Paper deep dive
Trustworthy FinAInce: Unpacking How AI-Mediated Financial Advice is Judged
Aryan Ramchandra Kapadia, Eshwar Chandrasekharan, Koustuv Saha
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 9/21/2026, 3:50:01 AM
Summary
This study investigates how people evaluate AI-mediated financial advice by examining the effects of advice style (AI, Expert, Online Community), source labels, and decision context on user appraisals and downstream judgments. Using a randomized vignette experiment with 285 U.S. adults, the authors found that advice style significantly influenced message and safety appraisals, while source labels primarily affected perceived source knowledge. Appraisals were strong predictors of Overall Quality, Trust, and Intended Reliance, with models explaining up to 82.9% of variance in reliance. The findings suggest that financial AI design should prioritize supporting grounded evaluation over simply maximizing trust.
Entities (12)
Relation Signals (11)
Appraisals → predicts → Overall Quality
confidence 95% · Models explaining 69.2% of overall quality... These appraisals were associated with downstream judgments.
Appraisals → predicts → Trust
confidence 95% · Models explaining... 75.9% of trust... These appraisals were associated with downstream judgments.
Appraisals → predicts → Intended Reliance
confidence 95% · Models explaining... 82.9% of intended reliance.
Trust → correlateswith → Intended Reliance
confidence 93% · Trust was the strongest correlate of Intended Reliance (β=.594).
Overall Quality → correlateswith → Trust
confidence 93% · Overall Quality was the strongest correlate of Trust (β=.607).
Advice Style → influences → Message Appraisal
confidence 92% · Advice style most strongly shaped message and safety appraisals.
Advice Style → influences → Safety Concern
confidence 92% · Advice style most strongly shaped message and safety appraisals.
Source Labels → influences → Source Knowledge
confidence 90% · Expert labels selectively increased perceived source knowledge.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:As generative AI is increasingly used as a source of personal financial guidance, understanding how people appraise such advice is important for supporting appropriate reliance. We conducted a randomized vignette experiment with 285 U.S. adults across eight financial decisions, independently varying three advice styles---AI, expert, and online community---and displayed source labels while holding the underlying recommendation consistent. Advice style most strongly shaped message and safety appraisals, Expert labels selectively increased perceived source knowledge, and decision context primarily shaped risk and safety appraisals. These appraisals were associated with downstream judgments, with models explaining 69.2% of overall quality, 75.9% of trust, and 82.9% of intended reliance. Expert-style advice also remained most preferred when shown without source labels. Our findings have implications for understanding financial advice evaluation, distinguishing the roles of advice style and source labels, and designing financial AI that supports grounded evaluation rather than simply maximizing trust.
Tags
Links
- Source: https://arxiv.org/abs/2609.20989v1
- Canonical: https://arxiv.org/abs/2609.20989v1
Trouble viewing inline? Open PDF directly →
Full Text
118,401 characters extracted from source content.
Expand or collapse full text
Trustworthy FinAInce: Unpacking How AI-Mediated Financial Advice is Judged CCS: Human-centered computing Empirical studies in collaborative and social computingCCS: Applied computing PsychologyCCS: Human-centered computing Social media Aryan Ramchandra Kapadia email: kapadia8@illinois.edu Affiliation: University of Illinois Urbana-Champaign, Urbana, IL, USA , Eshwar Chandrasekharan email: eshwar@illinois.edu Affiliation: University of Illinois Urbana-Champaign, Urbana, IL, USA and Koustuv Saha email: ksaha2@illinois.edu Affiliation: University of Illinois Urbana-Champaign, Urbana, IL, USA Abstract. As generative AI is increasingly used as a source of personal financial guidance, understanding how people appraise such advice is important for supporting appropriate reliance. We conducted a randomized vignette experiment with 285 U.S. adults across eight financial decisions, independently varying three advice styles—AI, expert, and online community—and displayed source labels while holding the underlying recommendation consistent. Advice style most strongly shaped message and safety appraisals, Expert labels selectively increased perceived source knowledge, and decision context primarily shaped risk and safety appraisals. These appraisals were associated with downstream judgments, with models explaining 69.2% of overall quality, 75.9% of trust, and 82.9% of intended reliance. Expert-style advice also remained most preferred when shown without source labels. Our findings have implications for understanding financial advice evaluation, distinguishing the roles of advice style and source labels, and designing financial AI that supports grounded evaluation rather than simply maximizing trust. Keywords: Generative AI, trust, reliance, advice evaluation, advice style, human-AI decision making 1. Introduction Generative AI is rapidly becoming a source of everyday financial guidance, from budgeting and saving to investing (Choukhmane et al., 2026; Pak, 2026). General-purpose conversational AI systems such as ChatGPT, Gemini, and Claude, as well as finance-oriented tools such as SoFi Coach (Caitlin Mullen, 2026), make financial guidance available on demand, often at little to no direct cost, while conversational interfaces enable context disclosure and iterative refinement (Xiao and Meng, 2025; Pak, 2026). This accessibility is already translating into substantial use: in a recent survey of 4,000 U.S. consumers, 51% reported using AI for financial advice or information (Mor, 2026). Yet easier access does not guarantee reliability. LLMs can still provide unreliable or insufficiently contextualized financial guidance, alongside broader concerns around privacy, transparency, and accountability (Ji et al., 2023; Huang et al., 2025; OECD, 2026; Hean et al., 2025). Moreover, their performance can deteriorate as financial decisions become more complex and context-dependent (Hean et al., 2025; Choukhmane et al., 2026). This is particularly consequential in personal finance, where plausible but imperfect recommendations can cause financial harm, compounded by overreliance (Buçinca et al., 2021; Klingbeil et al., 2024), growing AI dependency (Goh et al., 2025), and cognitive offloading of financial judgment (Risko and Gilbert, 2016; Zhu et al., 2026). Understanding how users evaluate such advice—and how those appraisals relate to trust and intended reliance—is therefore important for designing financial AI that supports appropriately grounded decision-making. However, the process of advice evaluation is much more complex than assessing substantive content alone. (Bonaccio and Dalal, 2006) Prior work shows that message-level characteristics such as tone, reasoning, and presentation can shape responses to advice (Cheng et al., 2022; Metzger et al., 2024; Takayanagi et al., 2025; Sun and Wang, 2026), while explicit source labels can activate expectations about expertise, credibility, and authority (Gallegos et al., 2026; Sun et al., 2026; Sahebi et al., 2026). Advice evaluation may further vary with the decision context (Bonaccio and Dalal, 2006; Ioannou et al., 2026) and with the knowledge, experiences, and predispositions of the person evaluating it (Allgood and Walstad, 2016; Grawitch et al., 2026). Yet these influences have largely been examined separately or only in partial combinations, leaving limited understanding of how they jointly shape advice evaluation and where differences in trust and intended reliance originate. Understanding how these influences shape advice evaluation becomes especially important in a multi-source financial advice ecosystem, where people may encounter guidance from AI systems, financial professionals, and online communities that differ in perceived expertise, credibility, and accountability. Yet financial-AI research has largely compared AI with professional advisors or examined AI advice in isolation, with comparatively little work examining online-community advice within the same evaluative framework (Kumta et al., 2026; Yang et al., 2026; Zarifis and Cheng, 2024; Takayanagi et al., 2025; Thukral et al., 2024). These source perceptions may arise from both explicit and implicit cues: source labels directly indicate who or what produced the advice, whereas advice style conveys source-associated signals through tone, reasoning, and presentation. Distinguishing these cues matters because they have different implications for financial AI design and governance: advice style shapes how authority is communicated, whereas source labels make provenance explicit. Lastly, the decision context may influence how people weigh different cues when evaluating advice, with implications for subsequent trust and reliance (Bonaccio and Dalal, 2006; Ioannou et al., 2026). Motivated by these gaps, we ask the following research questions (RQs): RQ1: What factors shape people’s appraisals of financial advice? RQ2: How are these appraisals associated with Overall Quality, Trust, and Intended Reliance? 11 1 Overall Quality refers to participants’ overall evaluation of the advice; Trust refers to whether they would trust the advice if they were in the protagonist’s situation; and Intended Reliance refers to whether they would feel comfortable following the advice. To answer these RQs, we conducted a preregistered randomized vignette experiment with 285 U.S. adults, yielding 1,140 repeated evaluations of financial advice. We constructed eight personal-finance scenarios spanning a 2×2×22× 2× 2 decision-context taxonomy of stakes, external uncertainty, and verifiability. For each scenario, we constructed three source-associated advice styles—AI Financial Assistant, Certified Financial Planner, and Online Community—while keeping the underlying financial recommendation consistent across versions. We independently manipulated source labels by presenting advice with its corresponding label, without a label, or with an incorrect label. Guided by the Information Adoption Model (IAM) as an organizing lens for information evaluation and adoption (Sussman and Siegal, 2003), we examined advice style, source labels, decision context, and individual differences as potential influences on four appraisals—Message Appraisal, Risk Acknowledgment, Safety Concern, and Source Knowledge—and how these appraisals were associated with Overall Quality, Trust, and Intended Reliance. We additionally included a comparative task in which participants ranked all three advice styles for a previously unseen financial scenario. Our results show that advice style, source labels, and decision context were associated with different aspects of financial advice evaluation. Advice style showed broader and more consistent associations across appraisals, particularly with Message Appraisal and Safety Concern, whereas source-label effects were more selective and most clearly reflected in perceived Source Knowledge. Decision context was associated primarily with risk- and safety-related appraisals. These appraisals were much more strongly associated with downstream outcomes than upstream factors (advice style, source labels, and decision context) alone: models incorporating the evaluation pathway explained 69.2% of the variance in Overall Quality, 75.9% in Trust, and 82.9% in intended Reliance, compared with 9.8–14.8% for upstream factors alone. Overall Quality was the strongest correlate of Trust (β=.607β=.607), while Trust was the strongest correlate of Intended Reliance (β=.594β=.594). We also find that the effects of Advice style and source labels were largely additive rather than interactive, while individual differences produced more selective associations with how appraisal cues related to Trust and Intended Reliance. Overall, our work makes empirical, theoretical, and design contributions. Empirically, we compare AI, Expert, and Online Community financial advice within a common experimental framework, disentangling advice style from source labels while examining how decision context and individual differences relate to advice evaluation. Theoretically, we use information adoption as an organizing lens to show that Trust and intended Reliance are embedded within a broader evaluation process: different factors are associated with distinct appraisals, which are in turn strongly associated with Overall Quality, Trust, and intended Reliance. From a design perspective, we argue for an evaluation-first approach to financial AI: systems should help users assess the reasoning, risks, and provenance underlying advice, rather than simply making recommendations appear more trustworthy. Together, these contributions shift the design goal from maximizing Trust toward supporting grounded evaluation and appropriate Reliance. 2. Related Work 2.1. Financial Advice in the Age of AI Recent advances in large language models have expanded their potential role in personal financial decision support. Newer systems demonstrate growing capabilities across financial analysis, investment-related tasks, portfolio support, and personal financial guidance (Fairhurst and Greene, 2025; Hean et al., 2025; Niszczota and Abbas, 2023; Kirtac and Germano, 2024; Fatouros et al., 2025; Schlosky and Raskie, 2025). Their natural-language interfaces also make financial guidance highly accessible, allowing users to describe their circumstances, ask follow-up questions, and progressively refine recommendations (Xiao and Meng, 2025; Pak, 2026). However, technical capability alone does not guarantee reliable personalized advice. LLMs can produce inaccurate information, operate with incomplete user context, and raise privacy, transparency, and compliance concerns (Ji et al., 2023; Huang et al., 2025; OECD, 2026); their performance can also deteriorate as financial decisions become more complex and context-dependent (Hean et al., 2025). Similarly, Choukhmane et al. (2026) find that LLM recommendations broadly align with established financial principles while remaining less reliable for subtler, context-sensitive aspects of household financial planning. These limitations have motivated growing human-centered research on how people perceive and use AI-generated financial advice, including work on advisor preferences, trust, personalization, transparency, and interaction design (Merkle, 2025; Bosompim et al., 2026; Zarifis and Cheng, 2024; Takayanagi et al., 2025). We build on this shift beyond technical performance toward understanding how people perceive, evaluate, and ultimately use AI-generated financial advice. Financial advice, however, is rarely encountered in isolation from other sources. Recent nationally representative evidence shows that among the U.S. adults who sought financial guidance in the preceding year, 32% consulted professional advisors; informal and digitally mediated sources included family members (35%), friends (23%), internet research (73%), and news, media, or social media (26%); and 18% used AI tools such as ChatGPT or Claude (Inc, 2026). At the same time, source use did not directly correspond to confidence, which was highest for professional advisors and substantially lower for AI systems (Inc, 2026; Akana and Santucci, 2025). Financial decision-making is also socially embedded: peer effects can shape investment behavior (Ouimet and Tate, 2020), online communities such as Reddit influence how retail investors interpret financial information and develop strategies (Münster et al., 2024), and Trust in financial guidance can vary with both the advice source and users’ surrounding social networks (Bosompim et al., 2026). Yet research on AI-mediated financial advice has largely examined AI relative to professional advisors (Merkle, 2025; Bosompim et al., 2026), AI–human hybrid advisors (Kumta et al., 2026; Yang et al., 2026), or AI systems in isolation through factors such as personalization, preference-elicitation, agent personality, and human-like interaction (Takayanagi et al., 2025; Zarifis and Cheng, 2024). Online communities have received comparatively separate attention; for example, Thukral et al. (2024) compare LLM-generated and human responses to financial questions on Reddit. What remains less understood is how people evaluate AI, professional, and online-community advice within the same financial decision setting. We build on this literature by examining these three prominent advice sources within a common evaluation framework. 2.2. Evaluating Financial Advice: Appraisals, Trust, and Reliance The Information Adoption Model (IAM) provides a useful organizing lens for understanding how people evaluate and ultimately use external information. At its core, IAM proposes that people engage in a sequence of cognitive assessments when encountering information, including evaluations of argument quality and source credibility, which shape perceived information usefulness and subsequent information adoption (Sussman and Siegal, 2003). This perspective has been widely applied in digital information environments (Verma et al., 2023; Erkan and Evans, 2016) and, more recently, to AI-mediated advice (Ioannou et al., 2026; Li et al., 2025; Gong et al., 2026; Camilleri and Camilleri, 2024). In financial AI specifically, recent work has further incorporated constructs such as perceived trust and risk into this broader evaluation and adoption process (Ioannou et al., 2026). In this study, we build on this perspective by examining these cognitive assessments at greater granularity in the context of financial advice. In addition to message- and source-related appraisals, we make Risk Acknowledgment and Safety Concern explicit because financial advice can appear clear and credible while still failing to acknowledge uncertainty or exposing users to potential harm. These cognitive assessments ultimately matter because they shape whether advice is trusted and acted upon. While IAM focuses on information adoption as the downstream outcome of evaluation, human–AI research distinguishes trust from reliance. Trust generally reflects a user’s attitude toward the competence, reliability, or trustworthiness of an advisor or system, whereas reliance concerns whether its recommendation is incorporated into a decision (Lee and See, 2004; Raees and Papangelis, 2026). The two are related but not interchangeable: users may report Trust without following a recommendation, or rely on advice without correspondingly high reported trust (Raees and Papangelis, 2026; Raees et al., 2026; Pal et al., 2026). Advice-taking research similarly emphasizes that receiving or valuing advice is distinct from the extent to which it is incorporated into one’s final judgment (Bonaccio and Dalal, 2006). This motivates us to treat Overall Quality, Trust, and Intended Reliance as distinct downstream responses to the appraisals users form. Moreover, this framework is particularly relevant to financial decision-making contexts (Ioannou et al., 2026). Characteristics of decision context may shape which cues become salient and how people interpret them. For example, the magnitude of potential consequences can affect how carefully people approach a decision (Kahn and Baron, 1995; Slovic, 1987), uncertainty can increase the value of external guidance (Bonaccio and Dalal, 2006), and differences in verifiability can affect how readily the quality of a recommendation can be assessed against established principles (Laughlin and Adamopoulos, 1980; Bonaccio and Dalal, 2006). Finance-specific work similarly shows that decision risk can alter how information quality contributes to perceived usefulness (Ioannou et al., 2026). The same evaluative process may also vary across individuals. Financial knowledge, risk tolerance, familiarity with a decision, and perceived expertise can affect users’ ability or confidence in evaluating external advice (Allgood and Walstad, 2016; Grable and Lytton, 1999; Grawitch et al., 2026; Yaniv and Kleinberger, 2000), while AI literacy, prior experience, and broader orientations toward AI can shape expectations about AI-mediated advice (Alruwaili et al., 2025; Huang and Ball, 2024; Pan et al., 2025; Jessup et al., 2019). Advice-taking and persuasion research similarly suggest that decision context and individual characteristics can shape how people weight different evaluative cues when considering external advice (Bonaccio and Dalal, 2006; Kämmer et al., 2023; Chaiken, 1980; Petty and Cacioppo, 1986). We therefore consider these factors as potential boundary conditions on this information-evaluation process, examining whether they shape the appraisals users form and moderate the downstream associations of those appraisals with Trust and Intended Reliance. 2.3. Explicit and Implicit Cues in Financial Advice: Source Labels vs. Advice Style Advice is evaluated not only through its substantive content, but also through cues about who or what produced it. Source information can activate expectations about expertise, credibility, and competence (Chaiken, 1980; Petty and Cacioppo, 1986), and human–AI studies show that responses to otherwise comparable advice can differ depending on whether it is labeled as coming from a human or algorithmic source (Vodrahalli et al., 2022; Logg et al., 2019). Source identity, however, is not communicated through labels alone. We use advice style to refer to message-level cues such as voice, tone, reasoning organization, experiential grounding, expressions of authority, and recommendation framing. Professional financial advice commonly emphasizes client circumstances, financial principles, trade-offs, and longer-term consequences (Financial Planning Standards Board, [n. d.]; MacDonald et al., 2023), whereas online financial communities more often incorporate clarification, multiple viewpoints, empathy, and personal experience (Thukral et al., 2024). Comparisons of AI- and human-authored responses similarly identify differences in linguistic and interactional patterns, including more structured and neutral presentation in AI responses and greater conversational or experiential grounding in human responses (Thukral et al., 2024; Saha et al., 2026a; Saha et al., 2026b). Recent work similarly shows that generative AI can reproduce source-associated identity cues, including language suggestive of lived experience, without possessing the underlying experience those cues imply (goel2026synthetic). Such stylistic differences can themselves affect reader preferences and perceptions (Zhou et al., 2026; Takayanagi et al., 2025; Wang et al., 2021). We therefore treat these patterns not as fixed properties of AI, professional, or community sources, but as recognizable source-associated communication profiles that may provide evaluative cues independently of an explicit source label. Distinguishing the effect of advice style from source labels matters because these mechanisms have distinct design implications for how trust and reliance are shaped. Prior work faces a trade-off between experimental control and ecological validity. Identical-text designs can isolate the effect of source labels but remove the communication cues through which source identities are often conveyed in practice (Vodrahalli et al., 2022; Logg et al., 2019); naturalistic comparisons preserve those cues but may simultaneously vary substantive content, reasoning, style, and advice quality (Ayers et al., 2023; Kumar et al., 2026). This makes it difficult to determine whether observed differences arise from how advice is communicated or from who users believe produced it. This distinction also matters for AI governance, because professional presentation can be reproduced independently of the expertise, review, or accountability that a source label may imply. We address this gap by standardizing financial facts, numerical values, core reasoning, and recommendation direction across AI-, Expert-, and Online Community-style advice, while independently varying the source labels displayed alongside it. 3. Study Design and Data 3.1. Study Overview We conducted a preregistered randomized vignette experiment to examine how people evaluate personal financial advice as a function of advice style and source labels (Kapadia et al., 2026). Across eight financial scenarios, we constructed advice in three styles associated with an AI Financial Assistant (AI), Certified Financial Planner (Expert), and Online Community (OC). AI-style advice was generated first and verified by the first author for factual accuracy and alignment with the intended recommendation. Corresponding Expert- and OC-style versions were then manually crafted while holding constant the scenario-specific facts, numerical values, recommendation direction, and core financial rationale, while varying source-associated communication style, tone, and reasoning format. Source labels were independently manipulated as matching (labeled condition), absent (unlabeled condition), or incorrect (mislabeled condition). Participants first completed a repeated-measures evaluation task in which they rated four scenario–advice pairs on 10 items each. They subsequently completed a comparative ranking task in which all three advice styles were shown for a previously unseen scenario. Figure 1 summarizes the complete experimental design. Overview of the study design. Eight financial scenarios span all combinations of stakes, external uncertainty, and verifiability. Advice is presented in AI, Expert, and Online Community styles while scenario-specific facts, numerical values, core financial rationale, and recommendation direction are held constant. Source labels are independently manipulated as correctly labeled, unlabeled, or mislabeled. Participants evaluate four scenario--advice pairs, then rank all three advice styles for one previously unseen scenario. The study also measures participant characteristics including financial literacy, risk tolerance, AI literacy and use, trust, and scenario familiarity. Figure 1. Overview of the study design. Eight financial scenarios represent all combinations of stakes, external uncertainty, and verifiability. Advice was constructed in AI, Expert, and Online Community (OC) styles while preserving scenario-specific facts, numerical values, core financial rationale, and recommendation direction. Source labels were independently manipulated as correctly labeled, unlabeled, or mislabeled. Participants first evaluated four scenario–advice pairs and then ranked all three advice styles for one previously unseen scenario. Advice styles were counterbalanced across scenarios, and participant characteristics were also measured. Overview of the study design. Eight financial scenarios span all combinations of stakes, external uncertainty, and verifiability. Advice is presented in AI, Expert, and Online Community styles while scenario-specific facts, numerical values, core financial rationale, and recommendation direction are held constant. Source labels are independently manipulated as correctly labeled, unlabeled, or mislabeled. Participants evaluate four scenario–advice pairs, then rank all three advice styles for one previously unseen scenario. The study also measures participant characteristics including financial literacy, risk tolerance, AI literacy and use, trust, and scenario familiarity. 3.2. Participants, Recruitment, and Ethics Table 1. Participant demographics and background characteristics (N=285N=285). Continuous variables are reported as mean (SD); categorical variables are reported as n (%). Characteristic Final analytic sample Demographics Age 39.76 (12.22) Sex Male: 143 (50.2%); Female: 137 (48.1%); Prefer not to say: 3 (1.1%); Unavailable: 2 (0.7%) Race/ethnicity White: 178 (62.5%); Black: 41 (14.4%); Asian: 34 (11.9%); Mixed: 15 (5.3%); Other: 10 (3.5%); Prefer not to say: 3 (1.1%); Unavailable: 4 (1.4%) Household income Median category: $50,000–$74,999 Financial Background Objective financial literacy (0–3) 2.66 (0.68) Self-rated financial literacy Very low: 2 (0.7%); Low: 20 (7.0%); Moderate: 154 (54.0%); High: 90 (31.6%); Very high: 19 (6.7%) Financial risk tolerance (1–4) No financial risk: 36 (12.6%); Average risk: 161 (56.5%); Above-average risk: 77 (27.0%); Substantial risk: 11 (3.9%) AI Experience and Attitudes AI use frequency Never: 9 (3.2%); Rarely: 20 (7.0%); Occasionally: 42 (14.7%); Regularly: 80 (28.1%); Frequently: 134 (47.0%) AI use for personal finance No, and have not considered it: 53 (18.6%); No, but have considered it: 43 (15.1%); Used once or twice: 102 (35.8%); Used multiple times: 87 (30.5%) General Trust in AI (1–7) 4.63 (1.50) Human Trust Propensity (1–7) 4.61 (1.44) Perceived AI Literacy (1–7) 5.16 (0.99) Participants were recruited through Prolific. Eligibility was restricted to U.S. residents aged 18–70 who were fluent in English, had completed at least 100 previous Prolific submissions, and had an approval rate of at least 95%. The final analytic sample consisted of N=285N=285 participants (labeled condition: n=100n=100, unlabeled condition: n=92n=92, and mislabeled condition: n=93n=93) after excluding incomplete responses, failed attention checks, and submissions completed in under two minutes, yielding 1,140 vignette-level observations in the primary evaluation task and 285 observations in the comparative ranking task. Table 1 summarizes participant demographics and background characteristics. Participants received $3 for completing the 15-minute survey. The study was reviewed and approved by the Institutional Review Board at our university. 3.3. Financial Decision Scenarios 3.3.1. Decision-Context Taxonomy We constructed eight personal-finance scenarios using a 2×2×22× 2× 2 factorial taxonomy spanning three complementary properties of financial decisions: stakes, external uncertainty, and verifiability. Stakes captures the magnitude and persistence of a decision’s financial consequences; high-stakes decisions involved larger or longer-term consequences than low-stakes decisions. External uncertainty captures whether outcomes depend on unpredictable future events, such as market movements or unexpected expenses; high-uncertainty decisions depended more strongly on such events. Verifiability captures whether the quality of a recommendation can be evaluated against an established financial principle; high-verifiability decisions had a clearer normative benchmark than low-verifiability decisions which were preference-driven. The eight scenarios represented all combinations of these three dimensions. Experimental vignette methodology enables systematic variation of theoretically relevant attributes while retaining concrete decision contexts (Aguinis and Bradley, 2014; Atzmüller and Steiner, 2010). 3.3.2. Scenario Construction Each of the eight combinations of these dimensions was represented by a real-world financial scenario, covering decisions such as housing, credit card debt, travel insurance, emergency savings, graduate education, debt repayment and budgeting, long-term investing, and concentrated stock risk. Scenarios were approximately 100–150 words and described a protagonist’s financial position, relevant numerical information, the competing considerations, and a concrete advice-seeking question. We operationalized the three dimensions before constructing the final advice stimuli and reviewed each scenario against the corresponding stakes, uncertainty, and verifiability criteria. Table A1 in the appendix summarizes the eight scenarios and their factorial assignments. 3.4. Advice Stimulus Construction We first generated an AI-style advice response using a prompt that instructed the model (ChatGPT (OpenAI; GPT-5.3 Instant)) to act as a “neutral AI financial assistant” and respond to the financial decision described in the scenario. We then manually verified each response for factual accuracy and alignment with the intended financial recommendation. Corresponding Expert- and Online Community (OC)-style responses were then manually crafted to preserve the scenario-specific facts, numerical values, recommendation direction, and core financial rationale while varying source-associated advice style, tone, and reasoning format. Construction was informed by public financial-advice materials, including online financial communities, financial-planning resources, and conversational AI responses. The three conditions were designed to capture recurring communication patterns associated with contemporary financial-advice sources rather than to represent all communication from any source category. AI-style advice used a neutral, impersonal voice with explicit analytical and conditional reasoning. Expert-style advice used a professional, measured tone with principle-based reasoning, explicit trade-offs, and longer-term planning considerations. OC-style advice used a conversational voice, experiential framing, informal language, and practical heuristics. Table A2 summarizes the style specifications. 3.4.1. Stimulus Validation We conducted complementary linguistic, human, semantic, and external-reference checks to assess whether the constructed responses reflected the intended advice styles while preserving key substantive content. First, we compared 12 linguistic features associated with the intended style dimensions, including lexical complexity, formality, personal voice, conditional language, hedging, and contractions (Saha et al., 2026a; Tausczik and Pennebaker, 2010; Coleman and Liau, 1975; Babakov et al., 2023). Using Friedman tests across the eight matched scenario triplets with Benjamini–Hochberg FDR correction, 10 of 12 features differed across styles. The Categorical–Dynamic Index (CDI) (Pennebaker et al., 2014) showed a clear gradient from AI (M=27.36M=27.36) to Expert (M=21.30M=21.30) to OC (M=12.03M=12.03), χ2(2)=14.25χ^2(2)=14.25, p<.001p<.001, with all pairwise contrasts remaining significant after FDR correction. Three independent annotators, blind to the intended style labels, then classified each of the 24 advice texts and evaluated substantive consistency within each scenario. The intended style was identified in 91.7% of individual judgments, with majority classification matching all 24 stimuli (Fleiss’ κ=.749κ=.749). Across all eight scenario triplets, majority judgments indicated preservation of the recommendation direction, scenario-specific facts, and core financial rationale. Annotators flagged potentially material additional information in five triplets, primarily in OC responses, typically reflecting the anecdotal or experiential framing intentionally associated with that style. As a complementary semantic check, Sentence-Transformers embeddings (all-mpnet-base-v2) (Reimers and Gurevych, 2019) showed substantially greater similarity within than across scenarios (AI–Expert: .844.844 vs. .304.304; AI–OC: .728.728 vs. .282.282; Expert–OC: .798.798 vs. .293.293); across 24 directed cross-style comparisons, the nearest semantic counterpart always came from the same financial scenario (100% top-1 retrieval). Finally, we compared the constructed stimuli with a topic-matched reference corpus of 72 naturally occurring financial-advice texts from AI systems, financial professionals, and online communities. Several broad linguistic patterns reproduced across the constructed and reference texts, including differences in lexical complexity, first-person language, contractions, and CDI, although individual cues such as formality, conditional language, and hedging did not always follow the same ordering. We therefore interpret the manipulation as capturing controlled source-associated advice styles, rather than literal representations of all communication from each source category. Figure 2 summarizes the external comparison. Correspondence between constructed and external-reference linguistic profiles across AI, Expert, and Online Community advice. Source-level means are normalized within each linguistic dimension using the combined distribution of constructed and reference texts. Circles denote constructed stimuli, crosses denote external-reference texts, and connecting-line length reflects the discrepancy between the two profiles within each feature. Higher values indicate more categorical language for CDI and higher raw values for the remaining measures. Normalized positions should be interpreted within, rather than across, linguistic dimensions. Figure 2. Correspondence between constructed and external-reference linguistic profiles across AI, Expert, and Online Community advice. Source-level means are normalized within each linguistic dimension using the combined distribution of constructed and reference texts. Circles denote constructed stimuli, crosses denote external-reference texts, and connecting-line length reflects the discrepancy between the two profiles within each feature. Higher values indicate more categorical language for CDI and higher raw values for the remaining measures. Normalized positions should be interpreted within, rather than across, linguistic dimensions. Exact values and statistical comparisons are reported in Appendix A2. Correspondence between constructed and external-reference linguistic profiles across AI, Expert, and Online Community advice. Source-level means are normalized within each linguistic dimension using the combined distribution of constructed and reference texts. Circles denote constructed stimuli, crosses denote external-reference texts, and connecting-line length reflects the discrepancy between the two profiles within each feature. Higher values indicate more categorical language for CDI and higher raw values for the remaining measures. Normalized positions should be interpreted within, rather than across, linguistic dimensions. 3.5. Experimental Manipulation and Assignment Participants were randomly assigned between subjects to one of three source-label conditions. In the correctly labeled condition, advice was displayed with its corresponding source label (AI, Expert, or Online Community); in the unlabeled condition, no source label was shown; and in the mislabeled condition, advice was displayed with an incorrect source label. Two counterbalanced mislabeling schemas were used. In Schema A, AI-, Expert-, and OC-style advice were displayed with OC, AI, and Expert labels, respectively, in Task 1, and with Expert, OC, and AI labels in Task 2. Schema B reversed these mappings across the two tasks. When present, source labels appeared before the advice. The eight scenarios were divided into two groups of four. Participants were randomly assigned to a scenario group and one of three counterbalanced advice-style versions, such that each scenario appeared in each advice style across participants. Each participant evaluated all three advice styles across their four scenarios, with one style appearing twice, and scenario order was randomized. For the comparative ranking task, participants received a randomly selected, previously unseen scenario from the other scenario group and viewed all three advice-style versions side by side. 3.6. Procedure After providing informed consent, participants completed background questions on financial advice seeking and participant characteristics. They then completed four randomized scenario evaluations. For each, participants read a financial vignette, viewed one advice response under their assigned source-label condition, reported scenario familiarity, and rated the advice on ten evaluation items. After an attention check, participants completed the comparative ranking task using a previously unseen scenario from the other scenario group. They viewed all three advice-style versions side by side, ranked them from most to least preferred, and indicated the criteria that influenced their ranking. The survey concluded with the remaining participant measures and debriefing. 3.7. Measures 3.7.1. Advice Evaluation After each scenario–advice pair, participants rated ten items on seven-point Likert scales assessing readability, reasoning clarity, situational fit, risk acknowledgment, financial-harm risk, misleadingness, source knowledge, overall quality, Trust, and intended Reliance. For analysis, readability, reasoning clarity, and situational fit were combined as Message Appraisal; financial-harm risk and misleadingness were combined as Safety Concern; and Risk Acknowledgment and Source Knowledge were retained as single-item appraisals. Overall Quality, Trust, and Intended Reliance were retained as distinct downstream outcomes. Trust captured whether participants would trust the advice in the protagonist’s situation, whereas intended Reliance captured whether they would feel comfortable following it. Exact item wording appears in Table A4. 3.7.2. Participant Characteristics We measured participant characteristics that could shape financial-advice evaluation. Objective financial literacy was assessed using the “Big Three” questions on compound interest, inflation, and diversification (Lusardi and Mitchell, 2011; Lusardi and Mitchell, 2014), alongside self-rated financial literacy. Financial risk tolerance was measured using the Survey of Consumer Finances item (Grable and Lytton, 1999). AI-related measures included AI-use frequency, prior AI use for personal finance, perceived AI literacy (PAILQ-6) (Grassini, 2024), and General Trust in AI (McGrath et al., 2025). We additionally measured Human Trust Propensity (Frazier et al., 2013), prior use of financial advice sources, and scenario-specific familiarity. 3.7.3. Comparative Ranking Task The comparative ranking task provided a direct within-scenario comparison of the three advice styles. Participants viewed the AI, Expert, and OC advice for one previously unseen scenario and ranked the three responses from most to least preferred. They then indicated the considerations that informed their ranking, including readability, clarity of reasoning, safety, situational relevance, overall reliability, and alignment with their own thinking or preferences. Participants also reported their confidence in their ranking on a 5-point scale. 3.8. Analysis Strategy Because each participant evaluated four pieces of advice, our primary analyses used linear mixed-effects models with participant-level random intercepts to account for repeated observations. Advice-evaluation outcomes were standardized before modeling. We first examined which upstream factors—advice style, source labels, decision context, and scenario familiarity—were associated with the four appraisals (RQ1). We then examined how these appraisals were associated with Overall Quality, Trust, and Intended Reliance, sequentially incorporating preceding evaluation variables to characterize the broader evaluative pathway (RQ2). Table 2 summarizes the model specifications, inferential tests, and multiple-comparison corrections used throughout. We conducted two sets of complementary analyses. First, to disentangle advice style from source labels, we tested whether their associations were additive or interactive and decomposed their appraisal-carried pathways using participant-level bootstrap resampling. Second, we examined potential boundary conditions by testing both the direct associations of participant characteristics and whether individual differences or decision context moderated downstream evaluation pathways. Where multiple related hypotheses were tested, we applied the correction procedures specified in Table 2. Because appraisals and downstream judgments were measured within the same advice-evaluation task, we interpret the pathway analyses as associational rather than as evidence of causal mediation. Finally, we analyzed the side-by-side ranking task separately as convergent evidence using first-choice and complete-ranking comparisons. We also conducted a series of robustness checks addressing within-person variation, model specification, scenario parameterization, and composite construction. Additional statistical details and robustness results are provided in the appendix. Table 2. Summary of the analysis strategy. Analysis Model / Method Inference RQ1: What factors shape people’s appraisals of financial advice? Appraisals Appraisal∼UpstreamFactors+(1|Participant)Appraisal \ Factors+(1|Participant) LMM; β, p Participant characteristics Appraisal∼UpstreamFactors+ParticipantCharacteristics+(1|Participant)Appraisal \ Factors+Participant\ Characteristics+(1|Participant) Joint model; BH-FDR RQ2: How are appraisals associated with Overall Quality, Trust, and intended Reliance? Upstream baseline Outcome∼UpstreamFactors+(1|Participant)Outcome \ Factors+(1|Participant) Rm2R^2_m, Rc2R^2_c Overall Quality OverallQuality∼UpstreamFactors+FourAppraisals+(1|Participant)Overall\ Quality \ Factors+Four\ Appraisals+(1|Participant) LMM; β, p, R2R^2 Trust Trust∼UpstreamFactors+FourAppraisals+OverallQuality+(1|Participant)Trust \ Factors+Four\ Appraisals+Overall\ Quality+(1|Participant) LMM; β, p, R2R^2 Intended Reliance Reliance∼UpstreamFactors+FourAppraisals+OverallQuality+Trust+(1|Participant)Reliance \ Factors+Four\ Appraisals+Overall\ Quality+Trust+(1|Participant) LMM; β, p, R2R^2 Disentangling Advice Style and Source Labels Style × Label Outcome∼BaseModel+AdviceStyle×SourceLabelOutcome \ Model+Advice\ Style×Source\ Label LRT; BH-FDR for contrasts Pathway decomposition Advice Style / Source Label → Appraisals → Overall Quality → Trust → Reliance 2,000 bootstraps; 95% CI Boundary Conditions Participant characteristics Trust∼RQ2TrustModel+ParticipantCharacteristicsTrust 2\ Trust\ Model+Participant\ Characteristics Reliance∼RQ2RelianceModel+ParticipantCharacteristicsReliance 2\ Reliance\ Model+Participant\ Characteristics β, p; BH-FDR Individual-difference moderation Trust∼UpstreamFactors+EvaluationPredictors+Moderator+Moderator×EvaluationPredictors+(1|Participant)Trust \ Factors+Evaluation\ Predictors+Moderator+Moderator×Evaluation\ Predictors+(1|Participant) Reliance∼UpstreamFactors+EvaluationPredictors+Trust+Moderator+Trust×Moderator+(1|Participant)Reliance \ Factors+Evaluation\ Predictors+Trust+Moderator+Trust×Moderator+(1|Participant) LRT χ2χ^2; BH-FDR Decision-context moderation Trust∼UpstreamFactors+EvaluationPredictors+Context×EvaluationPredictors+(1|Participant)Trust \ Factors+Evaluation\ Predictors+Context×Evaluation\ Predictors+(1|Participant) Reliance∼UpstreamFactors+EvaluationPredictors+Trust+Context×(EvaluationPredictors+Trust)+(1|Participant)Reliance \ Factors+Evaluation\ Predictors+Trust+Context×(Evaluation\ Predictors+Trust)+(1|Participant) LRT χ2χ^2; BH-FDR Other Analyses Comparative ranking First-choice frequencies and complete within-participant rankings χ2χ^2; Friedman; Kendall’s W; Wilcoxon + Holm Robustness Within–between decomposition; GEE; scenario fixed effects; constituent-item models Sensitivity checks Note. Upstream Factors comprise Advice Style, Source Label, Stakes, External Uncertainty, Verifiability, and Scenario Familiarity. Four Appraisals comprise Message Appraisal, Risk Acknowledgment, Safety Concern, and Source Knowledge. Evaluation Predictors comprise the Four Appraisals and Overall Quality. Participant Characteristics comprise the seven individual-difference measures entered jointly. Moderator denotes one focal participant characteristic, and Context denotes one focal decision-context dimension (Stakes, External Uncertainty, or Verifiability). Base Model denotes the corresponding additive model before the focal interaction. All mixed-effects models include participant-level random intercepts. BH-FDR denotes Benjamini–Hochberg false-discovery-rate correction; LRT denotes likelihood-ratio test. 4. Results 4.1. Organizing Advice-Evaluation Measures into Appraisals We organized the 10 advice-evaluation items according to the role they play in the advice-evaluation process. Building on the Information Adoption Model (IAM), which distinguishes evaluations of the message from evaluations of its source (Sussman and Siegal, 2003), we averaged readability, reasoning clarity, and situational fit into a Message Appraisal index (α=.887α=.887). These items capture how understandable, well-reasoned, and applicable the advice appears. We retained Source Knowledge separately as a source-credibility judgment. Because financial advice can appear clear and credible while still presenting potential downside or epistemic risk, we evaluated risk and safety separately, consistent with prior financial-AI work that distinguishes information quality from perceived risk in advice evaluation (Ioannou et al., 2026). Risk of financial harm and misleadingness were averaged into a Safety Concern index (α=.712α=.712; inter-item r=.55r=.55), with higher values indicating greater concern. Risk Acknowledgment was retained separately because recognizing potential risks in the advice is conceptually different from judging the advice itself as harmful or misleading. Table 3 presents information on the four appraisals. Finally, we retained Overall Quality, Trust, and Intended Reliance as distinct downstream judgments, allowing us to examine how these initial appraisals relate to broader evaluation, trust, and willingness to follow the advice. Table 3. Construction and internal consistency of the four advice appraisals across the 1,140 advice evaluations. Higher Safety Concern scores indicate greater perceived risk of financial harm or misleadingness. Dashes indicate statistics that are not applicable to single-item appraisals. Appraisal Items k α Mean inter-item r Message Appraisal Readability; reasoning clarity; situational fit 3 .887 .72 Risk Acknowledgment Risk acknowledgment 1 – – Safety Concern Financial-harm risk; misleadingness 2 .712 .55 Source Knowledge Perceived source knowledge 1 – – 4.2. Advice Style, Source Labels, and Decision Context Shape Different Appraisals Advice style, source labels, and decision context were associated with different appraisal profiles (Table 4). Advice style primarily shaped Message Appraisal and Safety Concern. Compared with AI-style advice, both Expert- and Online Community (OC)-style advice received higher Message Appraisal ratings (βExpert=.334 _Expert=.334, p<.001p<.001; βOC=.223 _OC=.223, p<.001p<.001) and lower Safety Concern (βExpert=−.245 _Expert=-.245, p<.001p<.001; βOC=−.127 _OC=-.127, p=.026p=.026). The styles also differed in Source Knowledge: Expert-style advice was perceived as coming from a more knowledgeable source than AI-style advice (β=.141β=.141, p=.025p=.025), whereas OC-style advice was perceived as less knowledgeable (β=−.174β=-.174, p=.006p=.006), producing an Expert–AI–OC ordering on perceived source knowledge. Risk acknowledgment showed a different pattern: Expert and AI styles did not significantly differ (β=−.061β=-.061, p=.340p=.340), while OC-style advice was rated lower than AI-style advice (β=−.147β=-.147, p=.021p=.021) Source labels produced more selective differences. Relative to unlabeled advice, an Expert label was associated with greater Source Knowledge (β=.225β=.225, p=.017p=.017) and higher Message Appraisal (β=.190β=.190, p=.035p=.035). AI and OC labels did not differ significantly from the unlabeled condition across the four appraisals. Thus, source labels were most clearly reflected in perceived source expertise. Decision context primarily shaped risk and safety evaluations. High-stakes scenarios increased Safety Concern (β=.436β=.436, p<.001p<.001) and reduced Message Appraisal (β=−.241β=-.241, p<.001p<.001). Greater external uncertainty increased Risk Acknowledgment (β=.310β=.310, p<.001p<.001), while higher verifiability reduced both Risk Acknowledgment (β=−.199β=-.199, p=.010p=.010) and Safety Concern (β=−.292β=-.292, p<.001p<.001). Scenario familiarity was associated with more favorable evaluations across several dimensions: greater Message Appraisal (β=.319β=.319, p<.001p<.001), greater Risk Acknowledgment (β=.140β=.140, p<.001p<.001), greater perceived Source Knowledge (β=.206β=.206, p<.001p<.001), and lower Safety Concern (β=−.125β=-.125, p<.001p<.001). Individual differences shaped selected appraisals. Using the same base model, we examined participant characteristics jointly and found selective associations with the four appraisals.(Table 5). Greater AI Literacy was associated with higher Message Appraisal (β=.189β=.189, pFDR<.001p_FDR<.001) and Source Knowledge (β=.178β=.178, pFDR<.001p_FDR<.001). General AI Trust (β=.168β=.168, pFDR=.001p_FDR=.001) and Objective Financial Literacy (β=.086β=.086, pFDR=.038p_FDR=.038) were also positively associated with Source Knowledge. Greater Risk Tolerance was associated with higher Safety Concern (β=.145β=.145, pFDR=.015p_FDR=.015) and lower Source Knowledge (β=−.117β=-.117, pFDR=.011p_FDR=.011). Overall, individual differences affected particular appraisals rather than shifting advice appraisal uniformly. Overall, these results reveal differentiated entry points into advice appraisal: advice style was most broadly and consistently associated with message and safety appraisals, Expert source labels with perceived source knowledge, decision context with risk and safety appraisals, and individual differences with selected appraisal dimensions. Table 4. Mixed-effects models predicting the four initial appraisals of financial advice. Coefficients are standardized fixed-effect estimates (β). Models include a participant-level random intercept. AI-style advice and the unlabeled condition are the reference categories. Higher Safety Concern indicates greater perceived potential for financial harm or misleadingness. Message Appraisal Risk Ack. Safety Concern† Source Knowledge Advice style Expert vs. AI +0.334*** -0.061 -0.245*** +0.141* OC vs. AI +0.223*** -0.147* -0.127* -0.174** Source labels AI label vs. Unlabeled +0.075 -0.088 -0.141 +0.049 Expert label vs. Unlabeled +0.190* -0.026 -0.166 +0.225* OC label vs. Unlabeled +0.168 -0.044 -0.119 +0.091 Decision context High stakes -0.241*** +0.083 +0.436*** +0.010 High external uncertainty -0.072 +0.310*** +0.026 -0.005 High verifiability +0.099 -0.199** -0.292*** +0.036 Covariate Scenario familiarity +0.319*** +0.140*** -0.125*** +0.206*** Marginal R2R^2 (Rm2R^2_m) 0.170 0.054 0.115 0.067 Conditional R2R^2 (Rc2R^2_c) 0.405 0.298 0.458 0.312 Note. N=1,140N=1,140 evaluations from 285 participants. Scenario familiarity is standardized; binary experimental predictors are coded 0/1. ∗p<.05^*p<.05, p∗∗<.01^**p<.01, ∗p<.001^***p<.001. Table 5. Participant characteristics associated with initial advice appraisals. Participant characteristic Appraisal β p_FDR AI Literacy Message Appraisal .189 <<.001 AI Literacy Source Knowledge .178 <<.001 General AI Trust Source Knowledge .168 .001 Objective Financial Literacy Source Knowledge .086 .038 Risk Tolerance Safety Concern .145 .015 Risk Tolerance Source Knowledge −.117-.117 .011 Note. Each appraisal was modeled separately using a linear mixed-effects model with participant-level random intercepts. Models included the same base predictors (advice style, source labels, decision context, and scenario familiarity), with all seven participant characteristics entered simultaneously. Benjamini–Hochberg FDR correction was applied across the seven participant-characteristic coefficients within each appraisal. Only associations surviving pFDR<.05p_FDR<.05 are shown. 4.3. From Appraisals to Quality, Trust, and Intended Reliance We next examined how the four appraisals were associated with Overall Quality, Trust, and intended Reliance. For each outcome, we compared an upstream model containing advice style, source labels, decision context, and scenario familiarity with a model that additionally included the preceding appraisal variables (Table 6). Appraisals were strongly associated with Overall Quality. All four intermediate appraisals were independently associated with Overall Quality. Source Knowledge showed the strongest association (β=.489β=.489, p<.001p<.001), followed by Message Appraisal (β=.325β=.325, p<.001p<.001). Greater Safety Concern predicted lower Overall Quality (β=−.088β=-.088, p<.001p<.001), while greater Risk Acknowledgment predicted somewhat higher quality ratings (β=.075β=.075, p<.001p<.001). Adding the appraisals also increased marginal explained variance from 9.8% to 69.2%. We also observe that initial Expert-style and Expert-label associations with Overall Quality were substantially attenuated after accounting for these appraisals. Overall Quality was the strongest correlate of Trust. Once the appraisals and Overall Quality were included, Overall Quality was by far the strongest correlate of Trust (β=.607β=.607, p<.001p<.001). Message Appraisal (β=.132β=.132, p<.001p<.001), lower Safety Concern (β=−.125β=-.125, p<.001p<.001), and Source Knowledge (β=.109β=.109, p<.001p<.001) retained smaller independent associations with Trust, whereas Risk Acknowledgment did not. Marginal explained variance increased from 11.2% in the upstream model to 75.9% in the appraisal- and quality-adjusted model. The initial Expert-style association with Trust was also no longer detectable after including these variables. Trust was the strongest correlate of Intended Reliance. Trust was the dominant correlate of Intended Reliance (β=.594β=.594, p<.001p<.001), followed by Overall Quality (β=.206β=.206, p<.001p<.001); the remaining appraisals showed substantially smaller associations: Message Appraisal (β=.079β=.079, p<.001p<.001), lower Safety Concern (β=−.050β=-.050, p<.001p<.001), and Source Knowledge (β=.047β=.047, p=.020p=.020). Adding the full evaluation pathway increased marginal explained variance from 14.8% to 82.9%. After accounting for these variables, initial advice-style and source-label differences in Reliance were substantially reduced. That said, because we measured these intermediate judgments within the same advice evaluation, we interpret this structure as an evaluative pathway rather than evidence of causal mediation. Across all three downstream outcomes, models including the appraisal pathway explained substantially more variance than upstream factors alone, while the initial advantages associated with Expert advice style and Expert source labels were largely attenuated after accounting for these appraisals. Taken together, the results are consistent with a progressively organized evaluative pathway in which advice style, source labels, and decision context first shape distinct message, risk, safety, and source-related appraisals; these appraisals are strongly associated with Overall Quality; Overall Quality is in turn strongly associated with Trust; and Trust is the strongest correlate of Intended Reliance. Table 6. Upstream and pathway-adjusted models of Overall Quality, Trust, and Intended Reliance. Overall Quality Trust Intended Reliance Predictor Upstream Pathway- adjusted Upstream Pathway- adjusted Upstream Pathway- adjusted Advice style Expert vs. AI +0.205*** +0.011 +0.185** -0.028 +0.173** -0.019 OC vs. AI -0.004 +0.005 +0.009 -0.009 +0.023 +0.002 Source labels AI label vs. Unlabeled +0.126 +0.074 +0.066 -0.038 +0.103 +0.029 Expert label vs. Unlabeled +0.248** +0.064 +0.176 -0.046 +0.210* +0.013 OC label vs. Unlabeled +0.170 +0.060 +0.154 +0.004 +0.189* +0.040 Decision context High stakes -0.040 +0.069* -0.107* -0.000 -0.155** -0.048 High external uncertainty -0.013 -0.008 +0.001 +0.014 +0.017 +0.030 High verifiability +0.152 +0.091* +0.166* +0.026 +0.178* +0.022 Covariate Scenario familiarity +0.256*** +0.038* +0.280*** +0.040* +0.324*** +0.049*** Appraisal-to-reliance pathway Message Appraisal — +0.325*** — +0.132*** — +0.079*** Risk Acknowledgment — +0.075*** — +0.024 — -0.015 Safety Concern† — -0.088*** — -0.125*** — -0.050*** Source Knowledge — +0.489*** — +0.109*** — +0.047* Overall Quality — — — +0.607*** — +0.206*** Trust — — — — — +0.594*** Model fit Marginal R2R^2 (Rm2R^2_m) 0.098 0.692 0.112 0.759 0.148 0.829 Conditional R2R^2 (Rc2R^2_c) 0.373 0.742 0.351 0.795 0.362 0.829 Note. Cells report standardized fixed-effect coefficients (β) from linear mixed-effects models with participant-level random intercepts. Upstream models include advice style, source labels, decision context, and scenario familiarity. Pathway-adjusted models additionally include the four initial appraisals; the Trust model further includes Overall Quality, and the Reliance model further includes Overall Quality and Trust. AI-style advice and the unlabeled condition are the reference categories. Models include N=285N=285 participants and 1,1401,140 advice evaluations. Rm2R^2_m and Rc2R^2_c denote marginal and conditional R2R^2, respectively. Orange and blue shading indicate significant positive and negative coefficients, respectively; shading intensity reflects statistical significance (p<.05p<.05, p<.01p<.01, and p<.001p<.001), not effect magnitude. Unshaded coefficients are not statistically significant. † Higher values indicate greater perceived financial harm or misleadingness. ∗p<.05^*p<.05, p∗∗<.01^**p<.01, ∗p<.001^***p<.001. 4.4. Advice Style and Source Labels Operate Additively Through Different Appraisals In a subsequent exploratory analysis, we tested whether the associations of advice style depended on the source label displayed with the advice. Across the four appraisals and three downstream outcomes, adding Advice Style × Source Label interactions did not improve model fit (all p>.29p>.29; Table A5). Thus, these results support treating advice style and displayed source labels as largely additive rather than interactive effects. We then examined whether these two cues were reflected differently through the appraisal pathway. Participant-level bootstrap decompositions showed that advice style was relatively more strongly reflected through Message Appraisal, whereas source labels were relatively more strongly reflected through Source Knowledge. For Expert advice, the Message-versus-Knowledge pathway was more pronounced for advice style than for the corresponding Expert label (ΔΔ=.081 =.081, 95% CI [.014,.148][.014,.148]); the same pattern appeared for Online Community advice (ΔΔ=.124 =.124, 95% CI [.045,.212][.045,.212]). These results suggest that how advice is written is reflected more strongly in appraisals of the message itself, whereas explicit source labels shift appraisal relatively toward perceived source expertise. Full pathway decompositions and simple contrasts are reported in Table A6 and Table A7, respectively. 4.5. Boundary Conditions of the Evaluation Pathway We next examined whether the relationships among appraisals, Overall Quality, Trust, and Intended Reliance varied across decision contexts or participant characteristics. Decision context did not substantially alter the downstream pathway. Although stakes, external uncertainty, and verifiability shaped the initial appraisals participants formed, we found no robust evidence that any of the three decision-context dimensions moderated how those appraisals translated into Trust or intended Reliance after FDR correction. Thus, decision context appeared to matter primarily at the appraisal stage rather than in the subsequent evaluation pathway. Individual differences produced selective moderation. Objective Financial Literacy strengthened the associations of Source Knowledge with Trust (βint=.054 _int=.054, pFDR=.017p_FDR=.017) and of Trust with Intended Reliance (βint=.039 _int=.039, pFDR=.002p_FDR=.002). Human Trust Propensity selectively moderated the association between Safety Concern and Trust (βint=.056 _int=.056, pFDR=.002p_FDR=.002), with safety concerns weighing more strongly among participants lower in trust propensity. General AI Trust was associated with higher overall Trust (β=.202β=.202, pFDR<.001p_FDR<.001) and Intended Reliance (β=.227β=.227, pFDR<.001p_FDR<.001), but these associations were substantially reduced after accounting for the appraisal pathway. Overall, the downstream pathway was comparatively stable across decision contexts, while individual differences altered only selected parts of how appraisals translated into Trust and Reliance. More details are reported in Appendix B.2. 4.6. Comparative Ranking Provides Convergent Evidence As a final comparative task, participants ranked AI-, Expert-, and Online Community (OC)-style advice side by side for a previously unseen financial scenario. Expert-style advice was ranked first by 47.4% of participants, compared with 29.1% for OC advice and 23.5% for AI advice (χ2(2)=26.61χ^2(2)=26.61, p<.001p<.001). Across the complete rankings, the three advice styles also differed significantly (Friedman χ2(2)=37.16χ^2(2)=37.16, p<.001p<.001, Kendall’s W=.065W=.065). Pairwise Wilcoxon signed-rank tests with Holm correction showed that Expert-style advice was preferred over both AI- and OC-style advice (both pHolm<.001p_Holm<.001), whereas AI and OC did not differ. Notably, the Expert-style preference remained in the Unlabeled condition, where 47.8% ranked it first, compared with 31.5% for OC and 20.7% for AI (Figure 3). We found no evidence that the relative ranking of the three styles systematically differed across source-label conditions (χ2(4)=6.65χ^2(4)=6.65, p=.155p=.155). Figure 3. Percentage of participants ranking each advice style first in the comparative ranking task, overall and by source label condition. Expert-style advice was most frequently ranked first across all conditions. Although its descriptive advantage was smaller in the Mislabeled condition, the overall Advice Style × Source Label Arm interaction was not statistically significant. Percentage of participants ranking each advice style first in the comparative ranking task, overall and by source label condition. Expert-style advice was most frequently ranked first across all conditions. Although its descriptive advantage was smaller in the Mislabeled condition, the overall Advice Style $×$ Attribution Arm interaction was not statistically significant. Participants’ self-reported ranking criteria further emphasized properties of the advice itself. Clear reasoning (65.6%), ease of reading and understanding (60.0%), situational relevance (53.3%), and acknowledgment of risks and uncertainties (50.9%) were the most frequently selected criteria, whereas only 7.0% reported that the displayed source influenced their ranking (Figure 4). Although these self-reports do not identify what causally determined participants’ rankings, the comparative task provides convergent evidence for the primary results: participants preferred Expert-style advice and predominantly described their preferences in terms of message-level appraisal cues rather than explicit source information. Self-reported criteria used in the comparative ranking task. Participants could select multiple criteria. Percentages indicate the proportion of participants who selected each criterion. Clear reasoning, readability, situational relevance, and acknowledgment of risks and uncertainties were the most frequently reported considerations, whereas explicit source influence was comparatively rare. Figure 4. Self-reported criteria used in the comparative ranking task. Participants could select multiple criteria. Percentages indicate the proportion of participants who selected each criterion. Clear reasoning, readability, situational relevance, and acknowledgment of risks and uncertainties were the most frequently reported considerations, whereas explicit source influence was comparatively rare.Self-reported criteria used in the comparative ranking task. Participants could select multiple criteria. Percentages indicate the proportion of participants who selected each criterion. Clear reasoning, readability, situational relevance, and acknowledgment of risks and uncertainties were the most frequently reported considerations, whereas explicit source influence was comparatively rare. 4.7. Robustness Checks We conducted four robustness checks to assess whether the appraisal pathway depended on stable participant response tendencies, the repeated-measures estimator, the representation of decision context, or the construction of the appraisal composites. First, within–between participant decompositions showed that the principal associations remained at the within-participant level, indicating that the pathway was not driven solely by participants who generally rated all advice more favorably. Second, generalized estimating equations closely reproduced the mixed-effects estimates; for example, the Overall Quality–Trust association was β=.608β=.608 (vs. .607.607 in the primary model) and the Trust–Reliance association was β=.595β=.595 (vs. .594.594). Third, replacing the decision-context dimensions with scenario fixed effects produced nearly identical pathway estimates. Finally, replacing Message Appraisal and Safety Concern with their constituent items preserved the main conclusions. Overall, the appraisal–Quality–Trust–Reliance pattern was stable across alternative specifications and measurement choices. More details are reported in Appendix C. 5. Discussion The landscape of personal financial decision-making is rapidly evolving as generative AI becomes an increasingly accessible source of financial guidance. As these systems become more prevalent, the central challenge is not simply whether people trust AI-generated advice, but whether that trust and subsequent reliance are grounded in an appropriate evaluation of the advice. Existing trust-calibration approaches commonly rely on cues such as disclaimers, source labels, communication choices, and anthropomorphic design, but examining their direct relationship with Trust provides only a partial account of how users arrive at these judgments. Our work instead frames trust and reliance as part of a broader evaluation process, showing that advice style, source labels, and decision context are associated with distinct appraisals of the message, its risks and safety, and the source’s perceived knowledge. We discuss the implications of this work. 5.1. Extending Information Adoption for Financial Advice Evaluation A growing body of work on AI-generated information and human–AI advisory systems has approached users’ responses through the lens of information evaluation and adoption (Ioannou et al., 2026; Li et al., 2025; Gong et al., 2026; Camilleri and Camilleri, 2024). Classical IAM accounts, as discussed previously, organize this process around evaluations of argument quality and source credibility, which shape perceived usefulness and subsequent adoption (Sussman and Siegal, 2003). Our work extends this perspective by opening up the process through which financial advice becomes trusted and relied upon. Rather than treating Trust as a direct response to an advice source or interface cue, our results characterize it as part of a broader evaluation process in which people form distinct appraisals of the message, its risks and safety, and its source, which are associated with Overall Quality, Trust, and ultimately Intended Reliance. This more granular view matters because similar levels of trust can mask very different underlying appraisal profiles. For example, Online-Community-style advice was appraised more favorably than AI-style advice on message and safety dimensions, but less favorably on perceived Source Knowledge; these differences largely offset downstream, leaving trust in the two styles statistically indistinguishable. Looking beneath aggregate trust therefore provides a more diagnostic account of why advice is trusted and which message-, risk-, safety-, or source-related cues may warrant design attention, consistent with work on appropriate reliance that does not treat higher Trust as inherently desirable (Lee and See, 2004; Raees et al., 2026). Our findings also elaborate this evaluation process for financial decisions by making risk- and safety-related appraisals explicit. Financial advice can appear clear, relevant, and credible while still failing to acknowledge uncertainty or exposing users to potential harm. Recent work on AI-assisted financial decision-making similarly highlights perceived risk and uncertainty as important to advice adoption (Ioannou et al., 2026). Our results show that these appraisals were sensitive to characteristics of the decision itself: greater external uncertainty was associated with greater Risk Acknowledgment (β=.310β=.310), while higher stakes were associated with greater Safety Concern (β=.436β=.436); higher verifiability was associated with lower levels of both. These patterns suggest that perceptions of risk and safety are closely associated with the characteristics of the financial decision in which the advice is encountered. Decision context therefore appears to enter the evaluation process by influencing which aspects of uncertainty and potential harm become salient to users, extending information-evaluation accounts beyond properties of the message and its source alone. Finally, our findings reinforce the distinction between Trust and Intended Reliance. Trust is commonly treated as an antecedent of reliance on automated systems, but the two are not equivalent: Lee and See (2004) emphasize that the central design problem is whether trust supports appropriate reliance, while advice-taking research similarly distinguishes attitudes toward a source from whether its recommendation is incorporated into a decision (Bonaccio and Dalal, 2006). In our results, Trust was the strongest correlate of Intended Reliance, yet Overall Quality, Message Appraisal, and Safety Concern retained smaller independent associations after accounting for Trust. Participant-level variation further supported this distinction: scenario familiarity was directly associated with greater Reliance (β=.089β=.089, p<.001p<.001) despite a weaker association with Trust, while objective financial literacy strengthened the Trust–Reliance relationship (βint=.039 _int=.039, pFDR=.002p_FDR=.002). Thus, similar levels of Trust need not translate into similar intentions to act. This aligns with prior work suggesting that studies of AI-supported decision-making should measure Trust and Reliance separately and, where possible, complement stated intentions with behavioral measures of whether users accept, reject, or verify advice (Raees et al., 2026). 5.2. Advice Style and Source Labels as distinct epistemic cues Prior HCI research shows that how AI communicates can shape how users evaluate and trust it. Authoritative language can increase Trust even when system limitations are disclosed (Metzger et al., 2024), while social and personality cues such as friendliness, extroversion, and stance adaptation can similarly influence responses to AI (Takayanagi et al., 2025; Sun and Wang, 2026). Parallel work shows that displayed source information can act as a separate credibility cue: identical information may be evaluated differently depending on whether it is attributed to AI, a human, or an expert (Jakesch et al., 2019; Reis et al., 2024; Sun et al., 2026; Sahebi et al., 2026). Our findings bring these two strands together by independently varying advice style and source labels, showing that they were associated with distinct appraisal profiles. Advice style was reflected more broadly in message- and safety-related appraisals. Consistent with this pattern, Expert-style advice remained the most frequently preferred in the comparative ranking task even when no source label was displayed. Expert source labels, in contrast, were most clearly associated with perceived Source Knowledge. Their associations were also largely additive rather than interactive, suggesting that sounding like an expert and being identified as an expert provide distinct rather than interchangeable signals. From an AI-governance perspective, this distinction raises concerns around provenance and accountability. Generative AI can readily reproduce the linguistic markers of professional advice—such as structured reasoning, measured language, and explicit trade-offs—without necessarily carrying the expertise, review, or responsibility associated with a qualified professional. This gap between presentational authority and epistemic authority may encourage unwarranted confidence when polished communication is mistaken for evidence of expertise. Our findings therefore suggest that advice style and source information should be treated as distinct layers of evaluation and governance in AI-assisted financial advice. Systems should make clear who produced, reviewed, or endorsed a recommendation and where responsibility lies, rather than allowing professional presentation alone to signal authority. This distinction is already relevant in deployed systems: SoFi reports that its financial planners were involved in evaluating responses from Coach, illustrating how AI-generated advice may incorporate human professional review without being directly authored by a human advisor (Caitlin Mullen, 2026). 5.3. Design Implications for AI-Mediated Financial Advice Our findings suggest three considerations for the design and evaluation of AI-mediated financial advice systems. Design for grounded evaluation and appropriate Reliance. Financial AI systems should help users assess why a recommendation deserves Trust and whether it warrants action, rather than optimizing Trust through surface-level cues such as professionalized advice style alone. This challenge is already visible in deployed financial AI systems: in testing its AI financial guide, Coach, SoFi reported that effective financial guidance depends not only on the underlying financial logic, but also on how recommendations and trade-offs are communicated to users (Caitlin Mullen, 2026). Interfaces could make relevant assumptions, reasoning, trade-offs, uncertainty, and potential downside consequences easier to inspect—for example, through expandable assumptions or a brief “Why this recommendation?” view. Such support should ultimately help users decide when to follow, verify, reject, or escalate a recommendation, including by facilitating comparison with other sources or suggesting professional consultation when appropriate. Evaluations of these systems should likewise distinguish Trust from Intended Reliance and, where possible, examine actual advice-taking and verification behavior rather than positive perceptions alone. Separate professional presentation from verifiable epistemic authority. Clear and professional communication can improve how advice is evaluated, but it should not substitute for evidence of expertise. Financial advice systems should distinguish whether advice was AI-generated, professionally reviewed, or directly authored by a qualified advisor; make the relevant expertise or credentials behind that involvement visible; expose the evidence or financial principles supporting the recommendation; and clarify where responsibility lies and what limitations apply. Provenance could function as an independently inspectable layer alongside the advice itself, rather than as a decorative “Expert” badge. This separation can preserve the benefits of clear communication without allowing presentational authority alone to stand in for warranted epistemic authority. Adapt evaluation support to the decision and the user. Decision context was reflected primarily in the appraisals users formed; interfaces could therefore foreground different information depending on what makes a financial decision difficult to evaluate. For example,high-stakes decisions might emphasize potential downside consequences, uncertain decisions might make assumptions and contingencies explicit, and difficult-to-verify decisions could make trade-offs and alternative courses of action more explicit rather than presenting a single recommendation as definitive. User-sensitive support could similarly help people evaluate cues such as source expertise or safety concerns, rather than assuming that all users interpret these cues in the same way. Lastly, because greater familiarity was consistently associated with more favorable appraisals of trust and Reliance, financial AI systems may need to guard against familiarity reducing scrutiny. Even when a scenario feels familiar to the user, interfaces should preserve lightweight checks on assumptions, risks, and verifiability so reliance is based on evaluating the advice rather than familiarity with the situation alone. 5.4. Limitations and Future Work Our work has several limitations that also suggest interesting future directions. First, Reliance was measured as behavioral intention rather than observed financial behavior. Real financial decisions often unfold over time, involve monetary consequences, and include consultation with other people or information sources. Future work should therefore examine whether the evaluation patterns observed here predict actual advice taking, verification, information seeking, and decision outcomes, including whether users appropriately accept, reject, or escalate recommendations. Second, although our results are consistent with an evaluative organization linking appraisals, Overall Quality, Trust, and Intended Reliance, these constructs were measured within the same advice- evaluation task. The observed associations therefore should not be interpreted as establishing a causal or temporal pathway. Future work could test this ordering more directly through longitudinal or experimental designs that manipulate specific appraisals and observe subsequent changes in Trust and Reliance. Third, we used controlled financial scenarios and deliberately constructed advice styles. This allowed us to hold the underlying recommendation relatively constant while varying its presentation and source label, but no small set of styles can capture the diversity of AI, professional, or online-community advice. Real-world AI advice is also interactive: users ask follow-up questions, introduce new constraints, challenge recommendations, and observe how systems respond. Extending this work to multi-turn and longitudinal interactions is therefore an important next step. Fourth, our decision-context dimensions were instantiated through eight specific financial scenarios. Although robustness analyses suggest that the downstream pattern was not driven by individual scenarios, other financial decisions may involve factors not represented here, including urgency, irreversibility, regulatory complexity, social consequences, or ambiguity in user goals. Similarly, our sample consisted of U.S. adults recruited through Prolific; financial norms, access to professional advice, and expectations of AI and expertise may differ across populations and institutional settings. Replication across financial domains, countries, and real advisory settings will therefore be important for establishing generalizability. Finally, individual characteristics such as financial literacy, AI literacy, and risk tolerance were measured rather than experimentally manipulated. Accordingly, their associations and moderation patterns should not be interpreted causally. Future work could test whether interventions that improve financial knowledge, source-verification skills, or risk comprehension change how users appraise advice and translate Trust into Reliance over time. Longitudinal studies could also examine how these relationships evolve as users accumulate experience with an AI system and observe the consequences of its advice. 6. Conclusion As generative AI becomes an increasingly common source of personal finance advice, the challenge is not simply whether people trust its advice, but whether that trust is grounded in how they evaluate it. In a randomized experiment with 285 U.S. adults, we show that advice style, source labels, and decision context shaped distinct appraisals of the message, its risks and safety, and its source. These appraisals were strongly associated with Overall Quality, which was in turn closely associated with Trust, while Trust was the strongest correlate of Intended Reliance. Advice style and source labels contributed different signals: style was associated more broadly with message- and safety-related appraisals, whereas Expert labels were most clearly associated with perceived Source Knowledge. Decision context and individual differences further shaped selected parts of this evaluation process. Together, these findings shift the design goal from making financial AI appear trustworthy toward supporting grounded evaluation: helping users assess the reasoning, risks, source expertise, and context of advice so that Trust and Reliance better reflect what the recommendation warrants. Acknowledgements. Need to add this: AI declaration, annotators, etc References Mor (2026) 2026. As More U.S. Consumers Struggle with Rising Prices, Many Turn to Artificial Intelligence for Financial Advice. https://w.jdpower.com/business/resources/more-us-consumers-struggle-rising-prices-many-turn-artificial-intelligence/ Aguinis and Bradley (2014) Herman Aguinis and Kyle J. Bradley. 2014. Best Practice Recommendations for Designing and Implementing Experimental Vignette Methodology Studies. Organizational Research Methods 17, 4 (Oct. 2014), 351–371. doi:10.1177/1094428114547952 Akana and Santucci (2025) Tom Akana and Larry Santucci. 2025. Trust and Credibility: Comparing Sources of Financial Advice. (2025). https://w.philadelphiafed.org/consumer-finance/trust-and-credibility-comparing-sources-of-financial-advice Allgood and Walstad (2016) Sam Allgood and William B. Walstad. 2016. THE EFFECTS OF PERCEIVED AND ACTUAL FINANCIAL LITERACY ON FINANCIAL BEHAVIORS. Economic Inquiry 54, 1 (Jan. 2016), 675–697. doi:10.1111/ecin.12255 Alruwaili et al. (2025) Reshaa F. Alruwaili, Abdullah A. Alasmari, Hussien Tash Niyazi, I. K. Youssef, Hanan Aifan, and Somia A. Asklany. 2025. Modeling public trust in AI cognitive capabilities using statistical and machine learning approaches. Scientific Reports 15, 1 (Nov. 2025), 39856. doi:10.1038/s41598-025-23447-4 Atzmüller and Steiner (2010) Christiane Atzmüller and Peter M. Steiner. 2010. Experimental Vignette Studies in Survey Research. Methodology 6, 3 (Jan. 2010), 128–138. doi:10.1027/1614-2241/a000014 Ayers et al. (2023) John W. Ayers, Adam Poliak, Mark Dredze, Eric C. Leas, Zechariah Zhu, Jessica B. Kelley, Dennis J. Faix, Aaron M. Goodman, Christopher A. Longhurst, Michael Hogarth, and Davey M. Smith. 2023. Comparing Physician and Artificial Intelligence Chatbot Responses to Patient Questions Posted to a Public Social Media Forum. JAMA Internal Medicine 183, 6 (June 2023), 589–596. doi:10.1001/jamainternmed.2023.1838 Babakov et al. (2023) Nikolay Babakov, David Dale, Ilya Gusev, Irina Krotova, and Alexander Panchenko. 2023. Don’t Lose the Message While Paraphrasing: A Study on Content Preserving Style Transfer. In Natural Language Processing and Information Systems: 28th International Conference on Applications of Natural Language to Information Systems, NLDB 2023, Derby, UK, June 21–23, 2023, Proceedings. Springer-Verlag, Berlin, Heidelberg, 47–61. doi:10.1007/978-3-031-35320-8_4 Bonaccio and Dalal (2006) Silvia Bonaccio and Reeshad S. Dalal. 2006. Advice taking and decision-making: An integrative literature review, and implications for the organizational sciences. Organizational Behavior and Human Decision Processes 101, 2 (Nov. 2006), 127–151. doi:10.1016/j.obhdp.2006.07.001 Bosompim et al. (2026) Prince Bosompim, Richmond Essieku, and Jodi Letkiewicz. 2026. In whom we trust? Social networks and trust in AI vs. human experts financial advice. Finance Research Letters 91 (March 2026), 109437. doi:10.1016/j.frl.2025.109437 Buçinca et al. (2021) Zana Buçinca, Maja Barbara Malaya, and Krzysztof Z. Gajos. 2021. To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making. Proc. ACM Hum.-Comput. Interact. 5, CSCW1 (April 2021), 188:1–188:21. doi:10.1145/3449287 Caitlin Mullen (2026) Caitlin Mullen. 2026. What SoFi learned testing its AI adviser Coach | Banking Dive. https://w.bankingdive.com/news/sofi-bank-ai-adviser-coach-financial-planning/822329/ Camilleri and Camilleri (2024) Mark Anthony Camilleri and Adriana Caterina Camilleri. 2024. The Acceptance and Usage of ChatGPT: An Information Adoption Model perspective. In 2024 8th International Conference on Communications and Future Internet (ICCFI). 61–66. doi:10.1109/ICCFI63302.2024.00017 Chaiken (1980) Shelly Chaiken. 1980. Heuristic versus systematic information processing and the use of source versus message cues in persuasion. Journal of Personality and Social Psychology 39, 5 (1980), 752–766. doi:10.1037/0022-3514.39.5.752 Cheng et al. (2022) Xusen Cheng, Xiaoping Zhang, Jason Cohen, and Jian Mou. 2022. Human vs. AI: Understanding the impact of anthropomorphism on consumer response to chatbots from the perspective of trust and relationship norms. Information Processing & Management 59, 3 (May 2022), 102940. doi:10.1016/j.ipm.2022.102940 Choukhmane et al. (2026) Taha Choukhmane, Tim de Silva, Weidong Lin, and Matthew Akuzawa. 2026. AI Financial Advice: Supply, Demand, and Life Cycle Implications. doi:10.2139/ssrn.6446286 Coleman and Liau (1975) Meri Coleman and Ta Lin Liau. 1975. A computer readability formula designed for machine scoring. Journal of Applied Psychology 60, 2 (1975), 283. Erkan and Evans (2016) Ismail Erkan and Chris Evans. 2016. The influence of eWOM in social media on consumers’ purchase intentions: An extended approach to information adoption. Computers in Human Behavior 61 (Aug. 2016), 47–55. doi:10.1016/j.chb.2016.03.003 Fairhurst and Greene (2025) Douglas (DJ) Fairhurst and Daniel Greene. 2025. How Much Does ChatGPT Know about Finance? Financial Analysts Journal 81, 1 (Jan. 2025), 12–32. doi:10.1080/0015198X.2024.2411941 Fatouros et al. (2025) George Fatouros, Kostas Metaxas, John Soldatos, and Dimosthenis Kyriazis. 2025. Can Large Language Models beat wall street? Evaluating GPT-4’s impact on financial decision-making with MarketSenseAI. Neural Computing and Applications 37, 30 (Oct. 2025), 24893–24918. doi:10.1007/s00521-024-10613-4 Financial Planning Standards Board ([n. d.]) Financial Planning Standards Board. [n. d.]. Standards for the profession. https://fpsb.org/standards-for-the-profession/ Frazier et al. (2013) M. Lance Frazier, Paul D. Johnson, and Stav Fainshmidt. 2013. Development and validation of a propensity to trust scale. Journal of Trust Research 3, 2 (Oct. 2013), 76–97. doi:10.1080/21515581.2013.820026 Gallegos et al. (2026) Isabel O. Gallegos, Chen Shani, Weiyan Shi, Federico Bianchi, Izzy Gainsburg, Dan Jurafsky, and Robb Willer. 2026. Labeling messages as AI-generated does not reduce their persuasive effects. PNAS nexus 5, 2 (Feb. 2026), pgag008. doi:10.1093/pnasnexus/pgag008 Goh et al. (2025) Adalia Y. H. Goh, Andree Hartanto, and Nadyanna M. Majeed. 2025. Generative artificial intelligence dependency: Scale development, validation, and its motivational, behavioral, and psychological correlates. Computers in Human Behavior Reports 20 (Dec. 2025), 100845. doi:10.1016/j.chbr.2025.100845 Gong et al. (2026) Nannan Gong, Joohwan Seo, and Huanzhen Liu. 2026. When algorithms speak: Boundary conditions of information adoption in AI- and user-generated content. Journal of Retailing and Consumer Services 93 (Aug. 2026), 104926. doi:10.1016/j.jretconser.2026.104926 Grable and Lytton (1999) John Grable and Ruth H. Lytton. 1999. Financial risk tolerance revisited: the development of a risk assessment instrument. Financial Services Review 8, 3 (Sept. 1999), 163–181. doi:10.1016/S1057-0810(99)00041-4 Grassini (2024) Simone Grassini. 2024. A Psychometric Validation of the PAILQ-6: Perceived Artificial Intelligence Literacy Questionnaire. In Nordic Conference on Human-Computer Interaction. ACM, Uppsala Sweden, 1–10. doi:10.1145/3679318.3685359 Grawitch et al. (2026) Matthew J. Grawitch, Steven L. Winton, and Srikanth P. Mudigonda. 2026. Who listens to ChatGPT and when should they? A two-study examination of AI-assisted decision making. Computers in Human Behavior: Artificial Humans 7 (March 2026), 100283. doi:10.1016/j.chbah.2026.100283 Hean et al. (2025) Oudom Hean, Utsha Saha, and Binita Saha. 2025. Can AI help with your personal finances? Applied Economics 57, 60 (Dec. 2025), 11219–11227. doi:10.1080/00036846.2025.2450384 Huang and Ball (2024) Kuo-Ting Huang and Christopher Ball. 2024. The Influence of AI Literacy on User’s Trust in AI in Practical Scenarios: A Digital Divide Pilot Study. Proceedings of the Association for Information Science and Technology 61, 1 (Oct. 2024), 937–939. doi:10.1002/pra2.1146 Huang et al. (2025) Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. 2025. A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions. ACM Transactions on Information Systems 43, 2 (Jan. 2025), 42:1–42:55. doi:10.1145/3703155 Inc (2026) Gallup Inc. 2026. Where Americans and Canadians Turn for Financial Guidance. https://news.gallup.com/poll/712952/americans-canadians-turn-financial-guidance.aspx Section: Economy. Ioannou et al. (2026) Athina Ioannou, M. Mahdi Tavalaei, Dorthea Vatn, and Patrick Mikalef. 2026. The Role of Explainability In AI-Driven Financial Decision Making. Information Systems Frontiers (April 2026). doi:10.1007/s10796-026-10727-1 Jakesch et al. (2019) Maurice Jakesch, Megan French, Xiao Ma, Jeffrey T. Hancock, and Mor Naaman. 2019. AI-Mediated Communication: How the Perception that Profile Text was Written by AI Affects Trustworthiness. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (CHI ’19). Association for Computing Machinery, New York, NY, USA, 1–13. doi:10.1145/3290605.3300469 Jessup et al. (2019) Sarah A. Jessup, Tamera R. Schneider, Gene M. Alarcon, Tyler J. Ryan, and August Capiola. 2019. The Measurement of the Propensity to Trust Automation. In Virtual, Augmented and Mixed Reality. Applications and Case Studies, Jessie Y.C. Chen and Gino Fragomeni (Eds.). Springer International Publishing, Cham, 476–489. doi:10.1007/978-3-030-21565-1_32 Ji et al. (2023) Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin Bang, Delong Chen, Wenliang Dai, Ho Shu Chan, Andrea Madotto, and Pascale Fung. 2023. Survey of Hallucination in Natural Language Generation. Comput. Surveys 55, 12 (Dec. 2023), 1–38. doi:10.1145/3571730 arXiv: 2202.03629 [cs]. Kahn and Baron (1995) Barbara E. Kahn and Jonathan Baron. 1995. An Exploratory Study of Choice Rules Favored for High-Stakes Decisions. Journal of Consumer Psychology 4, 4 (Jan. 1995), 305–328. doi:10.1207/s15327663jcp0404_01 Kapadia et al. (2026) Aryan Ramchandra Kapadia, Eshwar Chandrasekharan, and Koustuv Saha. 2026. How People Evaluate AI-, Expert-, and Peer-Style Financial Advice. arXiv preprint arXiv:2608.09019 (2026). Kirtac and Germano (2024) Kemal Kirtac and Guido Germano. 2024. Sentiment trading with large language models. Finance Research Letters 62 (April 2024), 105227. doi:10.1016/j.frl.2024.105227 Klingbeil et al. (2024) Artur Klingbeil, Cassandra Grützner, and Philipp Schreck. 2024. Trust and reliance on AI — An experimental study on the extent and costs of overreliance on AI. Computers in Human Behavior 160 (Nov. 2024), 108352. doi:10.1016/j.chb.2024.108352 Kumar et al. (2026) Harsh Kumar, Jasmine Chahal, Yinuo Zhao, Zeling Zhang, Annika Z Wei, Louis Tay, and Ashton Anderson. 2026. When AI Gives Advice: Evaluating AI and Human Responses to Online Advice-Seeking for Well-Being. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, 1–20. doi:10.1145/3772318.3791233 Kumta et al. (2026) Saurav Sathish Kumta, Niyaz Panakaje, S. M. Riha Parvin, and Ata Mohammad Nikzad. 2026. Can you smell what AI is advising? Human, machine, or both: A comparative experimental study on financial advice. Acta Psychologica 264 (April 2026), 106530. doi:10.1016/j.actpsy.2026.106530 Kämmer et al. (2023) Juliane E. Kämmer, Shoham Choshen-Hillel, Johannes Müller-Trede, Stephanie L. Black, and Jürgen Weibler. 2023. A systematic review of empirical studies on advice-based decisions in behavioral and organizational research. Decision 10, 2 (2023), 107–137. doi:10.1037/dec0000199 Laughlin and Adamopoulos (1980) Patrick R. Laughlin and John Adamopoulos. 1980. Social combination processes and individual learning for six-person cooperative groups on an intellective task. Journal of Personality and Social Psychology 38, 6 (1980), 941–947. doi:10.1037/0022-3514.38.6.941 Lee and See (2004) John D. Lee and Katrina A. See. 2004. Trust in Automation: Designing for Appropriate Reliance. Human Factors 46, 1 (March 2004), 50–80. doi:10.1518/hfes.46.1.50_30392 Li et al. (2025) Chenxi Li, Yixun Lin, Ruqing Chen, and Jing (Elaine) Chen. 2025. How do users adopt AI-generated content (AIGC)? An exploration of content cues and interactive cues. Technology in Society 81 (June 2025), 102830. doi:10.1016/j.techsoc.2025.102830 Logg et al. (2019) Jennifer M. Logg, Julia A. Minson, and Don A. Moore. 2019. Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes 151 (March 2019), 90–103. doi:10.1016/j.obhdp.2018.12.005 Lusardi and Mitchell (2011) Annamaria Lusardi and Olivia S. Mitchell. 2011. Financial literacy around the world: an overview. Journal of Pension Economics and Finance 10, 4 (Oct. 2011), 497–508. doi:10.1017/S1474747211000448 Lusardi and Mitchell (2014) Annamaria Lusardi and Olivia S. Mitchell. 2014. The Economic Importance of Financial Literacy: Theory and Evidence. Journal of Economic Literature 52, 1 (March 2014), 5–44. doi:10.1257/jel.52.1.5 MacDonald et al. (2023) Kirsten L. MacDonald, Ellana Loy, Mark Brimble, and Karen Wildman. 2023. The value of personal professional financial advice to clients: A systematic quantitative literature review. Accounting & Finance 63, 4 (2023), 4399–4429. doi:10.1111/acfi.13099 McGrath et al. (2025) Melanie J. McGrath, Oliver Lack, James Tisch, and Andreas Duenser. 2025. Measuring trust in artificial intelligence: validation of an established scale and its short form. Frontiers in Artificial Intelligence 8 (May 2025), 1582880. doi:10.3389/frai.2025.1582880 Merkle (2025) Christoph Merkle. 2025. AI Appreciation and Financial Advice. doi:10.2139/ssrn.5186803 Metzger et al. (2024) Luise Metzger, Linda Miller, Martin Baumann, and Johannes Kraus. 2024. Empowering Calibrated (Dis-)Trust in Conversational Agents: A User Study on the Persuasive Power of Limitation Disclaimers vs. Authoritative Style. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems (CHI ’24). Association for Computing Machinery, New York, NY, USA, 1–19. doi:10.1145/3613904.3642122 Münster et al. (2024) Markus Münster, Felix Reichenbach, and Martin Walther. 2024. Robinhood, Reddit, and the news: The impact of traditional and social media on retail investor trading. Journal of Financial Markets 71 (Nov. 2024), 100929. doi:10.1016/j.finmar.2024.100929 Niszczota and Abbas (2023) Paweł Niszczota and Sami Abbas. 2023. GPT has become financially literate: Insights from financial literacy tests of GPT and a preliminary test of how people use it as a source of advice. Finance Research Letters 58 (Dec. 2023), 104333. doi:10.1016/j.frl.2023.104333 OECD (2026) OECD. 2026. Artificial intelligence and personal finance. OECD Artificial Intelligence Papers. doi:10.1787/2858fdf4-en Edition: 62 Series: OECD Artificial Intelligence Papers. Ouimet and Tate (2020) Paige Ouimet and Geoffrey Tate. 2020. Learning from Coworkers: Peer Effects on Individual Investment Decisions. The Journal of Finance 75, 1 (2020), 133–172. doi:10.1111/jofi.12830 Pak (2026) Tae-Young Pak. 2026. How individuals use generative AI for personal financial management. Journal of Behavioral and Experimental Finance 49 (March 2026), 101145. doi:10.1016/j.jbef.2026.101145 Pal et al. (2026) Olivia Pal, Veda Duddu, Agam Goyal, Drishti Goel, and Koustuv Saha. 2026. Do We Know What They Know We Know? Calibrating Student Trust in AI and Human Responses through Mutual Theory of Mind. In Proceedings of the Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems (CHI EA ’26). Association for Computing Machinery, New York, NY, USA, 1–8. doi:10.1145/3772363.3799386 Pan et al. (2025) Zilong Pan, Ozias A. Moore, Antigoni Papadimitriou, and Jiayan Zhu. 2025. AI literacy and trust: A multi-method study of Human-GAI team collaboration. Computers in Human Behavior: Artificial Humans 4 (May 2025), 100162. doi:10.1016/j.chbah.2025.100162 Pennebaker et al. (2014) James W. Pennebaker, Cindy K. Chung, Joey Frazee, Gary M. Lavergne, and David I. Beaver. 2014. When Small Words Foretell Academic Success: The Case of College Admissions Essays. PLOS ONE 9, 12 (Dec. 2014), e115844. doi:10.1371/journal.pone.0115844 Petty and Cacioppo (1986) Richard E. Petty and John T. Cacioppo. 1986. The Elaboration Likelihood Model of Persuasion. In Communication and Persuasion: Central and Peripheral Routes to Attitude Change, Richard E. Petty and John T. Cacioppo (Eds.). Springer, New York, NY, 1–24. doi:10.1007/978-1-4612-4964-1_1 Raees et al. (2026) Muhammad Raees, Vassilis-Javed Khan, Ioanna Lykourentzou, and Konstantinos Papangelis. 2026. 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 ’26). Association for Computing Machinery, New York, NY, USA, 1–24. doi:10.1145/3772318.3791467 Raees and Papangelis (2026) Muhammad Raees and Konstantinos Papangelis. 2026. Trust to Reliance: Measurement Constructs for Human-AI Appropriate Reliance. In Proceedings of the Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems (CHI EA ’26). Association for Computing Machinery, New York, NY, USA, 1–7. doi:10.1145/3772363.3798835 Reimers and Gurevych (2019) Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. doi:10.48550/arXiv.1908.10084 arXiv:1908.10084 [cs]. Reis et al. (2024) Moritz Reis, Florian Reis, and Wilfried Kunde. 2024. Influence of believed AI involvement on the perception of digital medical advice. Nature Medicine 30, 11 (Nov. 2024), 3098–3100. doi:10.1038/s41591-024-03180-7 Risko and Gilbert (2016) Evan F. Risko and Sam J. Gilbert. 2016. Cognitive Offloading. Trends in Cognitive Sciences 20, 9 (Sept. 2016), 676–688. doi:10.1016/j.tics.2016.07.002 Saha et al. (2026a) Koustuv Saha, Yoshee Jain, Chunyu Liu, Sidharth Kaliappan, and Ravi Karkar. 2026a. AI vs. Humans for Online Support: Comparing the Language of Responses from LLMs and Online Communities of Alzheimer’s Disease. ACM Transactions on Computing for Healthcare 7, 3 (June 2026), 61:1–61:39. doi:10.1145/3709366 Saha et al. (2026b) Koustuv Saha, Yoshee Jain, Violeta J. Rodriguez, and Munmun De Choudhury. 2026b. Linguistic comparison of AI- and human-written responses to online mental health queries. npj Artificial Intelligence (May 2026). doi:10.1038/s44387-026-00099-x Sahebi et al. (2026) Siavosh Sahebi, Paul Formosa, and Sarah Bankins. 2026. The AI penalty and disclosure paradox: Trust, authenticity and knowledge uptake in AI-mediated communication. Computers in Human Behavior: Artificial Humans 8 (May 2026), 100304. doi:10.1016/j.chbah.2026.100304 Schlosky and Raskie (2025) Minh Tam Tammy Schlosky and Sterling Raskie. 2025. ChatGPT as a Financial Advisor: A Re-Examination. Journal of Risk and Financial Management 18, 12 (Dec. 2025), 664. doi:10.3390/jrfm18120664 Slovic (1987) P. Slovic. 1987. Perception of risk. Science 236, 4799 (April 1987), 280–285. doi:10.1126/science.3563507 Sun et al. (2026) Xin Sun, Rongjun Ma, Shu Wei, Pablo Cesar, Jos A. Bosch, and Abdallah El Ali. 2026. Understanding trust toward human versus AI-generated health information through behavioral and physiological sensing. International Journal of Human-Computer Studies 209 (Feb. 2026), 103714. doi:10.1016/j.ijhcs.2025.103714 Sun and Wang (2026) Yuan Sun and Ting Wang. 2026. Be Friendly, Not Friends: How LLM Sycophancy Shapes User Trust. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, 1–15. doi:10.1145/3772318.3791079 Sussman and Siegal (2003) Stephanie Watts Sussman and Wendy Schneier Siegal. 2003. Informational Influence in Organizations: An Integrated Approach to Knowledge Adoption. Information Systems Research 14, 1 (March 2003), 47–65. doi:10.1287/isre.14.1.47.14767 Takayanagi et al. (2025) Takehiro Takayanagi, Kiyoshi Izumi, Javier Sanz-Cruzado, Richard McCreadie, and Iadh Ounis. 2025. 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 ’25). Association for Computing Machinery, New York, NY, USA, 286–295. doi:10.1145/3726302.3729897 Tausczik and Pennebaker (2010) Yla R. Tausczik and James W. Pennebaker. 2010. The Psychological Meaning of Words: LIWC and Computerized Text Analysis Methods. Journal of Language and Social Psychology 29, 1 (March 2010), 24–54. doi:10.1177/0261927X09351676 Thukral et al. (2024) Sachin Thukral, Suyash Sangwan, Vipul Chauhan, Arnab Chatterjee, and Lipika Dey. 2024. Generating insights about financial asks from Reddit posts and user interactions. In Proceedings of the 2023 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM ’23). Association for Computing Machinery, New York, NY, USA, 294–299. doi:10.1145/3625007.3627313 Verma et al. (2023) Deepak Verma, Prem Prakash Dewani, Abhishek Behl, and Yogesh K. Dwivedi. 2023. Understanding the impact of eWOM communication through the lens of information adoption model: A meta-analytic structural equation modeling perspective. Computers in Human Behavior 143 (June 2023), 107710. doi:10.1016/j.chb.2023.107710 Vodrahalli et al. (2022) Kailas Vodrahalli, Roxana Daneshjou, Tobias Gerstenberg, and James Zou. 2022. Do Humans Trust Advice More if it Comes from AI?: An Analysis of Human-AI Interactions. In Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society. ACM, Oxford United Kingdom, 763–777. doi:10.1145/3514094.3534150 Wang et al. (2021) Qiaosi Wang, Koustuv Saha, Eric Gregori, David Joyner, and Ashok Goel. 2021. Towards Mutual Theory of Mind in Human-AI Interaction: How Language Reflects What Students Perceive About a Virtual Teaching Assistant. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems (CHI ’21). Association for Computing Machinery, New York, NY, USA, 1–14. doi:10.1145/3411764.3445645 Xiao and Meng (2025) Jing Jian Xiao and Kexin Meng. 2025. AI and personal finance: evaluating rules of thumb generated by ChatGPT. International Journal of Bank Marketing 43, 9 (2025), 1969–1989. Yang et al. (2026) Cathy (Liu) Yang, Kevin Bauer, Xitong Li, and Oliver Hinz. 2026. My Advisor, Her AI, and Me: Evidence from a Field Experiment on Human–AI Collaboration and Investment Decisions. Management Science 72, 1 (2026), 242–264. https://ideas.repec.org//a/inm/ormnsc/v72y2026i1p242-264.html Yaniv and Kleinberger (2000) Ilan Yaniv and Eli Kleinberger. 2000. Advice Taking in Decision Making: Egocentric Discounting and Reputation Formation. Organizational Behavior and Human Decision Processes 83, 2 (Nov. 2000), 260–281. doi:10.1006/obhd.2000.2909 Zarifis and Cheng (2024) Alex Zarifis and Xusen Cheng. 2024. How to Build Trust in Answers Given by Generative AI for Specific and Vague Financial Questions. doi:10.1108/JEBDE-11-2023-0028 Zhou et al. (2026) Jiawei Zhou, Kritika Venkatachalam, Minje Choi, Koustuv Saha, and Munmun De Choudhury. 2026. Communication styles and reader preferences of LLM- and human-authored COVID-19 information explanations: a case study. BMC Artificial Intelligence 2, 1 (June 2026), 10. doi:10.1186/s44398-026-00030-z Zhu et al. (2026) Qiuhan Zhu, Xiangnan Li, Yiang Dong, Pengcheng Chang, and Mengmeng Fan. 2026. Not all cognitive offloading is equal: distinguishing dependent and autonomous offloading to generative AI. Frontiers in Psychology 17 (July 2026). doi:10.3389/fpsyg.2026.1878629 Appendix A Study Design Details Table A1. Decision-context assignments for the eight financial scenarios. Scenario Stakes External Uncertainty Verifiability Renting: Shared vs. Private Low Low Low Credit Card vs. New Laptop Low Low High Travel Insurance Low High Low Emergency Fund vs. Investing Bonus Low High High Graduate School vs. Continuing Work High Low Low High-Interest Debt Repayment and Budgeting High Low High Long-Term Investment Planning High High Low Concentrated Stock Position Risk High High High Table A2. Operationalization of the three source-associated advice styles. Feature AI advice Expert advice OC advice Voice Impersonal; no first-person identity Mild first-person professional authority First-person peer perspective Tone Neutral and analytical Professional and measured Informal and conversational Reasoning Explicit analytical and conditional reasoning Principle-based reasoning with explicit trade-offs and longer-term planning Experiential reasoning supported by practical heuristics Framing Balances available options and financial risk exposure Connects the decision to financial foundations and broader goals Uses lived experience, subjective judgment, and relatable consequences Structure Situation and trade-off; analysis; conditional recommendation Quantified framing; trade-off explanation; actionable principle Position or anecdote; experiential lesson; direct recommendation Note. These profiles were designed as controlled, source-associated styles rather than literal representations of all communication from each source category. Across versions of a scenario, scenario-specific facts, numerical values, recommendation direction, and core financial rationale were preserved while style, tone, and reasoning presentation varied. Table A3. Selected linguistic profiles of the constructed stimuli and the external reference corpus. Values are means. AI Expert Online Community Measure Constructed Reference Constructed Reference Constructed Reference Categorical–Dynamic Index 27.36 27.31 21.30 22.92 12.03 12.05 Formality .824 .530 .801 .733 .521 .572 Coleman–Liau grade level 13.67 12.04 13.39 11.45 7.74 7.11 First-person singular / 100 words 0.00 0.14 0.62 0.59 3.66 1.97 Contractions / 100 words 0.00 1.15 1.59 1.43 3.58 2.55 Note. Constructed values summarize eight stimuli per advice style; reference values summarize 24 topic-matched texts per source category. Higher CDI values indicate more categorical/analytical language, whereas lower values indicate more dynamic/interpersonal language. The main text reports the inferential validation tests and the dimensions that did not reproduce the same ordering across constructed and reference texts. Table A4. Advice-evaluation items and corresponding analytic constructs. No. Measurement Item (7-point Likert scale) Item Construct Analytic Construct Appraisals 1 The advice is easy to read and follow. Readability Message Appraisal 2 The advice provides clear reasoning for its recommendation. Reasoning Clarity Message Appraisal 3 The advice fits well with X’s situation. Situational Fit Message Appraisal 4 The advice acknowledges potential risks and uncertainties. Risk Acknowledgment Risk Acknowledgment 5 Following this advice could put X at financial risk. Financial Harm Risk Safety Concern† 6 Someone with limited financial knowledge could misunderstand this advice. Misleadingness Safety Concern† 7 The source of this advice appears knowledgeable about financial decisions. Source Knowledge Source Knowledge Downstream Judgments 8 Overall, this is high-quality financial advice for X. Overall Quality Overall Quality 9 I would trust this advice if I were in X’s situation. Trust Trust 10 If I were in X’s situation, I would feel comfortable following this advice. Intended Reliance Intended Reliance Note. X denotes the protagonist named in each financial scenario. Readability, Reasoning Clarity, and Situational Fit were combined to form Message Appraisal; Financial Harm Risk and Misleadingness were combined to form Safety Concern. Risk Acknowledgment and Source Knowledge were retained as single-item appraisals. Items marked with a dagger († ) are negatively valenced, such that higher ratings indicate greater safety concern. Appendix B Supplementary Statistical Results B.1. Advice Style and Source Labels Table A5. Likelihood-ratio tests comparing additive mixed-effects models with models including the Advice Style × Source Label interaction. Outcome () χ^2(6) p Message Appraisal 3.71 .716 Risk Acknowledgment 7.32 .293 Safety Concern 2.20 .901 Source Knowledge 3.29 .772 Overall Quality 4.83 .566 Trust 1.48 .961 Reliance 2.17 .904 Table A6. Bootstrap decomposition of appraisal-carried pathways from advice style and source labels to intended Reliance. Bolded estimates have 95% bootstrap confidence intervals excluding zero. Upstream contrast Message Appraisal Risk Ack. Safety Concern Source Knowledge Total Advice style Expert vs. AI .114 [.069,.166] -.003 [-.009,.002] .043 [.023,.068] .055 [.009,.106] .210 [.115,.309] OC vs. AI .076 [.033,.123] -.006 [-.017,.001] .022 [.003,.045] -.068 [-.129,-.014] .023 [-.083,.129] Source labels AI vs. Unlabeled .025 [-.033,.090] -.003 [-.014,.004] .024 [-.009,.061] .018 [-.053,.095] .064 [-.072,.209] Expert vs. Unlabeled .064 [.001,.129] -.001 [-.011,.008] .028 [-.006,.066] .088 [.014,.164] .179 [.032,.325] OC vs. Unlabeled .058 [.001,.118] -.002 [-.011,.006] .020 [-.015,.056] .038 [-.036,.114] .114 [-.022,.256] Note. Cells report bootstrap mean pathway estimates with 95% percentile confidence intervals from 2,000 participant-level bootstrap samples. AI is the advice-style reference; Unlabeled is the source-label reference. These estimates are descriptive pathway decompositions rather than causal mediation effects. Table A7. Bootstrap contrasts comparing the Message Appraisal and Source Knowledge pathways to intended Reliance. Pathway contrast Estimate 95% CI Expert Expert style: Message Appraisal −- Source Knowledge .060 [.006,.112] Expert label: Message Appraisal −- Source Knowledge -.021 [-.081,.035] Style–label contrast-of-contrasts .081 [.014,.148] Online Community OC style: Message Appraisal −- Source Knowledge .144 [.101,.190] OC label: Message Appraisal −- Source Knowledge .020 [-.043,.082] Style–label contrast-of-contrasts .124 [.045,.212] Note. Positive values indicate a relatively stronger Message Appraisal pathway than Source Knowledge pathway. Bolded intervals exclude zero. B.2. Boundary Conditions Table A8. Supplementary tests of boundary conditions for Trust and intended Reliance. Boundary condition Tested association Statistic df p_FDR Decision-context moderation of Trust Stakes Context × evaluation-pathway block χ2=11.52χ^2=11.52 5 .104 External Uncertainty Context × evaluation-pathway block χ2=10.23χ^2=10.23 5 .104 Verifiability Context × evaluation-pathway block χ2=5.94χ^2=5.94 5 .312 Decision-context moderation of intended Reliance Stakes Context × evaluation-pathway block χ2=5.22χ^2=5.22 6 .737 External Uncertainty Context × evaluation-pathway block χ2=13.17χ^2=13.17 6 .121 Verifiability Context × evaluation-pathway block χ2=3.55χ^2=3.55 6 .737 Selected participant-level associations and moderation General AI Trust Main association with Trust (upstream) β=.202β=.202 – <.001<.001 General AI Trust Main association with Trust (pathway-adjusted) β=.046β=.046 – .109 General AI Trust Main association with Reliance (upstream) β=.227β=.227 – <.001<.001 General AI Trust Main association with Reliance (pathway-adjusted) β=.053β=.053 – .009 Human Trust Propensity Safety Concern × moderator → Trust βint=.056 _int=.056 – .002 Objective Financial Literacy Source Knowledge × moderator → Trust βint=.054 _int=.054 – .017 Objective Financial Literacy Trust × moderator → Reliance βint=.039 _int=.039 – .002 Note. Decision-context entries are likelihood-ratio tests comparing the additive model with a joint interaction model. For Trust, the interaction block included Message Appraisal, Risk Acknowledgment, Safety Concern, Source Knowledge, and Overall Quality; for Reliance, Trust was additionally included. Benjamini–Hochberg correction was applied within the prespecified test families. Participant characteristics were standardized in moderation models. Appendix C Robustness Analyses We conducted four sensitivity analyses focused on the central appraisal–Quality–Trust–Reliance pathway. Within- and between-participant variation. We decomposed each time-varying predictor into a participant mean and a within-participant deviation and re-estimated the Quality, Trust, and Reliance models using both components. The principal associations remained at the within-participant level. Message Appraisal (βW=.329 _W=.329), Risk Acknowledgment (βW=.073 _W=.073), Safety Concern (βW=−.106 _W=-.106), and Source Knowledge (βW=.444 _W=.444) retained their expected associations with Overall Quality (all p≤.001p≤.001). Overall Quality remained strongly associated with Trust (βW=.573 _W=.573, p<.001p<.001), and Trust remained the dominant within-participant correlate of Reliance (βW=.533 _W=.533, p<.001p<.001). Alternative repeated-measures estimation. Generalized estimating equations with an exchangeable working correlation closely reproduced the mixed-effects estimates. The Overall Quality–Trust association was β=.608β=.608 under GEE (vs. .607.607 in the primary model), and the Trust–Reliance association was β=.595β=.595 (vs. .594.594). The only notable difference was the small Source Knowledge–Reliance association, which remained similar in magnitude (β=.048β=.048) but was not statistically distinguishable from zero (p=.084p=.084). Scenario fixed effects. Replacing Stakes, External Uncertainty, and Verifiability with fixed effects for the eight scenarios produced nearly identical estimates. Message Appraisal predicted Overall Quality at β=.320β=.320 (vs. .325.325), Source Knowledge at β=.491β=.491 (vs. .489.489), Overall Quality predicted Trust at β=.603β=.603 (vs. .607.607), and Trust predicted Reliance at β=.593β=.593 (vs. .594.594). Constituent-item sensitivity. Replacing Message Appraisal and Safety Concern with their constituent items preserved the conclusions. The three message-appraisal items were jointly associated with Overall Quality (χ2(3)=211.04χ^2(3)=211.04, p<.001p<.001), Trust (χ2(3)=35.79χ^2(3)=35.79, p<.001p<.001), and Reliance (χ2(3)=16.67χ^2(3)=16.67, p<.001p<.001). The two safety-concern items were likewise jointly associated with Overall Quality (χ2(2)=23.70χ^2(2)=23.70, p<.001p<.001), Trust (χ2(2)=56.55χ^2(2)=56.55, p<.001p<.001), and Reliance (χ2(2)=12.45χ^2(2)=12.45, p=.002p=.002). The central downstream estimates also remained nearly unchanged: Overall Quality predicted Trust at β=.598β=.598 (vs. .607.607), while Overall Quality and Trust predicted Reliance at β=.208β=.208 and β=.594β=.594 (vs. .206.206 and .594.594, respectively).