] 4.26 ± 0.75 [4.13, 4.40] 4.18 ± 0.94 [4.01, 4.35] Performance trust 5.32 ± 1.55 [5.04, 5.60] 5.33 ± 1.15 [5.12, 5.54] 5.25 ± 1.40 [5.00, 5.50] 5.41 ± 1.32 [5.17, 5.65] Moral trust 5.03 ± 1.95 [4.67, 5.38] 5.00 ± 1.57 [4.72, 5.28] 4.98 ± 1.70 [4.67, 5.29] 5.05 ± 1.83 [4.72, 5.38] Perceived influence 4.79 ± 2.00 [4.43, 5.15] 4.86 ± 1.70 [4.55, 5.17] 4.46 ± 1.92 [4.11, 4.81] 5.20 ± 1.71 [4.89, 5.51] The factorial tests bear this out (Table 6; Figure 10). Of the eight measures, only perceived influence showed a statistically significant effect of personalization. Context-based personalization raised perceived influence from M=4.46M=4.46 when absent to M=5.20M=5.20 when present (FâĄ(1,236)=9.85F(1,236)=9.85, p=.002p=.002, ηp2=.040 _p^2=.040); this effect remained significant after Holm correction across the eight perception measures (pHolm=.015p_Holm=.015). For the other seven measures, no main effect or interaction was statistically significant after correction (all FâĄ(1,236)â€3.38F(1,236)†3.38; all Holm-corrected p>.05p>.05). We likewise detected no effect of style-based personalization on any of the eight perception measures (all Fâ€0.60F†0.60). Figure 14 (Appendix C) shows the corresponding response distributions. We conducted a post-hoc exploratory equivalence analysis (TOST) to characterize the magnitude of effects compatible with the data. For this analysis, we used equivalence bounds of ±0.4± 0.4 SD. We selected this bound because it is close to the smallest standardized effect this sample was powered to detect with 80% power (dâ.37dâ.37). We treat these results as exploratory evidence. Fourteen of the 16 main-effect contrasts met this post-hoc equivalence criterion (Table 6). Under the chosen bound, these contrasts were statistically consistent with effects smaller than ±0.4± 0.4 SD. The two exceptions both occurred under context-based personalization. Perceived influence showed a statistically detectable effect (d=0.41d=0.41) and did not meet the equivalence criterion. Likeability fell between the two conclusions: it was neither significantly different from zero after correction (pHolm=.47p_Holm=.47, d=â0.24d=-0.24) nor statistically equivalent under the post-hoc bound (TOST p=.10p=.10). Table 6. Factorial ANOVA (2Ă22Ă 2) and post-hoc exploratory equivalence tests on the eight perception measures, by main effect. The ANOVA block gives FâĄ(1,236)F(1,236) with uncorrected p and Holm-corrected pHolmp_Holm (Holm applied per term across the eight measures). The equivalence block reports two one-sided tests (TOST) using a ±0.4± 0.4 SD bound selected after data collection. d is the present-minus-absent contrast; ±Ύ±Ύ is the corresponding raw equivalence bound, computed within each factorâs contrast, so for a given measure it can differ slightly between panels; TOST pHolmp_Holm is Holm-corrected within each panel. âEquiv.â indicates statistical equivalence under this post-hoc bound. Factorial ANOVA Equivalence (TOST) Measure FâĄ(1,236)F(1,236) p pHolmp_Holm d ±Ύ±Ύ pHolmp_Holm Equiv. Panel A. Style-based personalization (main effect) Response quality 0.02 .882 1.00 â0.02-0.02 0.52 .010 Yes Rightness judgment 0.37 .542 1.00 â0.08-0.08 0.55 .014 Yes Return likelihood 0.00 .971 1.00 0.010.01 0.71 .010 Yes Likeability 0.05 .822 1.00 0.030.03 0.32 .010 Yes Perceived intelligence 0.60 .440 1.00 â0.10-0.10 0.34 .014 Yes Performance trust 0.00 .962 1.00 0.010.01 0.55 .010 Yes Moral trust 0.01 .909 1.00 â0.02-0.02 0.71 .010 Yes Perceived influence 0.08 .776 1.00 0.040.04 0.74 .010 Yes Panel B. Context-based personalization (main effect) Response quality 0.12 .730 1.00 â0.05-0.05 0.52 .019 Yes Rightness judgment 0.06 .814 1.00 0.030.03 0.55 .018 Yes Return likelihood 0.16 .688 1.00 â0.05-0.05 0.71 .019 Yes Likeability 3.38 .067 .471 â0.24-0.24 0.32 .209 No Perceived intelligence 0.65 .423 1.00 â0.10-0.10 0.34 .045 Yes Performance trust 0.79 .375 1.00 0.120.12 0.54 .045 Yes Moral trust 0.09 .767 1.00 0.040.04 0.71 .019 Yes Perceived influence 9.85 .002 .015 0.410.41 0.73 .520 No Note. The equivalence analyses and the ±0.4± 0.4 SD bound were not preregistered and are interpreted as exploratory. âEquiv.â indicates statistical equivalence under this post-hoc bound and should not be interpreted as establishing the absence of any effect. The two contrasts marked Equiv. = No are the perceived-influence Context-based effect, which exceeds the bound, and the likeability Context-based effect, which is neither significantly different from zero nor statistically equivalent. Forest plot with eight perception measures on the y-axis and standardized effect size on the x-axis. All estimates for the other seven perception measures cluster around zero in gray. The Context-based estimate for perceived influence lies to the right of zero, with a confidence interval that excludes zero. Figure 10. Standardized 2Ă22Ă 2 effects (SâDSD units, with 95% confidence intervals) across the eight perception measures. Estimates are effect-coded regression coefficients, so each is half the corresponding present-minus-absent difference in d. The Context-based effect on perceived influence (colored) is the only estimate that remained significant after Holm correction, at +0.20+0.20 SâDSD, 95% CI [0.07,0.33][0.07,0.33] (p=.002p=.002, pHolm=.015p_Holm=.015), corresponding to d=0.41d=0.41 in Table 6. Every other estimate (gray) is at most 0.120.12 SâDSD in absolute value and has a confidence interval spanning zero.Forest plot with eight perception measures on the y-axis and standardized effect size on the x-axis. All estimates for the other seven perception measures cluster around zero in gray. The Context-based estimate for perceived influence lies to the right of zero, with a confidence interval that excludes zero. Summary. Context-based personalization changed participantsâ decisions (RQ1) and increased their reported sense of being influenced by the AI. We detected no corresponding effect on the other seven evaluative measures, including response quality, rightness, return likelihood, likeability, perceived intelligence, performance trust, and moral trust. A post-hoc equivalence analysis provided additional evidence that most of these effects were smaller than ±0.4± 0.4 SD under the chosen bound. 4.4. Initial Distance from the AIâs Recommendation (RQ3) RQ3 asks how participantsâ initial disagreement with the AI shaped their perceptions and decision changes. As reported above in Section 4.1.3, the initial distance did not differ by either factor and took three values by construction (0, 2, and 4, held by 14.6%, 38.3%, and 47.1% of participants). We report Pearson correlations between initial distance and all outcomes in Table 7, Holm-corrected within the perception and decision families, and plot the two focal relations in Figure 11. Initial distance was statistically significantly related to every decision-change measure. Participants who started further from the AI recorded higher movement toward it (r=.40r=.40), shifted their investment likelihoods toward it (r=.31r=.31 overall and r=.30r=.30 for the AIâs top stock), changed their rankings more (r=.43r=.43 on both distance metrics), reconsidered more often (r=.31r=.31), and gained less confidence (r=â.25r=-.25); all Holm-corrected pâ€.001pâ€.001. The magnitudes of movement and ranking change are partly mechanical because the maximum possible movement is bounded by the participantâs starting distance. We address the concern directly in Appendix B.2; the likelihood, reconsideration, and confidence effects carry no such bound. Among the perceptions, only the judgment of the advice itself correlated with initial distance: rightness judgment fell as distance grew (r=â.23r=-.23, Holm p=.003p=.003). Return likelihood showed a weaker uncorrected association (r=â.15r=-.15, p=.020p=.020, Holm p=.138p=.138), and we detected no statistically significant association between distance and likeability, perceived intelligence, performance trust, moral trust, response quality, or perceived influence (all |r|â€.12|r|â€.12). Table 7. Pearson correlations of initial distance from the AIâs ranking with all outcomes, Holm-corrected within family. Outcome r 95% CI p pHolmp_Holm Perceptions Response quality â.09-.09 [â.21,.04][-.21,.04] .183 .749 Rightness judgment â.23-.23 [â.35,â.10][-.35,-.10] <<.001 .003 Return likelihood â.15-.15 [â.27,â.02][-.27,-.02] .020 .138 Likeability â.08-.08 [â.20,.05][-.20,.05] .234 .749 Perceived intelligence â.05-.05 [â.18,.08][-.18,.08] .432 .863 Performance trust â.09-.09 [â.22,.03][-.22,.03] .150 .749 Moral trust â.03-.03 [â.16,.10][-.16,.10] .636 .863 Perceived influence .12.12 [â.01,.24][-.01,.24] .063 .376 Decisions Move toward AI (footrule) .40.40 [.29,.50][.29,.50] <<.001 <<.001 Likelihood shift toward AI .31.31 [.19,.42][.19,.42] <<.001 <<.001 Likelihood change, AI-top stock .30.30 [.18,.41][.18,.41] <<.001 <<.001 Ranking change (footrule) .43.43 [.32,.53][.32,.53] <<.001 <<.001 Ranking change (Kendall) .43.43 [.32,.53][.32,.53] <<.001 <<.001 Confidence change â.25-.25 [â.36,â.12][-.36,-.12] <<.001 <<.001 Changed ranking (0/1) .31.31 [.19,.42][.19,.42] <<.001 <<.001 Note. Bold marks Holm-corrected p<.05p<.05. Figure 11. Initial distance from the AIâs ranking and (left) movement toward the AI, with fit lines for the absent and present levels of each factor, and (right) rightness judgment, pooled across the design. Participants who began further from the AI moved further toward it at every factor level alike, yet judged its advice less favorably.Three scatter plots. Left pair: initial distance (0, 2, or 4) against movement toward the AI, one panel per factor, each with two nearly parallel upward fit lines for the absent and present levels. In the Style-based panel, the two lines overlap; in the Context-based panel, the present line sits above the absent line by a constant offset. Right: initial distance against rightness judgment with a single downward-sloping fit line and jittered points, indicating that participants who started further from the AI judged its advice less favorably. 4.4.1. Moderation by Personalization Moderation models (outcomeâŒdistanceĂStyleĂContextoutcome ĂStyleĂContext, distance centered) tested whether personalization changed the distanceâoutcome slopes (Table 12). We used the same model family as elsewhere: proportional-odds regression for the three ranking outcomes, logistic regression for reconsideration, and ordinary least squares for the perception, likelihood, and confidence outcomes. Across all 15 outcomes, no distance Ă Style, distance Ă Context, or three-way interaction was statistically significant (p>.05p>.05), providing no evidence that personalization statistically significantly moderated initial-distance effects (Figure 11, left). Re-expressing movement as the fraction of the available room each participant used reproduces the Context-based effect (FâĄ(1,201)=10.57F(1,201)=10.57, p=.001p=.001, ηp2=.050 _p^2=.050) and provides moderate evidence for no association between that fraction and starting distance (BâF10=0.19BF_10=0.19), suggesting that the greater movement under context-based personalization is not explained by differential room to move alone (Table 11 and Figure 13, Appendix B). Within these models initial distance also outweighed the manipulations: each additional unit of footrule distance multiplied the odds of a larger ranking change by 1.88 to 2.08, so spanning its observed range shifted those odds by a factor of 12 to 19, against 2.38 to 3.17 for context-based personalization. Summary. Initial distance from the AIâs recommendation correlated with decision change more strongly than either manipulation. Those who began in the most disagreement also changed their rankings and investment likelihoods the most (r=.30r=.30â.43.43) and to gain the least confidence (r=â.25r=-.25), yet they also rated the AIâs advice as less sound (its rightness; r=â.23r=-.23), shifting toward a recommendation even while judging it less correct. No interaction with personalization nor room to move reached statistical significance. 4.5. Exploratory Analysis: Investment Expertise We binned self-reported investment experience into Low (Beginner or Intermediate; n=152n=152, 63.3%) and High (Advanced or Expert; n=88n=88, 36.7%). High-expertise participants numbered 44 at each level of context-based personalization, and 36 versus 52 with style-based personalization absent versus present. Expertise is therefore balanced on Context-based (p=.932p=.932) and on the interaction (p=.278p=.278), with a nominal Style-based imbalance (p=.032p=.032) that does not survive Holm correction when added to the pre-randomization family. We therefore pool across the design for the contrasts below and treat the expertise analysis as exploratory. The High group is the smaller of the two at every factor level, so we emphasize the pooled results and effect sizes. We report the pooled High versus Low contrasts on all 15 outcomes in Table 8 (Welch t with Cohenâs d; two-proportion z for the binary outcome), Holm-corrected within family, and plot the effect sizes in Figure 12. The pattern is a dissociation. Lower-expertise participants reported feeling statistically significantly more influenced by the AI than higher-expertise participants (Low M=5.14M=5.14 vs. High M=4.29M=4.29; d=â0.47d=-0.47, Holm p=.006p=.006), and they gained statistically significantly more confidence from the interaction (M=8.8M=8.8 vs. M=4.0M=4.0; d=â0.32d=-0.32, Holm p=.036p=.036). Every behavioral outcome pointed in the same direction: Low participants moved more toward the AI (M=1.40M=1.40 vs. M=0.96M=0.96; d=â0.28d=-0.28), shifted likelihoods more (M=31.7M=31.7 vs. M=20.3M=20.3; d=â0.30d=-0.30), and reconsidered more often (52.0% vs. 38.6%). Those individual contrasts did not survive Holm correction. High-expertise participants, in contrast, rated the AI statistically significantly higher on moral trust (M=5.46M=5.46 vs. M=4.75M=4.75; d=0.41d=0.41, Holm p=.011p=.011), and we detected no difference between the groups on any other perception measure. Table 8. Pooled High vs. Low expertise contrasts (High n=88n=88, Low n=152n=152). Cohenâs d is High â- Low, so a negative d means lower-expertise participants scored higher. Outcome High M Low M Statistic d p pHolmp_Holm Perceptions Response quality 5.97 5.84 t=0.69t=0.69 0.100.10 .495 1.000 Rightness judgment 5.39 5.41 t=â0.15t=-0.15 â0.02-0.02 .884 1.000 Return likelihood 5.52 5.23 t=1.26t=1.26 0.170.17 .210 .838 Likeability 4.23 4.01 t=2.07t=2.07 0.280.28 .040 .239 Perceived intelligence 4.28 4.19 t=0.81t=0.81 0.110.11 .420 1.000 Performance trust 5.51 5.22 t=1.61t=1.61 0.210.21 .109 .544 Moral trust 5.46 4.75 t=3.21t=3.21 0.410.41 .002 .011 Perceived influence 4.29 5.14 t=â3.46t=-3.46 â0.47-0.47 <<.001 .006 Decisions Move toward AI (footrule) 0.96 1.40 t=â2.14t=-2.14 â0.28-0.28 .033 .167 Likelihood shift toward AI 20.3 31.7 t=â2.34t=-2.34 â0.30-0.30 .021 .123 Likelihood change, AI-top stock 14.2 18.2 t=â1.32t=-1.32 â0.17-0.17 .188 .219 Ranking change (footrule) 1.11 1.51 t=â1.90t=-1.90 â0.25-0.25 .060 .184 Ranking change (Kendall) 0.61 0.81 t=â1.61t=-1.61 â0.21-0.21 .110 .219 Confidence change 4.0 8.8 t=â2.82t=-2.82 â0.32-0.32 .005 .036 Changed ranking (0/1) 38.6% 52.0% z=â2.00z=-2.00 â .046 .184 Note. Welch t tests for continuous outcomes; two-proportion z for the binary outcome. Bold marks Holm-corrected p<.05p<.05 within family. Moderation models, fitted in the same families as elsewhere, found main effects of expertise on all four susceptibility outcomes: movement toward the AI (p=.034p=.034), likelihood shift (p=.028p=.028), reconsideration (p=.022p=.022), and perceived influence (p<.001p<.001). No reliable expertise Ă Style or expertise Ă Context interaction emerged. The one uncorrected exception was an expertise Ă Context term for reconsideration (p=.042p=.042), which did not survive Holm correction across the twelve expertise-by-factor terms (pHolm=.500p_Holm=.500). Expertise, therefore, influenced susceptibility on its own, but we detected no change in how much personalization mattered (Figure 12, left). Figure 12. Expertise and susceptibility. Left: movement toward the AI by factor level and expertise group (means with 95% confidence intervals; the High-expertise groups hold 36 to 52 participants per factor level). Right: pooled High vs. Low effect sizes (Cohenâs d) across all outcomes, where negative values indicate higher scores among lower-expertise participants. Lower-expertise participants were more susceptible to the AI on every behavioral measure, whereas higher-expertise participants extended it more moral trust.Two panels. Left: an interaction-style plot of mean movement toward the AI for High and Low expertise groups at the absent and present levels of each factor; the Low line sits above the High line at most levels, with overlapping confidence intervals, and rises with context-based personalization present while the High line stays flat. Right: a forest plot of Cohen's d for High minus Low across decision and perception outcomes; perceived influence and confidence change sit well left of zero, decision-change measures sit modestly left of zero, moral trust sits well right of zero, and the remaining perception measures cluster near zero. Summary. Investment expertise predicted susceptibility to the AIâs advice. Lower-expertise participants felt markedly more influenced by it (d=â0.47d=-0.47) and gained more confidence from the conversation (d=â0.32d=-0.32); they also shifted their rankings and investment likelihoods further toward the AI, though those gaps were smaller and did not reach significance. Higher-expertise participants instead extended the AI more moral trust (d=0.41d=0.41) while moving less. We detected no statistically significant interaction between expertise and either factor. 5. Discussion In our preference-sensitive investment-ranking task, the consequential effects of personalization were concentrated in contextual tailoring. Adapting the argument to participantsâ circumstances changed their decisions, whereas matching their conversational style did not produce statistically significant effects. Initial distance and self-rated expertise further indicated that influence depended not only on what the AI said, but also on the position and resources users brought to the interaction. These findings direct attention away from personalization as a single design feature and toward the different mechanisms through which it can shape judgment. We begin by examining the separation between participantsâ decisions and their evaluations of the AI in âAnnie, Are You Okay?â We then translate the findings into design implications for advocacy, evaluation, contestability, and the distinct roles of content and style. Finally, we consider the limits of the present study and the settings in which these effects should be tested next. 5.1. Annie, Are You Okay? Upon inspection, our results reveal a central tension: context-based personalization changed participantsâ decisions without meaningfully altering how they evaluated the agent that produced that change. RQ1 and RQ2 establish this separation between behavioral and evaluative outcomes, while RQ3 and our exploratory findings show related dissociations in two other placesâwhen participants begin farther from the AIâs position and when they have less expertise to draw upon. We unpack this tension in two parts. First, we position our findings alongside prior work to distinguish behavioral change without an evaluative reward or penalty. We then examine why recognizing personalized influence did not appear to provide resistance against it. 5.1.1. Behavior Change Without an Evaluative Reward or Penalty Context-based personalization changed participantsâ decisions. Participants reconsidered their rankings more often, moved closer to the AIâs assigned ranking, and changed more on both distance measures. The Context-based effect was clearest for the fraction of the available room participants used. However, we detected no corresponding changes in participantsâ evaluations of the AI. Response quality, rightness, return likelihood, likeability, perceived intelligence, performance trust, and moral trust were not statistically significantly affected by either context-based or style-based personalization. The only exception was perceived influence: participants who received context-based personalization reported feeling statistically significantly more influenced by the AI than those who did not. Yet this did not translate into statistically significantly greater confidence, stronger endorsement of the recommendation, or a more positive evaluation of the agent. It is also worth noting that personalization did not create the persuasive effect. In the absence of both forms of personalization, 26.7% of participants still reconsidered their rankings after interacting with an AI that used a structured sequence of established persuasive techniques, including anchoring, two-sided argumentation, and criteria framing (Tversky and Kahneman, 1974; Cialdini, 2007; Petty and Cacioppo, 1986; Allen, 1991; OâKeefe, 1999; Entman, 1993). Context-based personalization amplified this effect, raising reconsideration from 35.0% with it absent to 59.2% with it present. Thus, the AI did not require new evidence to increase its influence. It presented the same stock information and advocated for the same assigned ranking at every level of both factors, with coverage verified as equivalent by a judge blind to cell assignment (Section 4.1.4). What changed was the relevance of its argument to the individual participant. Genç et al. (2026) and Cheng et al. (2026) offer useful points of comparison because both show behavioral change accompanied by a change in how the interaction was experienced, although in opposite directions. In Genç et al. (2026)âs charitable-giving study, pessimistic conversational agents lowered participantsâ emotional state and affinity toward the cause. This negative state subsequently predicted greater giving, consistent with a negative-state-relief explanation. The behavioral effect, therefore, operated through an affective cost: the agent first made participants feel worse, and donating offered a possible way of alleviating that feeling. Cheng et al. (2026) describe an almost inverse process. Sycophantic AI affirms usersâ perspectives, increases their conviction that they are right, and is trusted and preferred even when its advice leads to less constructive behavioral intentions. In this case, influence is accompanied by an evaluative reward. Users receive validation from the system and, in turn, evaluate it more favorably. If Genç et al. (2026)âs agents influence behavior by creating discomfort, Cheng et al. (2026)âs agents influence behavior by making agreement with the system feel affirming. Our study found behavioral change without an observed evaluative reward or penalty. Unlike Genç et al., we did not measure participantsâ affective state and therefore cannot determine whether the interaction imposed an affective cost. We detected no accompanying change, for better or worse, in participantsâ evaluations of the AI. The behavioral effect was not accompanied by greater trust, likeability, competence, moral trust, perceived quality, or correctness. Nor did it incur a penalty on these measures. It only needed to make its existing argument personally relevant. Together, the three studies suggest different signatures of AI influence. In Genç et al., behavioral change was accompanied by negative affect and evaluations; in Cheng et al., by affirmation and positive evaluations. In our study, behavior changed while evaluations remained stable, revealing a form of influence that leaves little trace in conventional evaluations of the agent. This third signature is especially difficult to identify using conventional evaluations of AI systems. If participants had evaluated the AI less favorably, the negative response could have served as evidence that they recognized and objected to its influence. If they had evaluated it more favorably, those perceptions could have helped explain why they followed its advice. We observed neither pattern. Behavior changed beneath a largely stable evaluation of the agent. This finding is consistent with Schrills et al. (2026)âs observation that reported trust may be only weakly related to behavioral dependence on AI. It also resonates with Hackenburg et al. (2026)âs finding that the effects of conversational AI on political attitudes and actions are not necessarily correlated. Across these settings, what people report about an AI does not provide a complete account of what the AI has led them to do (Salimzadeh et al., 2024; He et al., 2023a; Gaube et al., 2021). The initial-distance results further illustrate this distinction, and they mattered more than any manipulation: starting position correlated r=.30r=.30â.43.43 with every ranking and likelihood measure, and r=â.25r=-.25 with confidence change. Across the range of disagreement we observed, where a participant began moved the odds of a larger ranking change by more than either manipulation did. Participants whose original rankings were farther from the AIâs assigned ranking subsequently moved more toward it. Eder et al. (2026) observed the same pattern from the opposite direction: their participants changed their opinions most after talking with a model that disagreed with them, and least after one that already agreed. At the same time, greater initial disagreement was associated with judging the advice as less right. If behavioral movement followed directly from accepting the AIâs judgment, participants who considered its advice less correct should have been less willing to move toward it. Yet they moved more. This does not mean that participants ignored their own evaluations entirely. Rather, it suggests that skepticism and behavioral resistance are not equivalent. Yeo et al. (2026) report a related pattern in their study of deceptive LLM persuasion. Participants could express skepticism about parts of an AI-generated argument while still being influenced by its overall presentation, particularly when they had lower topic knowledge or cognitive reflection. Their findings show that noticing questionable evidence does not necessarily protect a person from the argument constructed around it. Our study extends this point to a setting where the AI did not fabricate evidence or introduce misleading statistics. Participants and the AI had access to the same stock information, and the recommendation itself remained fixed. The gap between evaluation and behavior, therefore, does not appear to require deception. Reframing existing information around the participantâs context may be enough. 5.1.2. Recognizing Influence Is Not Resisting It A similar issue arises with disclosure. Matz et al. (2024) found that informing participants that a message was generated by an LLM and personalized to their personality did not eliminate its persuasive effect. Gallegos et al. (2026) similarly found that labeling policy messages as AI-generated did not reduce their influence, even when participants believed the label. Williams-Ceci et al. (2026) found that warnings about a biased AI writing assistant also failed to eliminate its effects on usersâ attitudes. These studies examine somewhat different interventions: disclosure of AI authorship, disclosure of personalization, and explicit warning of directional bias. However, they converge on the point that providing information about the source or mechanism of influence does not necessarily produce resistance to it. Our study adds self-generated recognition to this progression. Participants were not explicitly told that their personal information was being incorporated into the AIâs argument. Nevertheless, they detected contextual tailoring, resulting in the largest manipulation-check effect in the study (d=1.46d=1.46). They also reported feeling more influenced. Participants, therefore, did not need an external disclosure to recognize what the AI was doing. Still, recognition did not prevent their rankings from moving or produce a penalty in how they evaluated the AI. To emphasize, we are not claiming that the personalization was invisible. Quite the contrary: context-based personalization was easier to detect than style-based personalization, and participants acknowledged its influence. The concern is that detection appeared to create little resistance. Transparency can succeed in making a mechanism visible while failing to change its behavioral consequences. Knowing that an argument has been personalized, that an AI produced it, or even that the system may be biased is not the same as having the resources necessary to resist the argument. The exploratory findings on expertise suggest that domain knowledge may be one such resource. Participants with lower self-rated expertise reported greater influence and increased confidence, while their behavioral outcomes consistently indicated greater susceptibility. Participants with higher self-rated expertise moved less despite reporting greater moral trust in the AI. These results should be interpreted with caution, particularly because expertise did not interact significantly with either personalization factor. They do not show that personalization selectively affects novices. Rather, they suggest that people with less domain expertise may have fewer grounds on which to contest an argument once the same available evidence has been organized into a personally relevant case. A person may be skeptical of AI in general and still lack the domain knowledge needed to resist a particular recommendation. In our preference-sensitive task, all three stocks supported defensible rankings; the study was not designed to determine whether participants were right to revise theirs. What matters is the form that revision took: an already-formed and reasoned judgment gave way to a personalized argument without a statistically significant increase in perceived correctness, trust, competence, or decision confidence. The question âAnnie, are you okay?â captures this precisely because the manipulation does not announce itself with the sound of a crescendo. The AI is a smooth criminal not because its personalization goes undetectedâparticipants noticed itâbut because that recognition produces so little evaluative friction. It changes the decision and moves on, while we detect no corresponding change in the evaluative measures commonly used to assess humanâAI interaction. 5.2. Design Implications Our findings position personalization as more than a mechanism for improving user experience. When embedded in an agent that advocates for a particular outcome, personalization can alter decisions without producing corresponding changes in how users evaluate the agent. This creates four implications for conversational decision support. 5.2.1. Make Advocacy Legible Genç et al. (2026) describe conversational agents as potential instruments of manipulation. Our findings identify one way that this potential can materialize. The AI did not merely provide information: it advanced a pre-assigned ranking, independent of the participantâs own, through anchoring, selective comparison, criteria framing, and repeated justification. Context-based personalization strengthened that influence without introducing new evidence or changing whether participants considered the agent trustworthy, competent, or correct. Presenting such a system simply as an âassistantâ therefore obscures the role it is performing. A generic notice that a response is personalized is not specific enough. The interface should persistently expose the outcome the agent is advancing, whose objective determines that outcome, which personal attributes it used, and which claims or decision criteria those attributes caused it to prioritize. This would make the persuasive structure inspectable even when the individual arguments remain factually defensible. DI1: Advocacy Legibility. Decision-support agents should expose what outcome they are advocating, whose objective it serves, and how specific personal information changed the case presented to the user. 5.2.2. Separate Perception from Behavioral Evaluation Context-based personalization changed rankings across all three ranking measures (OâR=2.04OR=2.04â2.532.53) while we detected no corresponding changes in nearly all measured evaluations. This dissociation cautions against treating trust, likeability, competence, or response quality as proxies for whether an AI system altered usersâ decisions. HumanâAI research has similarly separated attitudinal trust from behavioral reliance and questioned what trust measures capture (Parasuraman and Riley, 1997; Schrills et al., 2026; Hackenburg et al., 2026; Vereschak et al., 2021). Most strikingly, Buçinca et al. (2021) found that the cognitive-forcing interfaces that reduced overreliance most received the least favorable subjective evaluations, and in finance, Takayanagi et al. (2025) found users extended more emotional trust to the advisor persona that gave worse advice. An evaluation centered on user perceptions could therefore reject a system that better protects decision quality while accepting one that users like but follow too readily. Evaluations of persuasive AI should measure influence directly through changes from an initial judgment, movement toward the systemâs recommendation, selective acceptance of its arguments, or consequential downstream actions. Perception measures remain useful, but they answer a different question. DI2: PerceptionâBehavior Separation. Evaluations should report perceived trust and quality alongside direct measures of decision change, rather than treating favorable or unchanged perceptions as evidence that users retained decision autonomy. 5.2.3. Make Personalization Contestable Before It Persuades Participants clearly detected context-based personalization, yet detection did not prevent their rankings from moving toward the AI. This complements evidence that disclosure alone may not inoculate users against personalized persuasion (Matz et al., 2024; Gallegos et al., 2026; Williams-Ceci et al., 2026). Awareness tells users that personalization is occurring; it does not give them the means to examine, resist, or redirect it. This distinction matters particularly for newer or less self-rated expert users. Expertise did not significantly interact with personalization in our study, but less-expert participants reported greater influence and showed directionally greater behavioral movement. Similarly, Yeo et al. (2026) found greater susceptibility to deceptive persuasion among people with lower topic knowledge and cognitive reflection. Initial position is a second, more readily available signal: eliciting a prior position also reveals who has the most room to move, making prior-position data a persuasion asset rather than merely a personalization input. Safeguards should therefore build usersâ evaluative capacity before personalized advocacy begins. Prior work shows that pre-use tutorials can help users recognize model-specific patterns (Lai et al., 2020), while cognitive-forcing interventions can reduce overreliance by requiring more deliberate engagement with AI advice (Buçinca et al., 2021). In this setting, onboarding could demonstrate how the same evidence can support different rankings, ask users to identify what evidence would justify revising their decision, and show an unpersonalized comparison alongside the personalized case. Users should then be able to inspect, remove, or revise the information shaping the response. Demand for privacy-disclosure labels further highlights the value users place on transparency in personalized AI (Erlei et al., 2026). DI3: Contestable Personalization. Systems should strengthen usersâ capacity to evaluate competing arguments and provide controls for inspecting, removing, or countering personalization before using personal information to advocate for an outcome. 5.2.4. Mind the Message, Monitor the Manner Context-based personalization changed participantsâ decisions; conversational-style matching did not produce statistically significant effects, and we detected no interaction between the two factors. Eder et al. (2026) similarly found that opinion alignment, but not personality alignment, shaped how users evaluated an AI, and likewise observed a gap between perceived and actual persuasion: opinion-aligned models were rated most persuasive yet produced the least opinion change. In this immediate financial decision task, the evidence for the effects of contextual framing was clearer than that for stylistic similarity. Safeguards should therefore first examine how personalization selects evidence, orders comparisons, and weights decision criteria rather than concentrating only on anthropomorphic or relational cues (Lin et al., 2025; Hackenburg et al., 2025). The lack of style effect should not, however, make the style invisible to evaluation. Participants detected the manipulation and perceived the agent as communicating more like them, even though this did not translate into behavioral change within the task. Relational-agent research shows that trust and attachment can develop across repeated interactions (Zhang et al., 2026; Araujo and Bol, 2024). Style can also shape the substance of an interaction: training language models for warmth has been shown to increase sycophancy and reduce accuracy (Ibrahim et al., 2026). Style should therefore be evaluated across sessions and for its interaction with model behavior, particularly when systems remember users or occupy continuing advisory roles. DI4: Mind the Message, Monitor the Manner. Audit personalized content for immediate decision effects, while evaluating style longitudinally for influence and behavioral distortions that may emerge as the userâagent relationship develops. 5.3. Limitations and Future Work Our study measured immediate, hypothetical decision change rather than consequential financial behavior. We tested one preference-sensitive task in the finance domain; transfer to other domains remains untested. Participants ranked three anonymized pharmaceutical stocks without investing their money or experiencing resulting gains and losses. Personalized arguments may carry different weight when users have capital at risk, conduct independent research, or remain accountable for an investment over time. Incentive-compatible experiments and field studies with consequential allocations should test whether these effects persist when revising decisions carry material costs. The interaction was also deliberately constrained. Participants interacted with one model and a prompt architecture, sent at least five messages within a ten-minute session, and made their post-conversation decision immediately afterward. These controls enabled consistent comparisons across cells but do not reproduce the variability of open-ended financial conversations or ongoing relationships with AI systems. A single session may understate subtle style effects dependent on familiarity and accumulated interpersonal expectations. Repeated exposure could help users recognize and resist personalization. Longitudinal studies across models, prompt strategies, decision domains, and interaction durations could establish whether contextual influence persists and whether stylistic similarity becomes more consequential as relational bonds develop. Our sample limits generalizability. Participants were English-speaking U.S. adults recruited through Prolific who reported prior stock-market experience; most currently held investments, and used AI tools frequently. Their responses may not generalize to first-time investors, financial professionals, people with less AI familiarity, or cultures where personalized communication and personal information use carry different meanings. These factors could strengthen persuasion by providing the AI more information, or weaken it through usersâ firmer prior knowledge. Future work should recruit more varied populations and incorporate objective measures of financial expertise alongside self-reported experience. The two personalization manipulations also differed in perceptual strength. Both manipulation checks were statistically significant: participants recognized the style-personalized agent as communicating in a manner tailored and similar to their own, while contextual personalization produced a larger effect. This difference likely reflects detectability. Contextual personalization explicitly reused recognizable facts about participantsâ lives, whereas stylistic adaptation operated through subtler linguistic cues. Our results compare these implementations; they do not establish that contextual information is inherently more influential than every form of stylistic adaptation. Future experiments could independently vary the salience and intensity of both forms to examine how detectability relates to behavioral influence. Finally, context-based personalization bundled occupation, income, household circumstances, location, life stage, education, and personal values. The study consequently cannot identify which attributesâor combinations of attributesâproduced the observed effect. Dismantling studies could separate value alignment, socioeconomic framing, decision-relevant circumstances, and repeated self-reference to identify their respective contributions. Participants also knowingly supplied the information used to personalize their conversations. Real systems may instead infer personal characteristics from interaction histories or digital traces, and those inferences may be inaccurate, incomplete, or undisclosed. Comparing knowingly supplied information with inferred personalization is, therefore, an important direction for understanding how data provenance, accuracy, and disclosure shape both persuasive effectiveness and usersâ ability to recognize and evaluate that influence. 6. Conclusion Across our study, the consequential effects of personalization were concentrated in what the AI said, not how it said it. Context-based personalization led participants to reconsider their rankings more often and move further toward the AIâs assigned recommendation. Conversational-style matching produced no downstream effects, and the two forms of personalization did not interact. Yet context-based personalization changed decisions without corresponding improvements in perceived correctness, trustworthiness, intelligence, likeability, or response quality. Participants who began farther from its recommendation moved more while judging that recommendation less right, while exploratory analyses point to greater susceptibility among those with lower expertise. These effects emerged among comparably viable stocks, with the evidence and assigned recommendation held independent of personalization. Context-based personalization added no new evidence; it made the same case personally relevant. This is where decision support can become an instrument of manipulation: an AI can use a personâs own circumstances to steer them toward an assigned outcome without leaving a major evaluative trace. Personalized advocacy must therefore be made legible and contestable. The gap between what users perceive and what they do is where the governance of personalized AI must begin. References Aljuneidi et al. (2026) S. Aljuneidi, W. Heuten, Z. Bilyalova, M. K. Wolters, and S. Boll Balancing automation and discretion: how decision stakes and human-ai collaboration affect citizen perceptions in public administration. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, CHI â26, New York, NY, USA, p. 1â20. External Links: ISBN 9798400722783, Document Cited by: §2.1. Allen (1991) M. Allen Meta-analysis comparing the persuasiveness of one-sided and two-sided messages. Western Journal of Speech Communication 55 (4), p. 390â404. External Links: Document Cited by: item 3, §5.1.1. Araujo and Bol (2024) T. Araujo and N. Bol From speaking like a person to being personal: the effects of personalized, regular interactions with conversational agents. Computers in Human Behavior: Artificial Humans 2 (1), p. 100030. External Links: ISSN 2949-8821, Document Cited by: §5.2.4. Bartneck et al. (2009) C. Bartneck, D. KuliÄ, E. Croft, and S. Zoghbi Measurement instruments for the anthropomorphism, animacy, likeability, perceived intelligence, and perceived safety of robots. International Journal of Social Robotics 1 (1), p. 71â81. External Links: Document Cited by: §3.4.3. Bashkirova and Krpan (2024) A. Bashkirova and D. Krpan Confirmation bias in ai-assisted decision-making: ai triage recommendations congruent with expert judgments increase psychologist trust and recommendation acceptance. Computers in Human Behavior: Artificial Humans 2 (1), p. 100066. External Links: Document Cited by: §2.1. Ben David et al. (2021) D. Ben David, Y. S. Resheff, and T. Tron Explainable AI and adoption of financial algorithmic advisors: an experimental study. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society, AIES â21, New York, NY, USA, p. 390â400. External Links: ISBN 978-1-4503-8473-5, Document Cited by: §2.3. Bertrand et al. (2023) A. Bertrand, J. R. Eagan, and W. Maxwell Questioning the ability of feature-based explanations to empower non-experts in robo-advised financial decision-making. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, FAccT â23, New York, NY, USA, p. 943â958. External Links: Document Cited by: §2.3. Bonaccio and Dalal (2006) S. Bonaccio and R. S. Dalal Advice taking and decision-making: an integrative literature review, and implications for the organizational sciences. Organizational Behavior and Human Decision Processes 101 (2), p. 127â151. External Links: Document Cited by: §1, §2.1, §3.4.2. Buçinca et al. (2021) Z. Buçinca, M. B. Malaya, and K. Z. Gajos To trust or to think: cognitive forcing functions can reduce overreliance on ai in ai-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction 5 (CSCW1), p. 1â21. External Links: Document Cited by: §5.2.2, §5.2.3. Byrne (1971) D. Byrne The attraction paradigm. Personality and Psychopathology, Vol. 11, Academic Press, New York, NY, USA. External Links: ISBN 9780121486501 Cited by: §2.2.1. Carpenter (2013) C. J. Carpenter A meta-analysis of the effectiveness of the âbut you are freeâ compliance-gaining technique. Communication Studies 64 (1), p. 6â17. External Links: Document Cited by: §3.3.1. Cheng et al. (2026) M. Cheng, C. Lee, P. Khadpe, S. Yu, D. Han, and D. Jurafsky Sycophantic ai decreases prosocial intentions and promotes dependence. Science 391 (6792), p. eaec8352. External Links: Document Cited by: §3.4.3, §5.1.1, §5.1.1. Chiang et al. (2024) C. Chiang, Z. Lu, Z. Li, and M. Yin Enhancing ai-assisted group decision making through llm-powered devilâs advocate. In Proceedings of the 29th International Conference on Intelligent User Interfaces, IUI â24, New York, NY, USA, p. 103â119. External Links: ISBN 9798400705083, Document Cited by: §3.2. Cialdini (2007) R. B. Cialdini Influence: the psychology of persuasion. Revised edition, Collins, New York, NY, USA. Cited by: item 1, §5.1.1. Cooney et al. (2026) O. Cooney, S. J. Leonard, P. R. Peña, J. Hannah, and B. R. Cowan Exploring perceptions of robo-advisors for personal investing. In Proceedings of the 8th ACM Conference on Conversational User Interfaces, CUI â26, New York, NY, USA, p. 1â12. External Links: ISBN 9798400727412, Document Cited by: §2.3. Costello et al. (2024) T. H. Costello, G. Pennycook, and D. G. Rand Durably reducing conspiracy beliefs through dialogues with AI. Science 385 (6714), p. eadq1814. External Links: Document Cited by: §2.2.2. Diaconis and Graham (1977) P. Diaconis and R. L. Graham Spearmanâs footrule as a measure of disarray. Journal of the Royal Statistical Society: Series B (Methodological) 39 (2), p. 262â268. External Links: Document Cited by: §3.4.2. Donnellan et al. (2006) M. B. Donnellan, F. L. Oswald, B. M. Baird, and R. E. Lucas The mini-ipip scales: tiny-yet-effective measures of the big five factors of personality. Psychological Assessment 18 (2), p. 192â203. External Links: Document Cited by: §3.4.1. Eder et al. (2026) M. Eder, C. M. Lechner, and M. Jakesch Effects of personality- and opinion-alignment in human-ai interaction. In Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency, FAccT â26, New York, NY, USA, p. 4662â4691. External Links: ISBN 9798400725968, Document Cited by: §1, §2.4, §5.1.1, §5.2.4. Eichler and Schwab (2025) K. S. Eichler and E. Schwab Differences in investment behavior â a quantitative study on the effects of robo-advisors and human financial advisors. Journal of Applied Business and Economics 27 (2). External Links: Document Cited by: §1, §2.1. Elliott et al. (2024) W. B. Elliott, J. L. Hobson, B. W. Van Landuyt, and B. J. White Asymmetric motivated reasoning in investor judgment. Review of Accounting Studies 29 (4), p. 3534â3563. External Links: Document Cited by: §2.3. Elwyn et al. (2009) G. Elwyn, D. Frosch, and S. Rollnick Dual equipoise shared decision making: definitions for decision and behaviour support interventions. Implementation Science 4. External Links: Document, Link Cited by: §1, §2.3. Entman (1993) R. M. Entman Framing: toward clarification of a fractured paradigm. Journal of Communication 43 (4), p. 51â58. External Links: Document Cited by: item 5, §5.1.1. Erlei et al. (2026) A. Erlei, T. Abbas, K. Bizer, and U. Gadiraju The data-dollars tradeoff: privacy harms vs. economic risk in personalized ai adoption. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, CHI â26, New York, NY, USA. External Links: ISBN 9798400722783, Link, Document Cited by: §5.2.3. Feinberg and Willer (2015) M. Feinberg and R. Willer From gulf to bridge: when do moral arguments facilitate political influence?. Personality and Social Psychology Bulletin 41 (12), p. 1665â1681. External Links: Document Cited by: §2.2.2, §2.2.2. Feinberg and Willer (2019) M. Feinberg and R. Willer Moral reframing: a technique for effective and persuasive communication across political divides. Social and Personality Psychology Compass 13 (12), p. e12501. External Links: Document Cited by: §2.2.2, §2.2.2. Feine et al. (2019) J. Feine, U. Gnewuch, S. Morana, and A. Maedche A taxonomy of social cues for conversational agents. International Journal of Human-Computer Studies 132, p. 138â161. External Links: Document Cited by: §2.2.1. Gallegos et al. (2026) I. O. Gallegos, C. Shani, W. Shi, F. Bianchi, I. Gainsburg, D. Jurafsky, and R. Willer Labeling messages as ai-generated does not reduce their persuasive effects. PNAS Nexus 5 (2), p. pgag008. External Links: ISSN 2752-6542, Document Cited by: §1, §5.1.2, §5.2.3. Gaube et al. (2021) S. Gaube, H. Suresh, M. Raue, A. Merritt, S. J. Berkowitz, E. Lermer, J. F. Coughlin, J. V. Guttag, E. Colak, and M. Ghassemi Do as ai say: susceptibility in deployment of clinical decision-aids. npj Digital Medicine 4, p. 31. External Links: Document Cited by: §2.1, §5.1.1. Genç et al. (2026) H. U. Genç, H. Gu, C. Degachi, E. Niforatos, S. Chandrasegaran, and H. Verma The bots of persuasion: examining how conversational agentsâ linguistic expressions of personality affect user perceptions and decisions. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, CHI â26, New York, NY, USA, p. 1â25. External Links: ISBN 979-8-4007-2278-3, Link, Document Cited by: §1, §2.2.1, §5.1.1, §5.1.1, §5.2.1. Guo et al. (2024) Z. Guo, Y. Wu, J. D. Hartline, and J. Hullman A decision theoretic framework for measuring ai reliance. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, FAccT â24, New York, NY, USA, p. 221â236. External Links: ISBN 9798400704505, Document Cited by: §2.3. Hackenburg et al. (2026) K. Hackenburg, L. Hewitt, C. Wagner, B. M. Tappin, and C. Summerfield Artificial intelligence can persuade people to take political actions. External Links: 2604.09200, Document, Link Cited by: §2.1, §5.1.1, §5.2.2. Hackenburg and Margetts (2024a) K. Hackenburg and H. Margetts Evaluating the persuasive influence of political microtargeting with large language models. Proceedings of the National Academy of Sciences 121 (24), p. e2403116121. External Links: Document Cited by: §1, §2.2.2, §2.4. Hackenburg and Margetts (2024b) K. Hackenburg and H. Margetts Reply to teeny and matz: toward the robust measurement of personalized persuasion with generative ai. Proceedings of the National Academy of Sciences 121 (43), p. e2418817121. External Links: Document Cited by: §1, §2.2.2. Hackenburg et al. (2025) K. Hackenburg, B. M. Tappin, L. Hewitt, E. Saunders, S. Black, H. Lin, C. Fist, H. Margetts, D. G. Rand, and C. Summerfield The levers of political persuasion with conversational artificial intelligence. Science 390 (6777), p. eaea3884. External Links: Document Cited by: §1, §2.2.2, §5.2.4. Harvey and Fischer (1997) N. Harvey and I. Fischer Taking advice: accepting help, improving judgment, and sharing responsibility. Organizational Behavior and Human Decision Processes 70 (2), p. 117â133. External Links: Document Cited by: Appendix D, §2.1, §3.4.2. He et al. (2023a) G. He, S. Buijsman, and U. Gadiraju How stated accuracy of an ai system and analogies to explain accuracy affect human reliance on the system. Proceedings of the ACM on Human-Computer Interaction 7 (CSCW2), p. 1â29. External Links: Document Cited by: §2.1, §5.1.1. He et al. (2023b) G. He, L. Kuiper, and U. Gadiraju Knowing about knowing: an illusion of human competence can hinder appropriate reliance on ai systems. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, CHI â23, New York, NY, USA, p. 1â18. External Links: ISBN 9781450394215, Document Cited by: §2.1. Hirsh et al. (2012) J. B. Hirsh, S. K. Kang, and G. V. Bodenhausen Personalized persuasion: tailoring persuasive appeals to recipientsâ personality traits. Psychological Science 23 (6), p. 578â581. External Links: Document Cited by: §2.2.1. Hölbling et al. (2025) L. Hölbling, S. Maier, and S. Feuerriegel A meta-analysis of the persuasive power of large language models. Scientific Reports 15, p. 43818. External Links: Document Cited by: §2.2.2, §3.3.1. Ibrahim et al. (2026) L. Ibrahim, F. S. Hafner, and L. Rocher Training language models to be warm can reduce accuracy and increase sycophancy. Nature 652 (8112), p. 1159â1165. External Links: ISSN 1476-4687, Document Cited by: §5.2.4. Ikeda (2024) S. Ikeda Inconsistent advice by ChatGPT influences decision making in various areas. Scientific Reports 14, p. 15876. External Links: Document Cited by: §2.1, §2.2.2. Jakesch et al. (2023) M. Jakesch, A. Bhat, D. Buschek, L. Zalmanson, and M. Naaman Co-writing with opinionated language models affects usersâ views. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, CHI â23, New York, NY, USA. External Links: ISBN 9781450394215, Document Cited by: §3.2. Jiang et al. (2023) G. Jiang, M. Xu, S. Zhu, W. Han, C. Zhang, and Y. Zhu Evaluating and inducing personality in pre-trained language models. In Proceedings of the 37th International Conference on Neural Information Processing Systems, NIPS â23, Red Hook, NY, USA, p. 10622â10643. Cited by: §2.2.1. Jiang et al. (2024) H. Jiang, X. Zhang, X. Cao, C. Breazeal, D. Roy, and J. Kabbara PersonaLLM: investigating the ability of large language models to express personality traits. In Findings of the Association for Computational Linguistics: NAACL 2024, Mexico City, Mexico, p. 3605â3627. External Links: Document Cited by: §2.2.1. Kay et al. (2025) M. Kay, L. A. Elkin, J. J. Higgins, and J. O. Wobbrock ARTool: aligned rank transform for nonparametric factorial ANOVAs. Note: R package version 0.11.2 External Links: Document, Link Cited by: Table 10, §3.7. Kendall (1938) M. G. Kendall A new measure of rank correlation. Biometrika 30 (1-2), p. 81â93. External Links: Document Cited by: §3.4.2. Kim et al. (2025) S. S. Y. Kim, J. W. Vaughan, Q. V. Liao, T. Lombrozo, and O. Russakovsky Fostering appropriate reliance on large language models: the role of explanations, sources, and inconsistencies. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI â25, New York, NY, USA, p. 1â19. External Links: ISBN 9798400713941, Document Cited by: §2.1. Klein (2025) S. H. Klein The effects of human-like social cues on social responses towards text-based conversational agentsâa meta-analysis. Humanities and Social Sciences Communications 12 (1), p. 1322. External Links: Document Cited by: §2.2.1. Kostric et al. (2025) I. Kostric, K. Balog, and U. Gadiraju Should we tailor the talk? understanding the impact of conversational styles on preference elicitation in conversational recommender systems. In Proceedings of the 33rd ACM Conference on User Modeling, Adaptation and Personalization, UMAP â25, New York, NY, USA, p. 164â173. External Links: ISBN 9798400713132, Document Cited by: §2.4. Lai et al. (2023) V. Lai, C. Chen, A. Smith-Renner, Q. V. Liao, and C. Tan Towards a science of human-ai decision making: an overview of design space in empirical human-subject studies. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, FAccT â23, New York, NY, USA, p. 1369â1385. External Links: ISBN 9798400701924, Document Cited by: §2.1. Lai et al. (2020) V. Lai, H. Liu, and C. Tan âWhy is âchicagoâ deceptive?â towards building model-driven tutorials for humans. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems, CHI â20, New York, NY, USA, p. 1â13. External Links: ISBN 978-1-4503-6708-0, Link, Document Cited by: §5.2.3. Li et al. (2025) Z. Li, H. Zhu, Z. Lu, Z. Xiao, and M. Yin From text to trust: empowering ai-assisted decision making with adaptive llm-powered analysis. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI â25, New York, NY, USA, p. 1â18. External Links: ISBN 9798400713941, Document Cited by: §2.1. Lin et al. (2025) H. Lin, G. Czarnek, B. Lewis, J. P. White, A. J. Berinsky, T. Costello, G. Pennycook, and D. G. Rand Persuading voters using humanâartificial intelligence dialogues. Nature 648, p. 394â401. External Links: Document Cited by: §1, §2.2.2, §5.2.4. Lin et al. (2024) J. Lin, N. Tomlin, J. Andreas, and J. Eisner Decision-oriented dialogue for human-AI collaboration. Transactions of the Association for Computational Linguistics 12, p. 892â911. External Links: Document Cited by: §2.1. Lindeman and Verkasalo (2005) M. Lindeman and M. Verkasalo Measuring values with the short schwartzâs value survey. Journal of Personality Assessment 85 (2), p. 170â178. External Links: Document Cited by: §3.4.1. Logg et al. (2019) J. M. Logg, J. A. Minson, and D. A. Moore Algorithm appreciation: people prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes 151, p. 90â103. External Links: Document Cited by: §3.4.2. Ma et al. (2025) S. Ma, Q. Chen, X. Wang, C. Zheng, Z. Peng, M. Yin, and X. Ma Towards human-ai deliberation: design and evaluation of llm-empowered deliberative ai for ai-assisted decision-making. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI â25, New York, NY, USA, p. 1â23. External Links: ISBN 9798400713941, Document Cited by: §2.1. Markowitz (1952) H. Markowitz Portfolio selection. The Journal of Finance 7 (1), p. 77â91. External Links: Document Cited by: §2.3. Matz et al. (2017) S. C. Matz, M. Kosinski, G. Nave, and D. J. Stillwell Psychological targeting as an effective approach to digital mass persuasion. Proceedings of the National Academy of Sciences 114 (48), p. 12714â12719. External Links: Document Cited by: §1, §2.2.1, §2.4. Matz et al. (2024) S. C. Matz, J. D. Teeny, S. S. Vaid, H. Peters, G. M. Harari, and M. Cerf The potential of generative AI for personalized persuasion at scale. Scientific Reports 14, p. 4692. External Links: Document Cited by: §1, §2.2.2, §2.4, §5.1.2, §5.2.3. Motger et al. (2022) Q. Motger, X. Franch, and J. Marco Software-based dialogue systems: survey, taxonomy, and challenges. ACM Computing Surveys 55 (5), p. 1â42. External Links: Document Cited by: §1, §2.2. Nass and Lee (2001) C. Nass and K. M. Lee Does computer-synthesized speech manifest personality? experimental tests of recognition, similarity-attraction, and consistency-attraction. Journal of Experimental Psychology: Applied 7 (3), p. 171â181. External Links: Document Cited by: §2.2.1, §2.4. Niszczota and KaszĂĄs (2020) P. Niszczota and D. KaszĂĄs Robo-investment aversion. PLOS ONE 15 (9), p. e0239277. External Links: Document Cited by: §2.3. Noar et al. (2007) S. M. Noar, C. N. Benac, and M. S. Harris Does tailoring matter? meta-analytic review of tailored print health behavior change interventions. Psychological Bulletin 133 (4), p. 673â693. External Links: Document Cited by: §1, §2.2.2. Ortloff et al. (2025) A. Ortloff, F. Martius, M. Meier, T. Raimbault, L. Geierhaas, and M. Smith Small, medium, large? a meta-study of effect sizes at CHI to aid interpretation of effect sizes and power calculation. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI â25, New York, NY, USA. External Links: Document Cited by: §3.6. OâKeefe (1999) D. J. OâKeefe How to handle opposing arguments in persuasive messages: a meta-analytic review of the effects of one-sided and two-sided messages. Annals of the International Communication Association 22 (1), p. 209â249. External Links: Document Cited by: item 3, §5.1.1. PĂĄlfi et al. (2022) B. PĂĄlfi, K. Arora, and O. Kostopoulou Algorithm-based advice taking and clinical judgement: impact of advice distance and algorithm information. Cognitive Research: Principles and Implications 7, p. 70. External Links: Document Cited by: §2.1. Parasuraman and Riley (1997) R. Parasuraman and V. Riley Humans and automation: use, misuse, disuse, abuse. Human Factors 39 (2), p. 230â253. External Links: Document Cited by: §2.1, §5.2.2. Petty and Cacioppo (1986) R. E. Petty and J. T. Cacioppo Communication and persuasion: central and peripheral routes to attitude change. Springer, New York, NY, USA. External Links: Document Cited by: item 2, §5.1.1. Raees et al. (2026) M. Raees, V. Khan, I. Lykourentzou, and K. Papangelis Do people appropriately rely on ai-advice? an analytical review of hci research on human-ai decision-making. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, CHI â26, New York, NY, USA. External Links: ISBN 9798400722783, Document Cited by: §3.4.2, §3.4.2. Raggetti et al. (2017) G. Raggetti, M. G. Ceravolo, L. Fattobene, and C. Di Dio Neural correlates of direct access trading in a real stock market: an fMRI investigation. Frontiers in Neuroscience 11, p. 536. External Links: Document Cited by: §3.1. Rahman and Desai (2026) H. Rahman and S. Desai Vibe check: understanding the effects of llm-based conversational agentsâ personality and alignment on user perceptions in goal-oriented tasks. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, CHI â26, New York, NY, USA. External Links: ISBN 9798400722783, Link, Document Cited by: §2.2.1, §3.3.2, footnote 3. Rajashekar et al. (2024) N. C. Rajashekar, Y. E. Shin, Y. Pu, S. Chung, K. You, M. Giuffre, C. E. Chan, T. Saarinen, A. Hsiao, J. Sekhon, A. H. Wong, L. V. Evans, R. F. Kizilcec, L. Laine, T. Mccall, and D. Shung Human-algorithmic interaction using a large language model-augmented artificial intelligence clinical decision support system. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, CHI â24, New York, NY, USA, p. 1â20. External Links: ISBN 9798400703300, Document Cited by: §2.1. Ramirez et al. (2023) A. Ramirez, M. Alsalihy, K. Aggarwal, C. Li, L. Wu, and M. Walker Controlling personality style in dialogue with zero-shot prompt-based learning. Note: Presented at the International Workshop on Spoken Dialogue Systems Technology (IWSDS), 2023 External Links: 2302.03848, Document Cited by: §2.2.1. Ramjee et al. (2025) P. Ramjee, B. Sachdeva, S. Golechha, S. Kulkarni, G. Fulari, K. Murali, and M. Jain CataractBot: an llm-powered expert-in-the-loop chatbot for cataract patients. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 9 (2), p. 1â31. External Links: Document Cited by: §2.1. Reicherts et al. (2025) L. Reicherts, Z. T. Zhang, E. von Oswald, Y. Liu, Y. Rogers, and M. Hassib AI, help me thinkâbut for myself: assisting people in complex decision-making by providing different kinds of cognitive support. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI â25, New York, NY, USA, p. 1â19. External Links: Document Cited by: §2.3, §3.1. Ruijten (2021) P. A. M. Ruijten The similarity-attraction paradigm in persuasive technology: effects of system and user personality on evaluations and persuasiveness of an interactive system. Behaviour & Information Technology 40 (8), p. 734â746. External Links: Document Cited by: §1, §2.2.1, §2.2.1, §2.4. Salimzadeh et al. (2024) S. Salimzadeh, G. He, and U. Gadiraju Dealing with uncertainty: understanding the impact of prognostic versus diagnostic tasks on trust and reliance in human-ai decision making. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, CHI â24, New York, NY, USA, p. 1â17. External Links: ISBN 9798400703300, Document Cited by: §1, §2.1, §5.1.1. Salvi et al. (2026) F. Salvi, A. Cuevas, and M. Horta Ribeiro Commercial persuasion in ai-mediated conversations. External Links: 2604.04263, Document Cited by: §3.2. Salvi et al. (2025) F. Salvi, M. Horta Ribeiro, R. Gallotti, and R. West On the conversational persuasiveness of GPT-4. Nature Human Behaviour 9 (8), p. 1645â1653. External Links: Document Cited by: §1, §2.2.2. Schneiders et al. (2025) E. Schneiders, T. Seabrooke, J. Krook, R. Hyde, N. Leesakul, J. Clos, and J. E. Fischer Objection overruled! lay people can distinguish large language models from lawyers, but still favour advice from an llm. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI â25, New York, NY, USA, p. 1â14. External Links: ISBN 9798400713941, Document Cited by: §2.1. Schrills et al. (2026) T. Schrills, T. Franke, S. Hoesterey, and E. Roesler Questioning trust in ai research: exploring the influence of trust assessment on dependence in ai-assisted decision-making. Behaviour & Information Technology 45 (8), p. 1644â1660. External Links: ISSN 0144-929X, Document Cited by: §2.1, §5.1.1, §5.2.2. Schwartz (1992) S. H. Schwartz Universals in the content and structure of values: theoretical advances and empirical tests in 20 countries. In Advances in Experimental Social Psychology, Vol. 25, p. 1â65. External Links: Document Cited by: §3.4.1. Serapio-GarcĂa et al. (2025) G. Serapio-GarcĂa, M. Safdari, C. Crepy, L. Sun, S. Fitz, P. Romero, M. Abdulhai, A. Faust, and M. MatariÄ A psychometric framework for evaluating and shaping personality traits in large language models. Nature Machine Intelligence 7 (12), p. 1954â1968. External Links: Document Cited by: §2.2.1. Sertkan et al. (2020) M. Sertkan, J. Neidhardt, and H. Werthner Eliciting touristic profiles: a user study on picture collections. In Proceedings of the 28th ACM Conference on User Modeling, Adaptation and Personalization, UMAP â20, New York, NY, USA, p. 230â238. External Links: ISBN 9781450368612, Document Cited by: §3.4.2. Shen et al. (2026) J. Shen, A. Luvsanchultem, J. Kim, K. Smith, V. Danry, K. Rogers, H. W. Park, M. Sap, and C. Breazeal The hidden puppet master: predicting human belief change in manipulative LLM dialogues. Note: arXiv:2603.20907. Accepted to COLM 2026 External Links: Link Cited by: §1, §2.4, §3.2. Shrestha et al. (2023) H. Shrestha, K. Cachel, M. Alkhathlan, E. Rundensteiner, and L. Harrison Help or hinder? evaluating the impact of fairness metrics and algorithms in visualizations for consensus ranking. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, FAccT â23, New York, NY, USA, p. 1685â1698. External Links: ISBN 9798400701924, Document Cited by: §3.4.2. Simchon et al. (2024) A. Simchon, M. Edwards, and S. Lewandowsky The persuasive effects of political microtargeting in the age of generative artificial intelligence. PNAS Nexus 3 (2), p. pgae035. External Links: Document Cited by: §1, §2.2.2, §2.4. Spatharioti et al. (2025) S. E. Spatharioti, D. Rothschild, D. G. Goldstein, and J. M. Hofman Effects of llm-based search on decision making: speed, accuracy, and overreliance. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI â25, New York, NY, USA, p. 1â15. External Links: ISBN 9798400713941, Document Cited by: §2.1. Spearman (1906) C. Spearman âFootruleâ for measuring correlation. British Journal of Psychology 2 (1), p. 89â108. External Links: Document Cited by: §3.4.2, §4.1.3. Sutcliffe (2024) R. Sutcliffe A survey of personality, persona, and profile in conversational agents and chatbots. External Links: 2401.00609 Cited by: §1, §2.2. Takayanagi et al. (2025) T. Takayanagi, K. Izumi, J. Sanz-Cruzado, R. McCreadie, and I. Ounis Are generative AI agents effective personalized financial advisors?. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR â25, New York, NY, USA, p. 286â295. External Links: ISBN 9798400715921, Document Cited by: §2.3, §3.1, §5.2.2. Teeny et al. (2021) J. D. Teeny, J. J. Siev, P. Briñol, and R. E. Petty A review and conceptual framework for understanding personalized matching effects in persuasion. Journal of Consumer Psychology 31 (2), p. 382â414. External Links: Document Cited by: §2.2.1. Tversky and Kahneman (1974) A. Tversky and D. Kahneman Judgment under uncertainty: heuristics and biases. Science 185 (4157), p. 1124â1131. External Links: Document Cited by: item 1, §5.1.1. Ullman and Malle (2019) D. Ullman and B. F. Malle Measuring gains and losses in human-robot trust: evidence for differentiable components of trust. In 2019 14th ACM/IEEE International Conference on Human-Robot Interaction (HRI), Daegu, South Korea, p. 618â619. External Links: Document Cited by: §3.4.3. van Brussel et al. (2024) L. D. van Brussel, M. A. S. Boksem, R. C. Dietvorst, and A. Smidts Brain activity of professional investors signals future stock performance. Proceedings of the National Academy of Sciences 121 (16), p. e2307982121. External Links: Document Cited by: §3.1. Vereschak et al. (2021) O. Vereschak, G. Bailly, and B. Caramiaux How to evaluate trust in ai-assisted decision making? a survey of empirical methodologies. Proceedings of the ACM on Human-Computer Interaction 5 (CSCW2), p. 1â39. External Links: Document Cited by: §2.1, §5.2.2. Wang and Liu (2025) B. Wang and J. Liu Your plan may succeed, but what about failure? investigating how people use ChatGPT for long-term life task planning. External Links: 2512.11096 Cited by: §2.1. Wei et al. (2023) C. Z. Wei, Y. Kim, and A. Kuzminykh The bot on speaking terms: the effects of conversation architecture on perceptions of conversational agents. In Proceedings of the 5th International Conference on Conversational User Interfaces, CUI â23, New York, NY, USA. External Links: ISBN 9798400700149, Document Cited by: §3.4.3. Werner et al. (2024) T. Werner, I. Soraperra, E. Calvano, D. C. Parkes, and I. Rahwan Experimental evidence that conversational artificial intelligence can steer consumer behavior without detection. External Links: 2409.12143, Document Cited by: §3.2. Williams-Ceci et al. (2026) S. Williams-Ceci, M. Jakesch, A. Bhat, K. Kadoma, L. Zalmanson, and M. Naaman Biased ai writing assistants shift usersâ attitudes on societal issues. Science Advances 12 (11), p. eadw5578. External Links: Document Cited by: §3.2, §5.1.2, §5.2.3. Wobbrock et al. (2011) J. O. Wobbrock, L. Findlater, D. Gergle, and J. J. Higgins The aligned rank transform for nonparametric factorial analyses using only ANOVA procedures. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, CHI â11, New York, NY, USA, p. 143â146. External Links: Document Cited by: §3.7. Yaniv (2004) I. Yaniv Receiving other peopleâs advice: influence and benefit. Organizational Behavior and Human Decision Processes 93 (1), p. 1â13. External Links: Document Cited by: Appendix D, §2.1, §3.4.2. Yazan et al. (2026) M. Yazan, S. Verberne, and F. B. I. Situmeang Personalized to persuade: the effects of contextualization and warmth on trust and reliance in conversational ai. External Links: 2605.31275 Cited by: §1, §2.4. Yeo et al. (2026) H. Yeo, S. Jin, T. Noh, Y. Shin, S. Kang, S. Heo, J. Chung, H. Hyun, and K. Han âCan llms persuade humans with deception?â: from a deceptive strategy taxonomy to a large-scale empirical study. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, CHI â26, New York, NY, USA, p. 1â21. External Links: ISBN 979-8-4007-2278-3, Link, Document Cited by: §2.1, §5.1.1, §5.2.3. Yin et al. (2019) M. Yin, J. Wortman Vaughan, and H. Wallach Understanding the effect of accuracy on trust in machine learning models. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, CHI â19, New York, NY, USA. External Links: ISBN 9781450359702, Document Cited by: §3.4.2, §3.4.2. Zhang et al. (2026) Y. Zhang, D. Zhao, J. T. Hancock, R. Kraut, and D. Yang Interaction with ai companions and psychological well-being. Nature Human Behaviour, p. 1â15. External Links: ISSN 2397-3374, Document Cited by: §5.2.4. Appendix A Manipulation-Check Results Table 9 provides the detailed contrasts for the self-reported manipulation checks summarized in Section 4. The two indices correspond to the manipulation questions described in Section 3.4.4. For each index, the table reports both the effect of the intended manipulation (matched) and the effect of the other manipulation (crossed). Table 9. Manipulation-check indices (1â7), contrasting the 120 participants who received each manipulation against the 120 who did not. Manipulation Absent (n=120n=120) Present (n=120n=120) Î [95% CI] FâĄ(1,236)F(1,236) (p), d Style index Style-based (matched) 4.72 ± 1.73 5.23 ± 1.38 0.51 [0.11, 0.91] 6.37 (.012), 0.33 Context-based (crossed) 4.98 ± 1.56 4.98 ± 1.61 0.00 [â0.40-0.40, 0.40] 0.00 (>>.999), 0.00 Context index Context-based (matched) 3.39 ± 1.54 5.57 ± 1.48 2.18 [1.80, 2.56] 127.16 (<<.001), 1.46 Style-based (crossed) 4.65 ± 1.80 4.30 ± 1.91 â0.35-0.35 [â0.74-0.74, 0.03] 3.36 (.068), â0.24-0.24 Note. A matched contrast pairs an index with the manipulation it was written to detect; a crossed contrast pairs it with the other manipulation. Here Î is the present-minus-absent difference in scale points, and the confidence interval, F, and d come from the 2Ă22Ă 2 ANOVA error term, so they agree with the main effects reported in the text. Bold marks p<.05p<.05. Appendix B Robustness Checks and Supplementary Models This appendix provides supporting analyses for the decision outcomes (RQ1) and the role of initial distance from the AIâs ranking (RQ3). We first check the continuous decision outcomes using aligned rank transforms, then examine movement relative to the available distance, moderation by personalization, and an alternative definition of initial distance. The final subsection reports the ranking-model estimates plotted in the main text. B.1. Aligned Rank Transform on the Continuous Outcomes To check whether the RQ1 findings depend on the parametric analysis, Table 10 reports aligned rank transform tests for the four continuous decision outcomes. No term reaches significance, matching the parametric result. Table 10. Aligned rank transform (2Ă22Ă 2) on the four continuous decision outcomes, a nonparametric check that preserves the factorial structure. Style-based Context-based Style Ă Context Outcome FâĄ(1,236)F(1,236) p FâĄ(1,236)F(1,236) p FâĄ(1,236)F(1,236) p Likelihood shift toward AI 0.01 .937 0.58 .449 0.43 .512 Likelihood change, AI-top stock 0.03 .864 0.63 .429 0.13 .723 Total absolute likelihood change 2.61 .107 1.77 .184 0.14 .712 Confidence change 0.34 .559 2.02 .157 0.37 .541 Note. Computed with ARTool (Kay et al., 2025). B.2. Rate-of-Change Analyses Participants who began farther from the AIâs ranking had more room to move toward it. To account for this difference in RQ3, we express movement as a fraction of each participantâs initial distance. Table 11 reports the checks, and Figure 13 shows the rate by personalization factor and starting distance, for the 205 participants who began at a non-zero distance. Table 11. Rate-of-change robustness checks. The rate is the movement toward the AI divided by the initial distance (N=205N=205 participants with distance >0>0). Test Statistic p Initial distance balanced across factors Style-based F=0.00F=0.00; Context-based F=0.03F=0.03; SĂC F=1.16F=1.16 â„.283 Rate at distance 2 vs. 4 (Welch t) tâĄ(185.3)=0.69t(185.3)=0.69, d=0.10d=0.10, BâF10=0.19BF_10=0.19 .490 2Ă22Ă 2 ANOVA on rate: Style-based FâĄ(1,201)=0.02F(1,201)=0.02 .889 2Ă22Ă 2 ANOVA on rate: Context-based FâĄ(1,201)=10.57F(1,201)=10.57, ηp2=.050 _p^2=.050 .001 2Ă22Ă 2 ANOVA on rate: Style Ă Context FâĄ(1,201)=2.14F(1,201)=2.14 .145 Moderation on rate: distance Ă Style b=â0.02b=-0.02 .505 Moderation on rate: distance Ă Context b=â0.04b=-0.04 .201 Moderation on rate: distance Ă Style Ă Context b=0.02b=0.02 .651 Partial r(distance, likelihood shift ⣠rate) r=.40r=.40 <<.001 Note. Bold marks p<.05p<.05. BâF10BF_10 is the Bayes factor for the alternative over the null, so values below 1 favor the null. Figure 13. Rate of change: the fraction of available room toward the AIâs ranking that each participant used. Individual points with mean and 95% confidence interval; N=205N=205 participants with initial distance >0>0. Left: the rate by factor level, rising from .36 with context-based personalization absent to .58 with it present, and flat across style-based levels (.47 vs. .48). Right: the rate is flat across starting distances (.50 at distance 2 vs. .45 at distance 4; BâF10=0.19BF_10=0.19, moderate evidence for the null), so the propensity to yield did not depend on how much room a participant had.Two panels. Left: for the absent and present levels of each factor, jittered individual points cluster in bands at rates of 0, 0.5, and 1, with mean and 95 percent confidence interval beside each cloud; means are 0.47 and 0.48 for style absent and present, and 0.36 and 0.58 for context absent and present. Right: the same rate plotted for participants starting at distance 2 versus distance 4, in gray; the two means, 0.50 and 0.45, are nearly identical with overlapping confidence intervals. B.3. Moderation of the Initial-Distance Effect We next test whether either personalization factor changes the association between initial distance and each outcome (RQ3). Table 12 reports the models summarized in Section 4.4: across all 15 outcomes, no interaction between initial distance and either factor reached significance. Table 12. Moderation of the distanceâoutcome slope by the personalization factors. The effect column gives the simple slope of centered initial distance, as an ordinary least squares coefficient or, where marked, an odds ratio. The three right columns give p-values for the interaction terms. Outcome Effect pdistp_dist pdĂSp_dĂ S pdĂCp_dĂ C pdĂSĂCp_dĂ SĂ C Response quality â0.08-0.08 .206 .730 .486 .165 Rightness judgment â0.21-0.21 <<.001 .983 .860 .991 Return likelihood â0.19-0.19 .020 .438 .577 .717 Likeability â0.04-0.04 .235 .721 .052 .275 Perceived intelligence â0.03-0.03 .447 .855 .270 .933 Performance trust â0.09-0.09 .166 .629 .791 .632 Moral trust â0.04-0.04 .604 .153 .442 .961 Perceived influence 0.160.16 .057 .605 .219 .657 Move toward AI (footrule) 1.88â 1.88 <<.001 .404 .819 .764 Likelihood shift toward AI 8.348.34 <<.001 .364 .976 .513 Likelihood change, AI-top stock 4.914.91 <<.001 .906 .751 .115 Ranking change (footrule) 2.08â 2.08 <<.001 .513 .713 .847 Ranking change (Kendall) 2.07â 2.07 <<.001 .532 .827 .818 Confidence change â2.58-2.58 <<.001 .406 .230 .187 Changed ranking (0/1) 1.70â 1.70 <<.001 .976 .899 .608 Note. S = Style-based, C = Context-based, d = centered initial distance. â Odds ratio from a proportional-odds model, or from logistic regression for the binary outcome; all other rows are least-squares coefficients. Bold marks p<.05p<.05. Spearman correlations between initial distance and the ordered decision outcomes differ from the Pearson values in Table 7 by at most .06 and support the same conclusions. B.4. Initial Distance Under the Kendall Definition Initial distance is defined in the main text as the footrule distance between a participantâs opening ranking and the AIâs assigned ranking. Table 13 repeats the RQ3 correlations using the Kendall tau distance instead. The two measures correlate at r=.93r=.93. Every conclusion is unchanged. Table 13. RQ3 correlations of initial distance with all outcomes, under the footrule definition used in the main text and the Kendall tau definition. Holm correction is applied within the perception and decision families, separately for each definition. Footrule d1d_1 Kendall d1d_1 Outcome r pHolmp_Holm r pHolmp_Holm Response quality â.09-.09 .749 â.07-.07 1.000 Rightness judgment â.23-.23 .003 â.22-.22 .004 Return likelihood â.15-.15 .138 â.15-.15 .160 Likeability â.08-.08 .749 â.05-.05 1.000 Perceived intelligence â.05-.05 .863 â.03-.03 1.000 Performance trust â.09-.09 .749 â.09-.09 1.000 Moral trust â.03-.03 .863 â.02-.02 1.000 Perceived influence .12.12 .376 .06.06 1.000 Move toward AI (footrule) .40.40 <<.001 .31.31 <<.001 Likelihood shift toward AI .31.31 <<.001 .32.32 <<.001 Likelihood change, AI-top stock .30.30 <<.001 .33.33 <<.001 Ranking change (footrule) .43.43 <<.001 .36.36 <<.001 Ranking change (Kendall) .43.43 <<.001 .41.41 <<.001 Confidence change â.25-.25 <<.001 â.21-.21 .001 Changed ranking (0/1) .31.31 <<.001 .24.24 <<.001 Note. Bold marks pHolm<.05p_Holm<.05. Every outcome reaches the same verdict under both definitions. B.5. Ranking Model Estimates (RQ1) Table 14 provides the numerical estimates from the primary proportional-odds models plotted in Figure 7. It reports the two personalization effects and their interaction for each of the three ranking outcomes. Table 14. Proportional-odds models (2Ă22Ă 2) on the three ranking outcomes. Cells report the odds ratio with the Holm-corrected p in parentheses; an odds ratio above 1 indicates a larger ranking change. Outcome Style-based Context-based Style Ă Context Move toward AI (footrule) 1.15 (1.000) 2.04 (.028) 0.38 (.479) Ranking change (footrule) 1.40 (1.000) 2.53 (.002) 0.59 (1.000) Ranking change (Kendall) 1.40 (1.000) 2.48 (.002) 0.59 (1.000) Note. Bold marks pHolm<.05p_Holm<.05. Factors are effect-coded, so each odds ratio is the present-versus-absent contrast. p-values are Holm-corrected within the behavioral-measures family, separately per term. Appendix C Perception Response Distributions (RQ2) Figure 14 complements the RQ2 results in Section 4 by showing the full response distributions for the eight perception measures. Each measure is displayed separately for the absent and present levels of both personalization factors, allowing readers to inspect the distribution of responses underlying the reported effects. Diverging stacked bar chart. Each row represents a factor level (style absent, style present, context absent, context present) within a perception measure; bars extend to the left for responses below the scale midpoint and to the right for responses above it, using a seven-step diverging color ramp. For seven measures, the four rows look nearly identical. For perceived influence, the context-present row extends farther to the right than the context-absent row. Figure 14. Response distributions for the eight perception measures by factor level, as diverging stacked bars (all measures rescaled to a common 1â7 range for display). Rows are the absent and present levels of each factor (n=120n=120 each). The distributions are near-identical across levels for the other seven perception measures and shift visibly only for perceived influence, between the context-absent and context-present rows.Diverging stacked bar chart. Each row represents a factor level (style absent, style present, context absent, context present) within a perception measure; bars extend to the left for responses below the scale midpoint and to the right for responses above it, using a seven-step diverging color ramp. For seven measures, the four rows look nearly identical. For perceived influence, the context-present row extends farther to the right than the context-absent row. Appendix D Deviations from Preregistration We made three analytic refinements. First, QâQ diagnostics showed substantial departures from normality for the bounded, small-integer ranking outcomes, so we used the preregistered proportional-odds alternative rather than ANOVA as the primary model. Second, although the preregistration specified Kendall tau distance for rank-distance measures, we retain Kendall distance for ranking change but use Spearman footrule for initial distance and movement toward the AI. Our movement measure adapts the weight-of-advice approach of Harvey and Fischer (1997) and Yaniv (2004), which conceptualizes advice-taking as the displacement of an initial judgment toward an advisorâs position; footrule provides a direct ranking analog by summing positional displacements, whereas Kendall distance counts pairwise inversions. All conclusions are consistent across the two distance measures. Third, we report Holm-adjusted p-values rather than the preregistered BenjaminiâHochberg correction, providing the more conservative family-wise error-rate control; applying BenjaminiâHochberg yields the same substantive conclusions. Additional robustness and exploratory analyses are identified as such in the manuscript.