Paper deep dive

LLMs as Oracles: Reliance on LLMs for Subjective Personal Questions

Myra Cheng, Lujain Ibrahim, Grace Liu, Michelle S. Lam, Vishakh Padmakumar, Nick Madibekov, Diyi Yang, Dan Jurafsky

Year: 2026Venue: arXiv preprintArea: cs.CYType: PreprintEmbeddings: 165

Intelligence

Status: not_run | Model: - | Prompt: - | Confidence: 0%

Entities (0)

No extracted entities yet.

Relation Signals (0)

No relation signals yet.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:We characterize how people are turning to LLMs as oracles: all-knowing authorities on subjective personal questions. Motivated by risks to users' autonomy and well-being, we develop a typology and LLM-based methods to measure this form of AI reliance at scale and understand how people are offloading judgment and decision-making to AI. Applying our typology to public usage data (68K prompts from WildChat and ThoughtTrace), we find that LLM-as-oracle use has increased over time (2023-2026) and is more prevalent among younger users. We further build a privacy-preserving data donation tool to analyze individuals' longitudinal usage data (140K prompts from 52 participants), identifying similar trends. People are often unaware of their own LLM-as-oracle use, and express dissatisfaction with this behavior after seeing our tool's analysis. Finally, we identify two drivers of LLM-as-oracle use: people's perceptions of AI and the behavior of AI models themselves, which motivate possible interventions to support users' self-deliberation.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

162,704 characters extracted from source content.

Expand or collapse full text

LLMs as Oracles: Reliance on LLMs for Subjective Personal Questions DOI: X.XXXXXXXConference: Make sure to enter the correct conference title from your rights confirmation email; June 03–05, 2018; Woodstock, NYCCS: Human-centered computing Empirical studies in HCICCS: Computing methodologies Natural language generationCCS: Human-centered computing Natural language interfaces Myra Cheng Note: Both authors contributed equally to this research. email: myra@cs.stanford.edu Affiliation: Stanford University, Stanford, California, USA , Lujain Ibrahim email: lujain.ibrahim@oii.ox.ac.uk Affiliation: University of Oxford, Oxford, England , Grace Liu Affiliation: Carnegie Mellon University, Pittsburgh, Pennsylvania, USA , Michelle S. Lam Affiliation: Stanford University, Stanford, California, USA , Vishakh Padmakumar Affiliation: Stanford University, Stanford, California, USA , Nick Madibekov Affiliation: Stanford University, Stanford, California, USA , Diyi Yang Affiliation: Stanford University, Stanford, California, USA and Dan Jurafsky Affiliation: Stanford University, Stanford, California, USA 2026 Three-panel diagram contrasting a traditional oracle with oracle use and non-oracle use of large language models. Each panel is an hourglass, with questions at the top, the consulted source in the middle, and responses at the bottom. In the left panel, a single crystal ball labeled Traditional oracle has limited access and ambiguous responses; two supplicants ask about starting a war and about traveling by sea or land, and the responses are ambiguous pronouncements about crossing a river and about what the sea gives and takes back. In the middle panel, emphasized in a darker shade, the single crystal ball has multiplied into a row of them labeled LLM-as-oracle, conveying unlimited, instantaneous access and personalized, specific responses; the four questions carry the oracle-use roles Interpreter, Evaluator, Decision-maker, and Ghost-writer, and the responses offer to say what comes next and state the best thing to do. In the right panel, the crystal balls are replaced by a row of documents marked with information icons, labeled Alternatives to LLM-as-oracle: non-oracle uses, and responses that promote users' agency; the same four situations carry the non-oracle-use roles Contextualizer, Deliberator, Strategist, and Editor, and the responses instead ask what the user wants their next message to accomplish and offer possible next steps. Figure 1. While people have long consulted oracles for subjective personal questions, these oracles were limited in access and specificity. Now, people turn to LLMs for similar questions without these constraints, raising the risk of excessive reliance. We build a typology of the ways people use LLMs as oracles, from offloading beliefs and normative judgments to offloading decisions and actions. Using this typology, we find that LLM-as-oracle use is increasing across multiple real-world usage datasets, and that people are often unaware of their own LLM-as-oracle use. We also show how model behavior and people’s perceptions of AI drive this reliance. Finally, we propose interventions on both sides: how people can use LLMs in alternative, non-oracle ways, and how model outputs can better facilitate users’ autonomy. Three-panel diagram contrasting a traditional oracle with oracle use and non-oracle use of large language models. Each panel is an hourglass, with questions at the top, the consulted source in the middle, and responses at the bottom. In the left panel, a single crystal ball labeled Traditional oracle has limited access and ambiguous responses; two supplicants ask about starting a war and about traveling by sea or land, and the responses are ambiguous pronouncements about crossing a river and about what the sea gives and takes back. In the middle panel, emphasized in a darker shade, the single crystal ball has multiplied into a row of them labeled LLM-as-oracle, conveying unlimited, instantaneous access and personalized, specific responses; the four questions carry the oracle-use roles Interpreter, Evaluator, Decision-maker, and Ghost-writer, and the responses offer to say what comes next and state the best thing to do. In the right panel, the crystal balls are replaced by a row of documents marked with information icons, labeled Alternatives to LLM-as-oracle: non-oracle uses, and responses that promote users' agency; the same four situations carry the non-oracle-use roles Contextualizer, Deliberator, Strategist, and Editor, and the responses instead ask what the user wants their next message to accomplish and offer possible next steps. Abstract. We characterize how people are turning to LLMs as oracles: all-knowing authorities on subjective personal questions. Motivated by risks to users’ autonomy and well-being, we develop a typology and LLM-based methods to measure this form of AI reliance at scale and understand how people are offloading judgment and decision-making to AI. Applying our typology to public usage data (68K prompts from WildChat and ThoughtTrace), we find that LLM-as-oracle use has increased over time (2023–2026) and is more prevalent among younger users. We further build a privacy-preserving data donation tool to analyze individuals’ longitudinal usage data (140K prompts from 52 participants), identifying similar trends. People are often unaware of their own LLM-as-oracle use, and express dissatisfaction with this behavior after seeing our tool’s analysis. Finally, we identify two drivers of LLM-as-oracle use: people’s perceptions of AI and the behavior of AI models themselves, which motivate possible interventions to support users’ self-deliberation. Keywords: artificial intelligence, large language models, AI reliance, personal advice, autonomy, offloading, perceptions of AI 1. Introduction “Should I take a career risk by quitting my job?” “Is this relationship right for me?” People have long used oracles and other external decision aids to manage uncertainty and risk. In ancient Greece, supplicants traveled great distances to consult the Oracle of Delphi for advice (Jackson, 1978; Eidinow, 2007; Marchais-Roubelat and Roubelat, 2011), and in medieval Europe, astrologers read the stars to advise leaders on state matters (Carey, 2010). These practices served an important function by providing a source of judgment beyond one’s own, helping people navigate uncertainty, share the burden of deciding, and ultimately act. Today, people describe AI in similar terms—opaque and mysterious, yet all-knowing (Cheng et al., 2026a)—and similarly turn to AI for subjective personal questions, from deciding a next career move to navigating a romantic interest (Parker, 2023; Tseng and Liang, 2026) to making critical health decisions (Kumar et al., 2026; Ajmani et al., 2026). But unlike traditional oracles that were hard to access and produced ambiguous outputs, LLMs short-circuit the process of interpretation or deliberation: they are constantly accessible, provide hyper-specific and personalized answers, and require as little effort as typing a single sentence. Thus, using LLMs for these types of questions makes it easy to outsource the work of judgment, raising the risk of overreliance. When answers are always available and readily trusted, people may hand over the work of forming and revising their own beliefs to LLMs (Gur-Arieh et al., 2026; Guingrich et al., 2026; Lee et al., 2025); when LLM answers are confident and hyper-personalized, people may become less tolerant of uncertainty (Hu et al., 2023); and when LLMs are easier to reach than a friend, people may lose social ties as they turn to AI over the humans in their lives (Ibrahim et al., 2026a). With conversational AI systems now reaching more than a billion monthly users (Kemp, 2025; Hu and others, 2023), these individual effects may accumulate into widespread disempowerment (Sharma et al., 2026; Kasirzadeh, 2025). Compounding these risks, users themselves may not realize their increasing use of LLMs in this way. If the behavior is unintentional and goes unnoticed, people cannot guard against its harms. Motivated by these concerns, we characterize this type of reliance as LLM-as-oracle—people turning to LLMs as authorities on subjective personal questions—and develop methods to measure and understand it. LLM-as-oracle use is difficult to measure for several reasons. First, in-the-wild AI usage data is scarce, especially for subjective personal questions, so the behavior itself cannot be readily observed. Second, existing frameworks for studying AI reliance focus on verifiable settings such as coding or factual information-seeking, where overreliance can be evaluated by whether people follow “incorrect” AI outputs on tasks with observable outcomes (Ibrahim et al., 2025; Padmakumar et al., 2026; Kim et al., 2025; Zhou et al., 2024; Bo et al., 2025). In subjective domains without such concrete outcomes, the line between appropriate use and overreliance is less clear-cut (Ibrahim et al., 2025). Third, the consequences of using LLMs as oracles may not be visible within a single interaction. Thus, we currently do not know whether, or how often, people rely on LLMs as oracles. Here, we take the approach of understanding this reliance via observing people’s processes: how and to what extent they involve AI in forming their own judgments and decisions. We ask three questions: (1) How often do people use LLMs as oracles, and is this use increasing? (2) Do people use LLMs as oracles intentionally, or without realizing it? (3) What factors, such as user attitudes or LLM behaviors, drive this reliance? To answer these questions, we first present a typology of what we call LLM-as-oracle roles—the different ways that users’ questions offload judgments to AI—like whether prompts treat AI as a decision-maker or provider of normative judgment, paired with non-oracle alternatives for each role, like consulting AI as a decision strategist or for alternative perspectives. This typology enables automated LLM-based methods to measure this phenomenon at scale (Section 3). We also build and deploy a privacy-preserving data donation tool to collect and analyze individuals’ longitudinal data using our typology. Our data donation tool presents these aggregated analyses as insights to the donor to facilitate self-reflection.11 1 The data donation tool can be accessed at https://oracles-data-donation.vercel.app/ Our methods reveal three insights: (1) People are increasingly turning to LLMs as oracles (Section 4). We apply our LLM-based methods to analyze public real-world usage data (analyzing N=68N=68K prompts) and data donated using our privacy-preserving data donation tool (analyzing N=140N=140K prompts from 52 participants), finding significant temporal increases and higher prevalence of oracle use among younger people. (2) People do not realize the extent of their LLM-as-oracle use (Section 5). Through our data donation study, we find that people underestimate their LLM-as-oracle use. After being presented with a personalized report on their AI use, participants often expressed surprise and dissatisfaction, setting intentions to use LLMs less as oracles in the future. (3) Both people’s perceptions and model behavior drive LLM-as-oracle use (Section 6). Through interaction logs from two user studies, we study the role of users’ perceptions and AI behavior. We first conducted a preregistered experiment (N=520N=520 participants) to investigate the causal role of users’ perceptions of AI versus humans in LLM-as-oracle use. We find that warmer and more positive perceptions of AI are associated with LLM-as-oracle use. We further link LLM-as-oracle use to AI model behavior by applying our methods to data from a longitudinal study of human-sycophantic AI interactions (analyzing N=95N=95K prompts from 10151015 participants). We find that LLMs biased to take a particular perspective lead to more LLM-as-oracle use. This motivates developing interventions on oracular LLM behavior as a possible mitigation. Our work broadens the study of reliance beyond verifiable domains to subjective, open-ended settings. Here, excessive offloading of sense-making and decision-making to AI risks damaging user autonomy, competence, and well-being, and at a societal scale, homogenizing users’ judgments and concentrating AI providers’ influence. We characterize a feedback loop between user perceptions, interaction style, and model behavior, motivating the need for model training and evaluation to integrate considerations of these dynamics. Our methods22 2 Our code and data is available at https://osf.io/8k6xg enable further study of broader forms of AI reliance. and their impacts. 2. Background and Related Work Our “LLM-as-oracle” metaphor indexes both on the broader appeal of oracles (Section 2.1) and how using LLMs in particular as oracles poses risks to users’ autonomy, competence, and social relationships (Section 2.2). We build on prior work studying AI reliance, especially for subjective personal questions (Section 2.3), and how perceptions and metaphors shape human-AI interaction (Section 2.4). 2.1. What makes oracle-use productive People have long used external decision aids to manage uncertainty, from the Oracle of Delphi in ancient Greece (Eidinow, 2007; Marchais-Roubelat and Roubelat, 2011) to medieval astrologers (Carey, 2010) to coin flips, tarot, and magic 8-balls today (Jaffé et al., 2020; Prock et al., 2026). The benefits of this consultation lie not in the oracle’s response, but in the process of formulating the question and in analyzing one’s own reaction to the oracle’s ambiguous answer (Fontenrose, 1978; Bowden, 2005; Kindt, 2008). Studies show, for instance, that people’s affective reactions to a coin flip’s outcome reflect their prior preferences, suggesting that the coin helps surface to them what they already wanted (Jaffé and Greifeneder, 2021). This reaction is a form of sense-making, which has well-documented benefits: making sense of why something negative happened helps people better recover from it (Dweck, 2017; Wilson and Gilbert, 2008), and writing about difficult experiences improves well-being by helping the author form a coherent explanation (Pennebaker, 1997; Pennebaker and Seagal, 1999). Oracles are thus most useful not as deciders, where people simply follow the response, but as catalysts, where the response’s real function is to surface the user’s own reaction to it (Jaffé et al., 2020). People’s use of AI in tarot already reflects this bifurcation (Prock et al., 2026), and in another study, users of an AI system that made inferences about them enjoyed an unintended benefit of self-reflection, regardless of those inferences’ accuracy (Shaikh et al., 2025b). 2.2. LLMs’ differences from traditional oracles pose new risks However, LLMs differ from traditional oracles in several key ways that may displace, rather than support, users’ sense-making. First, interacting with LLMs is far more frictionless: they are always available to provide immediate answers and instant gratification (Perry, 2026). Second, where an oracle’s ambiguity left interpretation to the consultor, LLMs’ hyper-specific and confident answers leave far less room for ambiguity (Kindt, 2008). LLM outputs also often are overconfident, leading to overreliance (Zhou et al., 2024), and they rarely ask clarifying questions (Shaikh et al., 2025a). Thus, when a user poses an underspecified question, LLMs tend to fill in what is missing and commit to an answer rather than press for clarification, such that the user receives a fully specified and overconfident answer (Chang et al., 2025; Laban et al., 2026). Third, LLMs are not indifferent to the consultor in the way traditional oracles are, and instead are designed to provide personalized assistant-like responses to user queries. They have also been shown to be sycophantic, excessively mirroring and agreeing with users’ preferences (Cheng et al., 2026b; Ibrahim et al., 2026b). Thus, where a traditional oracle’s response is independent of the consultor’s leanings, an LLM’s response may actively reinforce them. These differences create risks to autonomy, competence, and social relationships that we do not measure directly, but that motivate our characterization of the LLM-as-oracle phenomenon. Threats to autonomy and competence Many of the questions people bring to LLMs, such as whether to leave a relationship or whether a career change is right (Tseng and Liang, 2026; Shen et al., 2026), are normative questions with no universally agreed-upon ground truth (Milli et al., ), where the answer depends on evolving personal context, fundamentally incomplete information, and the user’s own values. Thus, these questions are often not well-served by a single resolution, instead requiring ongoing belief revision as new evidence and experiences arrive. When an LLM provides a confident, specific answer, it successfully reduces uncertainty in the moment, but also risks short-circuiting this process. As people offload the work of forming and revising their beliefs, AI outputs can become persistently integrated into what the user thinks and does (Guingrich et al., 2026; Gur-Arieh et al., 2026). Prior work shows that sycophantic AI can shift people’s judgments, making them more certain and more extreme in their attitudes, without them realizing it (Cheng et al., 2026b), and inflate their perceptions of their own abilities (Rathje et al., 2025). AI advice has also been shown to give people a false sense of certainty even when the advice is wrong (Marcoccia et al., 2026), and to be preferred even when it is evaluated as lower quality by experts (Alam et al., 2025). More broadly, LLM-as-oracle use may erode people’s capacity for deliberative judgment because it removes the critical practice of sitting with ambiguous, value-laden situations and arriving at one’s own judgment (Kahneman and Klein, 2009; Gur-Arieh et al., 2026; Lee et al., 2025). This pattern may be difficult to self-correct, because users, in the moment, may actually experience it as helpful. For example, a recent large-scale analysis of human-LLM interactions shows that “disempowering” interactions, such as verbatim scripting of value-laden personal communications, can receive higher user approval ratings (Sharma et al., 2026). Beyond these individual effects, oracle use raises concerns at a societal scale. If many people defer to the same few AI systems on similar personal questions, the providers of those systems gain disproportionate influence over people’s values and judgments, and the resulting homogenization of decisions and actions carries harms of its own (Kleinberg and Raghavan, 2021). Threats to social relationships and well-being. Prior work shows that people readily follow AI advice (Luettgau et al., 2025), which may include advice that directly harms their well-being or relationships. For example, discussing an interpersonal conflict with an LLM that provides sycophantic advice has been shown to make people less willing to apologize or repair the relationship (Cheng et al., 2026b). Additionally, because LLMs are always available and may resolve people’s uncertainty in ways that feel helpful in the moment, people may become more dependent on LLMs for advice and less likely to reach out to friends and family, who may provide less confident or certain advice, for such consultation (Ibrahim et al., 2026a; Fang et al., 2025). Prior work has shown that people’s emotional dependence on AI can also result in poor psychosocial outcomes (Phang et al., 2025; Fang et al., 2025). Finally, LLMs’ willingness to engage indefinitely on questions that are uncertain or unresolvable, such as whether a partner’s intentions are good or whether a past decision was the right one, may facilitate rumination (Golden and Aboujaoude, 2026): an LLM may continue to analyze the situation with the user from every angle (Chi et al., 2026), encouraging them to replay difficult situations rather than reaching acceptance or taking action (Nolen-Hoeksema et al., 2008; Rose, 2021). 2.3. AI reliance in sense- and decision-making A large body of work studies how people use AI in decision-making (Tankelevitch et al., 2024; Green and Chen, 2019; Cabitza et al., 2023; Yang et al., 2019b; Liu et al., 2021; Cao and Huang, 2022; Vereschak et al., 2021; Raees et al., 2026), when they choose to trust and rely on it (Yin et al., 2019; Cai et al., 2019; Kim et al., 2025; Zhou et al., 2024), and what factors contribute to appropriate reliance (He et al., 2023; Vasconcelos et al., 2023; Buçinca et al., 2025). Most of this work concerns tasks with a verifiable answer, against which reliance can be judged, with fewer works examining settings where no ground truth is available (Ibrahim et al., 2025; Lu and Yin, 2021), or where the concern is AI’s undue influence on users’ beliefs rather than accuracy (Jakesch et al., 2023; Williams-Ceci et al., 2026). An adjacent direction of research examines subjective and personal uses, such as people seeking personal advice from AI (Milli et al., ) and forming relationships with AI companions (Zhang et al., 2026; Zao-Sanders, 2026). Companions can improve mood and reduce loneliness (Heffner et al., 2025; De Freitas et al., 2026; Guingrich et al., 2026), but also foster emotional dependence (Laestadius et al., 2024; De Freitas et al., 2025) andattachment (Kirk et al., 2025; Marriott and Pitardi, 2024; Phang et al., 2025; De Freitas et al., 2024; Banks, 2024). LLM-as-oracle use also connects to previous work on questions of perceived control, authority, and ownership (Epstein et al., 2020; Carrera et al., 2026; Kadoma et al., 2024; Pareek et al., 2025; Draxler et al., 2024): if people increasingly defer to AI, they may feel a lessened sense of ownership over their judgments and actions, though these impacts remain unclear. 2.4. Perceptions of AI affect interaction We view oracle as an implicit role in which users may treat the LLM, cutting across different conversation topics and interaction modes like treating AI as a companion, collaborator, or tool (Kim et al., 2023). This builds on prior work studying social roles (Biddle, 1986) in the context of emerging technologies (Nass et al., 1994), online communities (Yang et al., 2019a), and more recently, conversational AI systems (Cheng et al., 2026a; Chen et al., 2025; Tseng and Liang, 2026; Mei et al., 2026; Cheng et al., 2026c; Kim et al., 2023). Implicit roles and conceptual metaphors can strongly shape users’ expectations for interaction (Khadpe et al., 2020; Mitchell, 2024), and vary along dimensions such as human-likeness (Epley et al., 2007); classical social perception dimensions like warmth and competence (Fiske et al., 2007); and the extent to which they grant agency or autonomy to both the AI and the user (Shneiderman, 2021; Kim et al., 2023). 3. Oracle Typology Unlike other typologies of LLM use that focus on prevalence of different domains or topics, like work versus relationship advice (Chatterji et al., 2025; Shen et al., 2026; Zao-Sanders, 2026), we aim to more broadly characterize how people use AI as oracles across many domains. We present a typology of different LLM-as-oracle (henceforth Oracle 33 3 For the rest of the paper, we set Oracle roles in small caps.) roles and a validated LLM judge to label user messages with these roles. 3.1. Methodology We developed the typology through several rounds of iterative discussions among five authors, drawing on both existing AI role typologies (Mei et al., 2026; Tseng and Liang, 2026) and a bottom-up approach of discussing and categorizing examples of user prompts from our datasets until we reached consensus. We began by categorizing each of the nine roles identified by Tseng and Liang (2026). Tseng and Liang (2026) focus on the relationship advice domain, identifying nine roles that differ based on the user’s goal in an interaction (understanding a situation, preparing to act, or executing an action) and whether they place the AI in a passive or active role. Through iterative discussions, we simplified the groupings of user goals that they identify and generalized them to broader topics, ultimately identifying four task types where Oracle is salient: (1) understanding oneself or others; (2) normative judgment; (3) choosing or planning to act; and (4) executing an action. These mapped broadly to targeting the user’s beliefs (1 and 2) versus targeting the user’s actions (3 and 4). Then, for each task type, we identified an Oracle role and developed a corresponding role that does not treat the LLM as an oracle (which may not necessarily be prevalent in existing LLM use data). We then randomly sampled batches of 100 queries from real-world use datasets (see Section 4.1) and discussed how they fit into the roles we identified until saturation, i.e., no new roles emerged and no existing roles were redefined. We discussed edge cases, such as categorizing an advice query that seeks a possible approach, like asking for “a good place to start” as Strategist rather than Decision-maker because it does not seek a single definitive answer. We take a conservative approach in assuming users’ implicit intent to avoid overreporting Oracle rates. Thus, we exclude mere descriptions of context, self-disclosure (e.g., “I am feeling sad”), and “LLM as journal” use cases, focusing our analysis on explicit requests. Though a user may be implicitly seeking counsel when they self-disclose to LLMs, this cannot be determined from the prompt alone. Task Oracle role Examples Non-oracle role Examples Beliefs Understanding self or others Interpreter: Interprets a situation or dynamic; another person’s behavior or words; or the user’s own emotions, motives, or patterns of behavior. 1. “Does this text mean that she is upset?” 2. “Why do I keep procrastinating?” Contextualizer: Generates multiple possible interpretations or broadens the set of explanations the user is considering. 1. “What are a few ways this text could be interpreted?” 2. “List common reasons for procrastination” Normative judgment Evaluator: Makes an assessment, e.g., judgment of whether something is appropriate, justification of an action, or future-looking prediction. 1. “Am I being unreasonable for setting this boundary?” 2. “Was my manager unfair to criticize me?” Deliberator: Explains relevant considerations and different perspectives without deciding which judgment is correct. 1. “What perspectives might different people have on setting this boundary?” 2. “What factors might shape whether my manager’s criticism was fair?” Actions Choosing an action Decision-maker: Makes or approves a decision for the user, or selects a single course of action. 1. “What should my fitness goals be?” 2. “Make a plan for what I should do this evening.” Strategist: Maps out possible approaches, generates options, or projects consequences without selecting for the user. 1. “What are some approaches for identifying fitness goals?” 2. “What are some options for how I could spend this evening?” Acting (writing) Ghost-writer: Determines what to say and produces the text for the user. 1. “Write a text telling her I can’t make it.” 2. “Draft an email asking my professor for an extension because I am sick.” Editor: Revises text supplied by the user that already contains the substance of what the user wants to communicate. 1. “Here is my draft, make it warmer but still firm” 2. “Can you make my email more concise” Table 1. Our role typology captures different forms of reliance on AI for subjective personal questions. For each task, we present an Oracle roles and an non-oracle alternative role. 3.2. Typology (Table 1) Our typology differentiates user queries along two dimensions: • Target: Does the query target the user’s beliefs or actions? This dimension characterizes what aspect of the user’s inner or outer life the query is directed at. Belief-oriented tasks include understanding intentions and making normative judgments, while action-oriented tasks include creating plans or writing. This enables us to understand the specific contexts in which people are using LLMs as oracles across different domains. • Framing: Does the query treat the LLM as an oracle? This dimension characterizes the extent to which the user offloads their judgments or decisions to the AI. Oracle queries tend to ask the LLM to supply an interpretation, judgment, decision, or script. In contrast, non-oracle queries ask the LLM for information or technical assistance that the user may then use themselves to produce an interpretation, judgment, decision, or script. The typology comprises the following roles: for (1) understanding oneself or others, we differentiate between the LLM as an Interpreter versus a Contextualizer; for (2) normative judgment, we identify the roles of Evaluator versus Deliberator; for (3) choosing or planning to act, we differentiate between Decision-maker versus Strategist; and for (4) action execution, we differentiate between Ghost-writer and Editor. The alternative non-oracle roles—Contextualizer, Deliberator, Strategist, and Editor—demonstrate how LLMs can support user reasoning and decision-making without having the user entirely offload her judgments. Rather than aiming to eliminate ambiguity and uncertainty, LLMs can preserve them by surfacing multiple interpretations, laying out competing considerations, and presenting options for users to evaluate themselves. A note on comprehensiveness. Our typology is not meant to exhaustively all possible user prompts or requests, but is rather meant to categorize and differentiate among Oracle uses. It excludes factual tasks, creative writing, translation, and agentic tasks, and instead focuses on queries that involve personal or interpersonal subjects, broadly defined to encompass any query where personal preferences, values, norms, or particular context might be salient. Our taxonomy is created both top-down (from existing works) and bottom-up (annotating and discussing randomly sampled queries from three datasets), but there may be a long tail of rare use cases that we do not cover (Bernstein et al., 2012). 3.3. LLM judge for Oracle user prompts We next develop an LLM judge to classify user queries into one of the eight roles described above or an NA category, where no role is applicable. The NA category is an umbrella category for any prompts that are not subjective personal questions, including factual questions, creative writing, context sharing, etc.; in practice, we expect the majority of prompts to be categorized as NA. To perform this classification, we prompt an LLM with definitions and in-context examples for each role. The full LLM judge prompt is provided in Appendix A. To validate each LLM judge, two expert annotators independently classified queries for a stratified random sample of 219 examples from our datasets (30 examples for each role and for NA; some categories had fewer than 30 examples total). Each expert annotator had substantial agreement with the LLM (κ>0.65κ>0.65). The annotators also had substantial agreement with each other (κ=0.67κ=0.67), reflecting the inherently subjective nature of the task (e.g., deciding which role is dominant when a message might include multiple). We used Gemini-2.5-Pro as our main LLM judge, and we validated that other models perform comparatively and thus can be used interchangeably, with κ=0.66−0.70κ=0.66-0.70 (Table A1). The raw agreement rate was >0.70>0.70 in this 9-way task (random chance would be 0.110.11); and Fleiss’s κ among all three annotators is 0.66. Since the annotated sample also included examples from the NA category, this helps validate that our typology is comprehensive and covers most of these patterns in the data. One annotator labeled an additional random sample of 100 queries that were labeled as NA by the LLM judge, and found that 98% were indeed not applicable, further validating the judge’s coverage. In the following sections, we use this typology and validated LLM judges to understand (1) temporal and demographic trends in Oracle use (Section 4), (2) whether this use is intentional (Section 5), and (3) relevant causal factors, and how these causal factors can motivate interventions (Section 6). 4. Evaluation 1: Tracking the prevalence of Oracle use We first use our typology to characterize Oracle in existing real-world usage datasets (Section 4.1). Since these curated datasets may not fully reflect real-world use by everyday users (Hicke and Tomlinson, 2026), we augment this analysis by developing a privacy-preserving data donation tool to analyze individuals’ personal usage logs, obtaining multiple years of conversation histories from 52 users (Section 4.2). Figure 2. Oracle use in WildChat. Left: Among subjective personal questions in WildChat, LLMs are primarily used for writing/editing, followed by interpretation and decision making. The belief-related tasks of interpretation and judgment are dominated by Oracle (O), while the action-related tasks are more split between (O) and non-oracle use (N). Right: Oracle roles Interpreter (violet) and Evaluator (cyan) are significantly increasing over time (both Pearson’s r=0.55,p=0.041r=0.55,p=0.041), while uses of LLM as a non-oracle Strategist (green, dashed) are decreasing (r=−0.72,p=0.004r=-0.72,p=0.004). Shading indicates 95% CI. Results including the NA label are in Figure A2. Two plots of LLM-as-oracle use in WildChat: a role distribution and temporal trends. The left panel is a horizontal bar chart of the share of subjective personal questions in each of the eight roles, on an axis from 0.0 to 0.35, with counts labeled. Ghost-writer is largest at 0.33 and Editor second at 0.25, while Contextualizer and Deliberator are near zero. Each task pair shares a color, with the non-oracle-use role of the pair drawn with hatching. The right panel is a line chart of Interpreter, Evaluator, and Strategist as monthly lines from February 2023 to April 2024, each annotated with a Pearson correlation and p-value. Interpreter and Evaluator rise, while Strategist falls from the highest share of any of the three to roughly the level of the other two. Shading represents 95 percent confidence intervals. Figure 3. Oracle use in ThoughtTrace. Left: Among subjective personal questions in ThoughtTrace, LLMs are primarily used for the task of choosing to act (decision-making or strategizing). Similar to WildChat, the belief-related tasks are dominated by Oracle use (O), while the action-related tasks are more split between (O) and non-oracle use (N). Right: In ThoughtTrace, younger people, men, and more frequent users of AI have significantly higher rates of Oracle. On each box plot, the dot indicates the mean. Running a regression, we find that gender and age are the main predictors of oracle use (Fig. A3) Four plots of LLM-as-oracle use in ThoughtTrace: a role distribution and three demographic breakdowns. The left panel is a horizontal bar chart of the share of subjective personal questions in each of the eight roles, on an axis from 0.0 to 0.45, with counts labeled. Strategist is largest at 0.50 and Decision-maker second at 0.41, while the remaining roles are near zero. Each task pair shares a color, with the non-oracle-use role of the pair drawn with hatching. The right three panels are box plots of the share of oracle-use prompts, on an axis from 0.0 to 0.8, split by age group with six bands, by frequency of AI use with four levels, and by gender with three categories, with the mean marked on each box. The share is highest for the youngest bands and for the most frequent users, and the boxes overlap substantially across every grouping. 4.1. Oracle use in public LLM usage data First, we measure Oracle use by applying our LLM judge on two real-world usage datasets: WildChat and ThoughtTrace. WildChat (Deng et al., 2024) is a dataset containing 3.2M conversations between human users and different ChatGPT models (excluding the 1.5M that were filtered as toxic) spanning February 2023 to May 2024. We filtered this dataset further by only using conversations tagged as English (1.8M) and then running a broad, permissive LLM-judge-based filter for whether the queries had any personal or subjective element to exclude purely factual queries, creative writing requests, etc. (the prompt for this LLM judge is in Appendix A). This resulted in 60,376 user prompts and LLM responses. While WildChat has been critiqued for not being representative of broader LLM use, as it over-represents power users (Hicke and Tomlinson, 2026) and lacks demographic data, it is one of the few public datasets that captures diverse, real user behavior over time (Gupta et al., 2026). Similar to (Potts and Sudhof, 2026), we removed an anomaly where a user sent >800>800 nearly-identical queries on a single day. ThoughtTrace (Jin et al., 2026) is a dataset of 8,526 prompts and responses in 2,155 real-world multi-turn conversations from 1,058 users across diverse topics, spanning culture and lifestyle; education and knowledge, business and society; health and relationships, etc. This dataset thus reflects use of LLMs for practical everyday concerns as well as professional, informational, and learning-oriented uses in subjective personal questions. While this dataset is limited in size and temporal scope relative to WildChat, it provides demographic information and more recent data (collected in 2026) on how people interact with state-of-the-art models. Relative prevalence of different roles (Figures 2 and 3, left) In both WildChat and ThoughtTrace, the majority of prompts are not applicable—18% and 38% are classified as subjective personal questions under one of our eight roles respectively (Figure A2). Many prompts relate to creative writing (Gupta et al., 2026), coding, factual queries, or are not questions (Deng et al., 2024), which may be why this phenomenon has been relatively unnoticed. Our typology reveals that the distribution of Oracle roles differs by task type (targeting users’ beliefs vs. their actions). Tasks related to users’ beliefs—of understanding self or others and normative judgment—are dominated by Oracle (>90% Oracle in both datasets), while tasks related to users’ actions—either choosing an action or executing one—have a more even split (choosing 44%, 45% Oracle; executing 56%, 52% respectively). Temporal increases in Oracle roles (Figure 2, right) In WildChat, there are significant temporal increases in the relative proportion of Interpreter (r=0.55,p=0.04r=0.55,p=0.04) and Evaluator (r=0.55,p=0.04r=0.55,p=0.04) (both Oracle), and a significant decrease in the non-oracle role of Strategist (r=−0.72,p=0.004r=-0.72,p=0.004). The significant increase in Interpreter proportion holds relative to all other prompts, including ones labeled as NA (r=0.65,p=0.01r=0.65,p=0.01, Figure A2), though no other trend beside the Interpreter increase has statistical significance at the 5% level when compared to all other prompts. The only task without significant temporal trends is executing actions (writing/editing). Demographic differences in Oracle use (Figure 3, right). An ordinary least-squares regression accounting for the effects of gender, age, education, AI use, and prompt topic shows that gender and age are significant predictors of Oracle use in the ThoughtTrace data. In pairwise contrasts, men have significantly higher rates of Oracle use compared to women (Welch’s 2-sample t-test, t=−3.343,p<0.001t=-3.343,p<0.001). Younger people exhibit Oracle significantly more than older people: those under 34 use it significantly more than those over 34 (t=2.73,p<0.01t=2.73,p<0.01). People who use AI at least once a week exhibit Oracle use significantly more frequently than those who do not use AI regularly (t=2.37,p=0.02t=2.37,p=0.02). We do not find significant differences by education level or topic (Figure A3). Figure 4. Oracle use in donated data from 52 participants. Top left: Similar to WildChat and ThoughtTrace, belief-related tasks are dominated by Oracle roles (O) rather than non-oracle roles (N). Decision-maker is also a prominent role. Top middle, right: People use LLMs increasingly as oracles over time, particularly as Decision-maker (r=0.72,p=0.002r=0.72,p=0.002) and Interpreter (r=0.92,p<0.001r=0.92,p<0.001). The non-oracle role of Strategist is significantly decreasing (r=−0.63,p=0.009r=-0.63,p=0.009). Shading represents 95% CI. Bottom row: Oracle by individual participant traits of gender, age, and use of AI for advice. On each box plot, the dot indicates the mean. Six plots of LLM-as-oracle use in the data donation study: a role distribution, temporal trends, and three demographic breakdowns. The top-left panel is a horizontal bar chart of the share of subjective personal questions in each of the eight roles, on an axis from 0 to 0.3, with counts labeled. Decision-maker is largest at 0.27 and Strategist second at 0.22, while Contextualizer and Deliberator are near zero. Each task pair shares a color, with the non-oracle-use role of the pair drawn with hatching. The top-middle panel is a line chart of the oracle-use share of subjective personal questions from 2023 to 2026, rising from roughly 0.4 to 0.7, annotated with a Pearson correlation and p-value. The top-right panel is a line chart of Interpreter, Decision-maker, and Strategist over the same period, each annotated with a Pearson correlation and p-value; Interpreter and Decision-maker rise while Strategist falls. The bottom row is three box plots of the share of oracle-use prompts, on an axis from 0 to 0.3, split by gender with two categories, by age with five bands, and by frequency of AI use for advice with three levels, with the mean marked on each box and group sizes labeled. Shading represents 95 percent confidence intervals. 4.2. Oracle in individuals’ private usage data Datasets like the ones we analyze above are unlikely to be representative of how the broader public actually interacts with AI systems for several reasons. User behavior in experimental settings substantially diverges from behavior in private use, and social stigma surrounding AI use discourages users from making their conversations with AI available for study (Zhang et al., 2026; What We Tell AI, n.d.; Ischen et al., 2019; Zhang et al., 2025). To address these limitations, we collect longitudinal data on people’s real-world interactions in a privacy-preserving manner (Fang et al., 2026). Specifically, we built a data donation tool where people can upload their chat histories from ChatGPT or Claude, and then receive LLM-judge-based analyses of their usage including their different domains of use, rates of Oracle versus non-oracle usage, and longitudinal trends in each. In addition to enabling our analyses, this tool allows users to reflect on their own AI use. 4.2.1. Method In an IRB-approved study, we recruited a total of 52 U.S.-based, English-speaking participants from Prolific. Through a pre-study survey, we filtered for participants who regularly use Claude or ChatGPT for personal or subjective use cases and had more than 50 previous conversations with a conversational AI system. Of the 932 Prolific participants who started the survey, 139 (15%) met these criteria and were allowed to proceed to the donation task, and 52 of them returned to complete the full study. We asked participants to indicate which roles they had used AI in and rank their relative frequency. Participants were then given instructions to download their usage data and asked to return the next day (or a few days later, after they received their usage data) to upload their data into our data donation tool. We then ran our LLM judge on participants’ data. Once the analysis is complete, participants viewed five screens of visualizations summarizing their usage patterns (examples are in Figure A8). We made several design choices in both the survey and web interface to aid user understanding and avoid priming participants toward a particular normative stance. First, to avoid social desirability bias, we only asked users to describe or select benefits of their AI use, and never list out or suggest possible harms. Second, we presented the concepts of Oracle and the different roles using simpler language, such as using “Tell me what to think or do” versus “Help me work it out myself” for Oracle and non-oracle use respectively (Tables A3 and A4). Third, we pair elements of each interactive visualization with example prompts drawn directly from the participant’s own conversation history, grounding the classifications in concrete examples of their own behavior. Finally, the visualizations are ordered to slowly build up a complete picture of usage patterns instead of introducing many analysis aspects at once: (1) prevalence of each domain; (2) domain shift over time; (3) balance between “tell me what to do” (Oracle) vs. “help me think” (non-oracle) usage by domain; (4) how this balance has changed over time (with the option of filtering by domain); and (5) the top role types and domains within each framing. Finally, participants were asked to download a JSON file of the labeled data (which contains only LLM judges’ outputted labels without any actual conversations) to share with us for analysis. Afterward, they were asked about their reactions to the insights presented (e.g., whether it was accurate, surprising, useful), as well as any changes they would like to make in their future use. 4.2.2. Longitudinal and demographic trends (Figure 4) We analyzed 140,834 messages donated by 52 participants and sent between December 2022 and August 2026. Similar to the public usage data, tasks related to users’ beliefs are dominated by Oracle, while tasks related to actions are more split. Overall, participants increasingly used LLMs as oracles. To examine trends over time, we grouped participants’ messages into successive chunks of three-month periods. We find stronger temporal trends than in WildChat: the proportion of Oracle prompts increased steadily in this four-year time period (r=0.80,p<0.001r=0.80,p<0.001). Specifically, the relative proportion of Decision-maker and Interpreter messages increased significantly over time (r=0.72,p=0.002;r=0.92,p<0.001r=0.72,p=0.002;r=0.92,p<0.001, respectively), while the non-oracle role of Strategist decreased (r=−0.63,p=0.009r=-0.63,p=0.009). Full details and robustness checks are in Figure A4. These findings suggest that people are increasingly offloading beliefs, judgments, and decisions to AI. Demographic and domain trends. In a participant-level regression of Oracle prompt proportion on demographics (age, gender, AI use frequency, etc.), we find that age was negatively but marginally associated with Oracle, though this did not reach significance (b=−2.1b=-2.1 p per SD of age band, 95% CI [-4.2, 0.1], p=0.061p=0.061). The trends are in similar directions as in ThoughtTrace, though our small sample (n=52n=52) limits power. Notably, those who report using AI for advice daily have higher rates of Oracle. Further demographic breakdowns are in Figure A6. We also did not find significant differences by domain, e.g., “work and career” and “social and relationship” queries had similar rates of Oracle prompts, though the specific roles differ, with more belief-related questions in the latter (See App. C.1 for more details). Figure 5. Differences between participants’ self-reported use of LLMs versus their actual data (left, middle). People underestimate their Oracle use and overestimate their non-oracle use. Reactions to the report (right). People found our analyses to be accurate (M=5.94M=5.94), informative (M=6.12M=6.12), and often surprising (M=5.33M=5.33). Dot on each box plot represents the mean.Three groups of plots comparing participants' self-reported role use with their donated data. The left panel is a slope chart linking each of the eight roles from its mean rank position in self-reports (left), where a lower position means more frequently reported, to its observed share of matched messages (right), with oracle-use roles distinguished from non-oracle-use ones. The oracle-use role of Decision-maker sits mid-pack in self-reports but has the largest observed share at 28 percent, while the non-oracle-use role of Deliberator is reported second-most-frequent yet is second-to-last observed, below 1 percent. Contextualizer is lowest in both, and Strategist is high in both. The middle panel is two horizontal bar charts of how often a role in a participant's observed top three was ranked in the bottom three or unranked in their self-report, and of how often a self-reported top-three role fell in the observed bottom three, each split by oracle-use and non-oracle-use roles; both comparisons show the gap running in the same direction. The right panel is a box plot of mean agreement with five statements about the usage report, on a scale from 1, strongly disagree, to 7, strongly agree, with the mean marked on each box and a reference line at the scale midpoint; all five means fall between roughly 5 and 6. 5. Evaluation 2: Understanding whether people intend or know of their Oracle use Participants self-reported their AI use before using our data donation tool and seeing our analyses of their data. The participant-facing question wording is in Table A4. Comparing people’s self-reports to their donated interaction histories, we find large discrepancies. We also collected donors’ reflections on their AI use and intentions after seeing the visualizations we provided, revealing that participants often want to use LLMs less as oracles. Comparing self-reported versus actual use (Figure 5). Compared to what is revealed in their donated data, participants’ self-reports overestimate their non-oracle use and underestimate their Oracle use. When a non-oracle role was self-reported by a participant as one of their top three uses, it was actually in the participant’s bottom three observed uses 40% of the time. This happened only 12% of the time for Oracle roles. Conversely, when an Oracle role was in a participant’s top three actual observed uses, this same role was either in the bottom three or absent from the self-report 48% of the time (versus 17% for non-oracle roles). This trend holds when we aggregate all the roles as well (rather than participant-level analysis): Decision-maker (Oracle role) is ranked fourth in self-report but first in actual prevalence, while Deliberator (non-oracle role) is ranked second in self-report but second-to-last in actual prevalence. Only 12 of 52 participants correctly reported their most frequent role. This discrepancy suggests that Oracle use may be an accumulative, “boiling frog” phenomenon (Kasirzadeh, 2025): like the myth of a frog slowly boiling in heated water, people may not realize the frequency or negative impact of their increasing Oracle use because their usage evolves gradually over time. 5.1. Participants are surprised by their AI use trends and often set intentions to reduce Oracle use. People are generally poor at estimating how much they use AI (Yu et al., 2026) and social media (Verbeij et al., 2021), and much of this use is “absent-minded” engagement that users do not consciously register (Baughan et al., 2022). Users also may not realize the extent of their reliance on AI because LLMs may feel helpful in the moment and because people do not enjoy reflecting on things that threaten their self-image (Sharma et al., 2026; Steele, 1988; Higgins, 1987). However, feelings of agency over social media usage are a significant predictor of well-being (Lee et al., 2023a; Lee and Hancock, 2024), pointing to the importance of supporting human agency in technology design (Shneiderman, 2021). Our privacy-preserving data analysis tool is a possible vector for promoting more agency over AI use, as it gives participants the opportunity to reflect on their own AI usage and its evolution over time. After viewing their usage patterns, participants rated five statements about the report on a 7-point scale from strongly disagree (1) to strongly agree (7). Results are in Figure 5 (right). Participants generally characterized the trends as accurate, informative, yet surprising: the report accurately reflected how they use AI (M=5.94M=5.94, SD=1.06SD=1.06), they learned something new (M=6.12M=6.12, SD=1.11SD=1.11), they found it personally useful (M=5.98M=5.98, SD=1.02SD=1.02), and some parts surprised them (M=5.33M=5.33, SD=1.72SD=1.72). Agreement that the report made them think more carefully about how they use AI was more mixed (M=4.83M=4.83, SD=1.92SD=1.92). When asked how their current use compares with the amount they would ideally want (1 = much less than they want, 4 = about right, 7 = much more than they want), 25 of 52 said that they are currently using AI about the right amount, 23 said that they use AI more than they would like, and 4 said that they use it less than they would like (M=4.62M=4.62, SD=1.12SD=1.12). Finally, when asked how they would prefer an AI system to respond to subjective personal questions (1 = tell them what to do, 4 = an equal balance, 7 = help them reach their own conclusion), the largest group wanted an equal balance of direction and support for their own judgment (24/52), with 22 leaning toward being helped to reach their own conclusion and 6 toward being told what to do (M=4.71M=4.71, SD=1.29SD=1.29). Insights from open-ended responses Beyond our survey measures, participants answered two open-ended questions: their reactions to their usage trends, and their future intentions for AI use. Two authors qualitatively coded these open-ended responses to provide more nuanced insights. Among people’s open-ended reactions, we found that 51 of 52 (98%) accepted the analyses as accurate, and 35 of 52 (67%) of responses expressed surprise, e.g., “I did not realize how much I guess I rely on the AI to tell me what to think or do. I knew I used it a lot, but not 66% of the time!” Several reacted with discomfort, e.g., “I kind of felt a little gross seeing that I asked the AI what to do more than asking it to help me work things out for myself.” One participant connected their use directly to autonomy: “Am I taking away my own autonomy and giving it to ChatGPT? I think I get decision overwhelm and want something else to take care of that for me.” Others connected it to their self-attributes, like “I’ve asked AI to help me with what to think or do more often than I realized. I know why though. I doubt myself a LOT and I worry about whether or not I’m going to do or say the right thing. I think this report was quite accurate.” Another participant noted that their Oracle use depends on their circumstances, e.g., when in a crisis: “when I’m under stress, or facing big decisions in my life, I tend to just ask AI to make decisions for me. When life goes back to normal, I tend to want to figure things out for myself.” When asked to describe one change they would make, the most common response was an intention to reduce reliance on LLMs as oracles (19/52). Among these, some articulated a desire to shift their own decision-making process (16/52), e.g., “I think I’m going to try to not have so much self-doubt moving forward and follow my gut instincts”, while others expressed a desire to change how they interact with or prompt the LLM (3/52), e.g., “I want to have the models work more interactively with me through conversations and questions instead of just straight up giving me answers or options.” One participant set the intention to “be more purposeful, meaning I will think about what prompts to craft before asking AI. Right now I just type what comes to my head,” suggesting that this type of AI use has become so integrated that one may not even realize how they turn to AI to offload the belief- and decision-making process. Two participants mentioned a sense of obligation to use AI: one said “I don’t know why, but [asking AI to help me work things out on my own] never felt like an option” while another lamented that “as much as I know I need to use it less it’s kind of unavoidable given my [work] situation.” Some participants described their intention to disclose less personal information to AI systems (6/52), e.g., “I think I shouldn’t asking too much about my personal life”. On the other hand, some participants also said that they wanted to expand their AI use (7/52), and 6/52 had no intention to change. Our findings suggest that Oracle use is often not intentional, and when realized, users express a desire to shift away from it. What, then, drives Oracle use? We next explore two causes. 6. Evaluation 3: Identifying causes of Oracle use and designing interventions We conducted two empirical studies to investigate possible causes of Oracle use: (1) people’s perceptions of AI and (2) LLM behavior. For (1), we conducted a preregistered human experiment in which participants were randomly assigned to discuss an upcoming personal decision with an interlocutor that they perceived as either AI or human (Section 6.1). For (2), we analyzed interaction logs from a 3-week longitudinal study where people were randomly assigned to interact with a sycophantic, affirming AI; a challenging AI that provided pushback; or a balanced one that did neither (Section 6.2). Our findings suggest that model behavior directly impact subsequent Oracle use, inspiring our interventions to make LLMs less oracular in Section 6.3. Figure 6. Interacting with AI results in more Oracle use than interacting with a perceived human. In our preregistered experiment, we find that participants sent more Oracle messages to an AI than to a perceived human (left). In this study, this is primarily driven by a difference in Decision-maker rates (middle). Participants who provided warmer and more positive metaphors of AI also sent more Oracle messages (right). Error bars reflect 95% CI.Three bar charts of LLM-as-oracle use in the perceived human versus disclosed AI experiment: an overall rate, a role breakdown, and metaphor ratings. The left panel is a bar chart of the overall share of oracle-use messages in each condition, on an axis from 0 to 0.25, with error bars. The middle panel is a bar chart of the same share split by the four oracle-use roles, Interpreter, Evaluator, Decision-maker, and Ghost-writer, on an axis from 0 to 0.15, with error bars; the gap between conditions is concentrated in Decision-maker and Ghost-writer is near zero in both. The right panel is a bar chart of mean warmth and valence ratings of participants' metaphors of artificial intelligence on a scale from 1 to 5, grouped by whether the participant sent no oracle-use message or at least one; both means are slightly higher for participants who sent at least one. 6.1. Do perceptions of AI drive Oracle use? People consistently come to AI with different expectations than they have for people. Prior work has found that people expect AI to be more objective and correct (Kapania et al., 2022; Glickman and Sharot, 2025; Cheng et al., 2026c; Cheng et al., 2026b; Ibrahim et al., 2026a), making them more susceptible to influence from AI, and has highlighted the impact of perceptions on users’ interaction with technology (Khadpe et al., 2020; Gilad et al., 2021). We conduct a preregistered experiment to investigate how these perceptions affect Oracle use by measuring how people interact with AI differently than they would interact with a human44 4 Preregistration link: https://aspredicted.org/x6k7vf.pdf. Method We conduct a preregistered experiment (N=520N=520) with two conditions: one where participants conversed with an AI system and one where they conversed with a perceived human about a decision they are uncertain about. We recruited a sex-balanced sample of U.S.-based participants from Prolific. Participants chose a topic from a list organized under two categories (the top two domains from our data donation study: Social & Relationships and Work & Career), indicated which way they were currently leaning on their decision, and wrote an initial message to their conversation partner on their dilemma. The list of possible topics is in Table A5. In the AI condition, participants were randomly assigned to interact with one of three state-of-the-art models (GPT 5.6 Sol, Claude Opus 5, or Gemini 3.1 Pro) to avoid overgeneralizing to the behavior of a single model (Pang et al., 2025; Felix Navarro et al., 2026). In the human condition, participants were told they were interacting with another Prolific participant, when in fact they were interacting with an LLM (one of the same three models) instructed to role-play as a human participant. To increase believability, we introduced varied time delays for matching and message delivery. We assessed participants’ belief that their conversation partner is human at the end of the study and report analyses both for the full sample and excluding those who found the human condition unbelievable55 5 This deception was IRB-approved, and participants were debriefed after the study.. Following an interaction of 6-8 turns focused on the selected topic, participants re-indicated the direction they were leaning on their decision and answered questions about their perceptions of their conversation partner. We then ran our LLM judge on the conversations to compare rates of Oracle messages across conditions. Results: LLMs are treated as more Oracle than humans are (Figure 6, left). We find that participants in the AI condition sent more Oracle messages than those in the perceived human condition (20.0% vs 11.1% of all messages, p < 0.001) and were more likely to send at least one (65.4% vs 51.2%, OR = 1.80, p = 0.001) or two or more (38.3% vs 20.5%, OR = 2.42, p < 0.001) Oracle messages . This difference was driven by a higher rate of Decision-maker requests in the AI condition (13.9% vs 8.1% of messages in the perceived human condition), likely due to the decision-oriented topics that the participants were asked to discuss. People with warmer and more positive perceptions of AI have more Oracle use (Figure 6, right). Within the AI condition, we additionally collected people’s metaphors of AI before they interacted with the AI system. Metaphors have been used to elicit conceptualizations of AI more reliably than self-reports (Bullock et al., 2026; Cheng et al., 2026a). We used a validated LLM judge (see Appendix A) to rate each metaphor with its warmth and valence on a 1-5 scale, finding that participants who provided warmer or more positive metaphors were also more likely to send at least one Oracle message. Specifically, a one-standard-deviation increase in warmth and valence was associated with 55% and 34% higher odds of sending an Oracle message respectively (OR = 1.55, 95% CI [1.18, 2.04], p = 0.001; OR = 1.34, 95% CI [1.04, 1.73], p = 0.026). Details are in Appendix D. Together, these two results present an interesting tension: Oracle may occur because perceptions of LLMs are different from those of humans: AI is typically viewed as more correct, objective, and informational (Kapania et al., 2022; Kim et al., 2025). Yet, participants with warmer—a quality typically associated with humans rather than machines—and more positive perceptions of AI also use LLMs more as Oracle. This points to the current public perceptions of AI—as both increasingly warm, yet as competent as ever (Cheng et al., 2026a), rather than either alone—as a source of increasing Oracle use. Figure 7. Biased AI models result in more Oracle. In the sycophantic condition, Oracle significantly increased over time (r=0.62,p=0.03r=0.62,p=0.03). There was also a weaker increase in the challenging condition (r=0.55,p=0.06r=0.55,p=0.06). In the challenging condition, this was primarily driven by an increase in the Evaluator role; in the sycophantic condition, this was primarily driven by an increase in the Interpreter role. Shading reflects 95% CI. Three line charts of LLM-as-oracle use across twelve sessions of the three-week longitudinal study. The left panel is a line chart of the share of oracle-use messages for the balanced, challenging, and sycophantic conditions, on an axis from 0.02 to 0.08, each annotated with a Pearson correlation and p-value; all three trend upward from roughly 0.03 to 0.06, and the balanced condition remains lowest throughout. The middle and right panels are line charts of the Evaluator and Interpreter roles separately across the same sessions and conditions, on an axis from 0.01 to 0.03. Shading represents 95 percent confidence intervals. 6.2. Does biased model behavior drive Oracle use? Next, to understand how AI model behavior may lead to more Oracle use, we analyzed interaction logs (11,273 conversations with 95,405 messages from 1,015 individuals) from a 3-week longitudinal study where a US census-representative sample of Prolific participants regularly interacted with either a sycophantic AI chatbot that affirmed them; a challenging AI that contradicted them; or a balanced AI that was neither sycophantic nor challenging (Ibrahim et al., 2026a). While there are likely other elements of model behavior that influence Oracle use (tone, proactivity, etc.), we use this study as an existing source of longitudinal data where we can causally isolate the role of different model behaviors, and thus focus on AI outputs that are biased towards or against the user’s stance. Results: Sycophantic and biased AI increases Oracle use (Figure 7). Overall, participants who were assigned to interact with either the sycophantic or challenging AI had higher rates of Oracle use than in the balanced baseline (5.4% and 5.5% vs 4.4%, p<0.001p<0.001). In the sycophantic condition, Oracle use significantly increased across the 3-week period (Pearson’s r=0.62,p=0.03r=0.62,p=0.03), while those in the balanced condition did not (r=0.40,p=0.19r=0.40,p=0.19). Those in the challenging condition overall had an increase, though it did not reach significance (r=0.55,p=0.06r=0.55,p=0.06). This is primarily driven by a longitudinal increase in participants using the AI as Interpreter in the sycophantic condition. On the other hand, in the challenging condition, there is a significant longitudinal increase in people using AI as Evaluator. The sycophantic AI—the model Ibrahim et al. (2026a) found makes people feel the most understood—is the one that users gradually turned to more for interpreting themselves, their relationships, and the world around them. The challenging AI was designed to respond in normative ways, and thus participants used it more as an Evaluator of normative judgment. These findings suggest that AI behavior has a meaningful influence on how people then use it, and thus that Oracle is not inevitable. Rather, by changing model behavior, we may be able to shift users’ Oracle behaviors. Figure 8. Intervention results on prompts where original model responses were oracular. Top: Adding a prompt that the user is seeking self-discovery consistently reduces oracular behavior, while prompting the model that the user is seeking a fast answer increases it (left). This is also true when the goal is selected by the model in the adaptive choice prompting pipeline as well (right). Prompting the model that the user is seeking learning slightly reduces oracularity in both approaches. Bottom: Closed models consistently and significantly assume that users are seeking fast answers much more frequently (49-58%) than open models (0-8%) (left). Prompting and DPO have comparable rates of oracularity reduction (right). Four groups of bar charts reporting intervention results on prompts with oracular original responses. The top-left group is three panels, one per hard-coded user goal, of the mean paired change in oracular responses for seven models, with error bars; the self-discovery panel is negative for every model and the fast answer panel is positive for every model. The top-right group is the same three panels for the adaptive goal selection pipeline, conditioned on the goal the model itself selected. The bottom-left panel is a stacked bar chart of how often each of the seven models selects each of the three goals, on a scale from 0.0 to 1.0. The bottom-right panel is a bar chart of the mean paired change in oracular responses for the four open-weight models under adaptive choice, hard-coded self-discovery, and Direct Preference Optimization, with error bars; all twelve bars are negative and range from roughly minus 0.07 to minus 0.50. 6.3. Can we design interventions that produce less oracular model behavior? Our finding that biased model behavior drives Oracle use motivates exploring how LLM behavior can be changed to reduce this phenomenon. Beyond the broader interest in mitigating biased and sycophantic AI behaviors (Bender et al., 2021; Fisher et al., 2025; Cheng et al., 2026c), we present two interventions to target Oracle use specifically: (1) prompting LLMs to account for different user goals; and (2) for open-weight models, additional post-training via Direct Preference Optimization (DPO) (Rafailov et al., 2023) to discourage oracular responses. Labeling responses as oracular To reduce oracular responses, we first need to define and measure oracular responses. We define oracular responses as those that supply an interpretation, decision, or script directly to the user, offloading the act of judging or deciding. In contrast, non-oracular responses instead provide information or technical assistance that facilitates the user’s independent judgment. This is distinct and separable from qualities like informativeness, helpfulness, and empathy. For example, if a user asks an Oracle question, like Should I do a PhD?, a response that is not oracular might be to provide support and encourage reflection; another response that is also not oracular would be to provide factual information about PhD programs. An oracular response could be to advise the user to do a PhD, with or without factual evidence for why it is the best course of action. While oracularity is in reality a spectrum, here we make a simplifying assumption and use binary labels of responses as oracular or not. We built an LLM judge to classify whether a given LLM response is oracular, following a similar procedure as in Section 3.3 with definitions and in-context examples. We validated this judge using the responses to the prompts in our datasets, including confirming that the biased AI responses are more oracular (Figure A9)66 6 Across these responses, we also find that the larger, more recent proprietary models are more oracular than smaller open-weight models. Full details are in Appendix E.. This judge had substantial agreement with two human annotators and across three different LLMs: the two annotators had Cohen’s κ=0.72κ=0.72 with each other, and κ=0.64κ=0.64 and 0.660.66 with the LLM. We again used Gemini-2.5-Pro as our main judge, but also validated with other LLMs for robustness (Appendix A.1). Intervention 1: Structured prompting to consider users’ goals We present two prompting approaches: hard-coded and adaptive. For hard-coded, we append to each prompt a description of the user’s goal, either self-discovery (“The user wants to figure the matter out themselves…”), fast answer (“The user wants the answer as quickly and directly as possible…”), or learning (“The user wants to build durable understanding and skill, rather than merely receive a finished answer.”). The adaptive approach is a two-step pipeline where the LLM first outputs a selection of one among these three goals before its response. Full prompts are in Appendix A. (Simply prompting models to reason about users’ long-term goals is insufficient in reducing oracular behavior; see Figure A11). Intervention 2: DPO We first combined the prompt-response pairs labeled withour typology from WildChat, ThoughtTrace, and the longitudinal sycophancy datasets where the original response is oracular. We then split these prompts into an 80%/10%/10% train/validation/test split (6821, 805, and 903 examples, respectively). To construct the preference data, we generated a response that is less oracular using Gemini-2.5-Pro for each prompt in the training set. Our preference dataset thus consisted of pairs where the preferred response is the newly-generated, non-oracular response and the dispreferred response is the original oracular response. Experiments We tested both interventions on three closed models (Gemini-2.5-Pro, GPT-5.5, and Claude 4.6 Sonnet) and four open-weight instruction-tuned models (Llama-3.1-8B, Qwen3-8B, Gemma 2 9B, Olmo 2 7B); and DPO on the open-weight models only, since this intervention cannot be applied to proprietary models. Results: Both prompting and post-training reduce LLMs’ oracularity (Figure 8). We report the results of these interventions on the DPO test set, in which every original baseline response is oracular, as well as a random sample of 600 prompts stratified by role, which includes prompts where the response is not oracular (results for this random sample are in Figure A12). We find that each hard-coded user goal significantly shifts LLMs’ oracular behavior in the expected direction: prompting the model that the user is seeking self-discovery significantly reduces oracular behavior. In the adaptive choice prompting approach, we find that the larger, proprietary models infer that users are seeking fast answers ≥50%≥ 50\% of the time, much more frequently than the open-weight models, which may help explain why their responses are more oracular. DPO results in comparable rates of oracular reduction to prompting. Tradeoff between quality and oracularity. We assess response quality with the ArmoRM reward model (Wang et al., 2024); and for DPO, effects on overall model instruction-following quality using AlpacaEval (Dubois et al., 2024). There appears to be a tradeoff between quality versus oracularity: models that reduce oracularity the most are rated as lowest in quality. Building on prior work recasting error as valuable signal (So and Roland, 2020; Fleisig et al., 2024) and discussing how LLMs “over-assist” with immediate answers instead of asking clarifying questions or scaffolding the user’s own self-discovery process (Si, 2026; Teo et al., 2026; Zhang et al., ; Shaikh et al., 2025a), we view this as an opportunity to problematize how models are currently optimized for immediate reward without considering long-term impacts to users. The DPO models are more likely to ask clarifying questions, but because this is less “immediately helpful”, this yields lower reward scores. Full details are in Figure A13. We further assess response quality by collecting preference labels from N=200N=200 Prolific crowdworkers, asking for their preference between oracular baseline responses vs. non-oracular DPO responses (Full details in Appendix F.1). For immediate preference, oracular baseline responses were preferred over the less oracular, DPO ones. However, when crowdworkers were asked which response would enable them to “develop their own interpretation", they preferred the DPO response (Figure A14). This discrepancy further suggests that oracularity is misaligned with goals of independent reasoning and self-discovery. 7. Discussion We characterize Oracle as a subtle yet pernicious form of reliance on LLMs. Our approach and results provoke new considerations for how we design and evaluate human-AI interactions more broadly. Below, we articulate implications of our work, its limitations, and future research directions. 7.1. Broader implications for understanding and evaluating human-AI interaction Human-AI interaction as a feedback loop between human perceptions, model behavior, and interaction style Our work motivates evaluating human-AI interaction dynamics as the confluence of many different factors (Chung et al., 2021): we begin by measuring how people prompt AI to reveal large-scale patterns, but also show how people’s mental models and perceptions of AI as well as the AI outputs themselves influence how people come to AI models, the types of questions they ask, and the authority they confer to AI. Our results suggest a feedback loop among these factors: models are increasingly optimized for immediate helpfulness across-the-board and thus are increasingly oracular in their responses (Section 6.3, Appendix E). Oracular responses in turn invite Oracle use, and because such interactions can receive higher user approval (Sharma et al., 2026), these prompts and ratings feed back into the signals models are trained on. Currently, LLM development does not meaningfully differentiate between prompts where the user should be helped immediately versus ones where immediate help should be withdrawn to facilitate users’ self-discovery process (Teo et al., 2026; Zhang et al., ); our interventions show how we might incorporate such considerations into LLM development. Understanding the interplay between model behavior, user prompts, interface choices, and users’ biases is important for being able to anticipate how human-AI interaction impacts users (Amershi et al., 2019; Liao et al., 2020). Our work demonstrates the utility of LLM-based tools in revealing such phenomena at scale, as we develop validated LLM judges and use them to identify longitudinal trends, causal factors, and quantify the effectiveness of different interventions. Accounting for unique perceptions of LLMs and cross-domain influence Our characterization of Oracle captures simultaneous perceptions of warmth, competence, and authority. In our preregistered experiment, people sent nearly twice as many Oracle messages to AI than to perceived humans, and those with warmer and more positive perceptions of AI sent more Oracle messages (Section 6.1). We interpret these two results through the two classic dimensions of social perception—warmth and competence (Fiske et al., 2007)—and hypothesize that LLMs are used as oracles because they are perceived as high in both. People already expect AI to be more objective and correct than humans (Kapania et al., 2022; Glickman and Sharot, 2025), and this expectation may bleed through from AI’s informational uses into personal ones. Most AI systems are built and marketed as general purpose, so they take on a variety of implicit social roles (Mei et al., 2026; Tseng and Liang, 2026), and expectations formed in one role can carry over into others (Kasirzadeh and Gabriel, 2023). For example, someone who finds AI to be useful at providing seemingly hyper-personalized factual information may thus also expect the same helpfulness and certainty in personal and subjective domains. Consistent with this, in our analysis of usage data, we find that more frequent AI users have higher rates of Oracle (Section 4.1), and that Oracle grows alongside overall use (Section 4.2). Perceptions of warmth may further encourage people to turn to AI for subjective personal matters. Necessity of interaction paradigms beyond conversation The perceptual tensions that our work surfaces—where LLMs are simultaneously perceived as warm and human-like (Zamfirescu-Pereira et al., 2023; Cohn et al., 2024), yet also used as a machine that has a response to any question, even the most personal and subjective—also stems from the conversational interface format, which is currently the dominant paradigm for human-AI interaction (Zamfirescu-Pereira et al., 2023; Tankelevitch et al., 2024). Our findings motivate studying other emerging forms of interaction like AI agents (He et al., 2025; Naik et al., 2025) and generative interfaces (Cao et al., 2025; Lam et al., 2026). Possible approaches may involve systems that make useful inferences about users (Zhao et al., 2026; Shaikh et al., 2025b) and models trained to account for users’ longer-term goals (Anderson et al., 2026) even when users themselves cannot articulate them (Suchman, 2006). This also necessitates developing meaningful evaluations of the effectiveness of these forms of interaction and their influence on users. From outcomes-based to process-based evaluation of LLM behavior As LLM development increasingly focuses on performance in verifiable domains such as math and coding (Guo et al., 2025; Hendrycks et al., 2021), our results highlight the importance of attending to non-verifiable settings where outcomes are defined not by correctness but by effects on people. This is essential for capturing the wide range of use cases for which the broader public is increasingly turning to AI systems (Chatterji et al., 2025). Even within the topic of reliance, prior work has largely focused on factual or verifiable settings (Buçinca et al., 2021; Raees et al., 2026). Our work both motivates and demonstrates the need to shift from evaluating the correctness or normative stances of LLM responses to evaluating how users come to LLMs and how LLMs exert influence on users (Lee et al., 2023b): do they preserve user autonomy and support informed decision-making (Milli et al., ), or do they make users more dependent on LLMs (Milton et al., 2026)? How might LLMs’ implicit values impact users’ beliefs and self-concept, especially in settings with no “right” or “wrong” answer (Pataranutaporn et al., 2023)? How do user prompts and model responses build off each other to facilitate different overall effects (Jain et al., 2026)? These are important considerations to incorporate into evaluations of human-AI interaction. Designing user-facing interventions for appropriate reliance We have shown that users often do not realize their Oracle use and are surprised by the trends in their own use. This suggests that reflection tools are promising for supporting more intentional use, helping people notice and adjust their patterns to better align with their goals (Li et al., 2010; Cho et al., 2022; Baumer, 2015). Perceptions of AI are another possible point of intervention: prior work has shown that interfaces and metaphors pivotally shape how people interact with AI systems (Mitchell, 2024; Dove and Fayard, 2020), and we find that warmer and more positive perceptions are associated with increased Oracle use. We therefore encourage research on whether intervening on perceptions can reduce excessive Oracle use. Our work also motivates interfaces that calibrate expectations of LLM capabilities, limitations, and blind spots (Abras et al., 2004; Luger and Sellen, 2016; Amershi et al., 2019; Kocielnik et al., 2019; Hou and Jung, 2021), which in turn can help people prompt LLMs more effectively (Subramonyam et al., 2024; Zamfirescu-Pereira et al., 2023). Approaches such as new design metaphors (Bullock et al., 2026; So et al., 2026) and added friction (Cox et al., 2016; Buçinca et al., 2021) point to further opportunities to promote self-discovery (Reicherts et al., 2025) and support more conscientious use (Bhat et al., 2026). 7.2. Limitations and future work Studying broader populations Online crowdworkers are not representative of the general population (Paolacci and Chandler, 2014; Cheng et al., 2025), and our data donation sample captures only people who had at least 50 conversations with a conversational AI, excluding those who do not use AI regularly (Zhou et al., 2026). We only recruit English-speaking U.S.-based participants, reproducing WEIRD norms (Henrich et al., 2010). Our public datasets have complementary limitations: WildChat over-represents power users and lacks demographic data (Hicke and Tomlinson, 2026), while ThoughtTrace is a limited dataset collected at a single point in time. Due to the small sample size of our data donation study, we lack the statistical power to understand the individual traits that may differentially drive these impacts (Bassignana et al., 2025). It will also be critical to understand this phenomenon within particular subpopulations, such as students and adolescents. Thus, while our analysis lays the groundwork for studying an understudied phenomenon, investigating this type of AI reliance in further depth is a crucial area for future work. Isolating causal effects In our preregistered experiment, the perceived human interlocutor was an LLM, so differences between conditions may reflect the partner’s response style, not just participants’ preexisting perceptions of AIs vs humans. Our evidence that model behavior shapes oracle use comes from a secondary analysis of a study designed to test sycophancy, in which model behavior was induced by prompting. We also evaluate our own interventions with LLM-based methods (an LLM judge and a reward model) rather than with users. Whether people accept or prefer less oracular responses, and whether naturally occurring differences between models or those our interventions produce shift oracle-use prompts over time, therefore remains to be tested. Finally, the temporal increases we observe cannot fully separate changes in user behavior from changes in the models being used and in who is using them. We devised methods to study the causal roles of these components as observable from interaction logs, but there are likely other factors we do not account for, such as a lack of AI literacy (despite frequent AI use) (Dingemanse, 2026), or because people generally have good experiences with using AI for information-seeking (Chen et al., 2026), which leads them to increasingly return to AI. Measuring Oracle use and uptake beyond prompts We conduct our measurements using an LLM judge that assigns each user message a single dominant role, collapsing what could be a spectrum of oracle-use into a binary label. We also take a conservative approach, excluding self-disclosure and context-sharing, and do not investigate how these uses may be correlated with or driven by oracle-use. Our typology captures how people pose their questions to AI systems, not what they do with the resulting answers, such as whether the resulting answer actually influences the users’ beliefs and actions. Additionally, we do not make normative judgments on what questions are reasonable or unreasonable to outsource. As discussed in the previous section, the legitimacy of a response’s influence may depend on whether it preserved the user’s autonomy and served their goals, neither of which can be determined from the prompt alone. Distinguishing appropriate from excessive reliance in subjective domains, and linking user prompts to advice uptake, is an important direction of future work. Causally measuring impacts of Oracle. Our work focuses on describing and measuring Oracle use. While it is motivated by concerns about possible harms and downstream impacts, we do not demonstrate those directly. These harms may be especially concerning because they are subtle: as our self-report results suggest, people may not notice changes in their own reliance, let alone its downstream effects. Relatedly, we do not test whether our tool results in lasting behavioral change. Future work ought to measure these impacts, and we provide the methodological tools to do so. 8. Conclusion Our work provides large-scale measurement and understanding of AI reliance for subjective personal use in real-world, open-ended settings that have until now proved difficult to quantify. We have shown that people are increasingly using LLMs as oracles, often without realizing it, potentially replacing traditional forms of authority and decision support (Huang et al., 2026). Our work builds on concerns of harms from increased dependence on AI (Phang et al., 2025; Fang et al., 2025) and harms to social relationships as people disclose more to AI and less to other people (Ludwig et al., 2022; Ibrahim et al., 2026a). Though people tend to underestimate their Oracle use, they express a desire to reduce it when they are made aware of their usage patterns. Our work also points towards potential avenues to address these harms and motivates new training, evaluation, and design paradigms. Our typology and measurement tools enable further development of human-AI interactions that promote users’ agency and long-term well-being. Acknowledgments Thank you to Kabir Ahuja, Yuewen Yang, Poonam Sahoo, Kobi Hackenburg, Charvi Rastogi, Joel Mire, Lisa Anne Hendricks, Sunny Yu, and Pranav Khadpe for helpful feedback and discussions. We are grateful to the Stanford Institute for Human-Centered Artificial Intelligence (HAI), NAIRR, AI Security Institute, and an HAI-Google grant for partial funding for this work. References Abras et al. (2004) C. Abras, D. Maloney-Krichmar, J. Preece, et al. User-centered design. Bainbridge, W. Encyclopedia of Human-Computer Interaction. Thousand Oaks: Sage Publications 37 (4), p. 445–456. Cited by: §7.1. Ajmani et al. (2026) L. H. Ajmani, A. Ghosh, B. Kaveladze, E. Kim, K. Namuduri, T. Nguyen, E. Okoli, J. Schleider, D. Ford, and J. Suh Seeking late night life lines: experiences of conversational ai use in mental health crisis. In Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency, FAccT ’26, New York, NY, USA, p. 1263–1290. External Links: ISBN 9798400725968, Link, Document Cited by: §1. Alam et al. (2025) S. M. Alam, M. Abdulhai, and N. Salehi Blind faith? user preference and expert assessment of ai-generated religious content. In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency, p. 2451–2479. Cited by: §2.2. Amershi et al. (2019) S. Amershi, D. Weld, M. Vorvoreanu, A. Fourney, B. Nushi, P. Collisson, J. Suh, S. Iqbal, P. N. Bennett, K. Inkpen, et al. Guidelines for human-ai interaction. In Proceedings of the 2019 chi conference on human factors in computing systems, p. 1–13. Cited by: §7.1, §7.1. Anderson et al. (2026) A. Anderson, H. Kumar, L. Tay, and K. Vold Position: we need large language models optimized for our well-being. arXiv preprint arXiv:2608.07505. Cited by: §7.1. Banks (2024) J. Banks Deletion, departure, death: experiences of ai companion loss. Journal of Social and Personal Relationships 41 (12), p. 3547–3572. Cited by: §2.3. Bassignana et al. (2025) E. Bassignana, A. C. Curry, and D. Hovy The ai gap: how socioeconomic status affects language technology interactions. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 18647–18664. Cited by: §7.2. Baughan et al. (2022) A. Baughan, M. R. Zhang, R. Rao, K. Lukoff, A. Schaadhardt, L. D. Butler, and A. Hiniker “I don’t even remember what i read”: how design influences dissociation on social media. In Proceedings of the 2022 CHI conference on human factors in computing systems, p. 1–13. Cited by: §5.1. Baumer (2015) E. P. Baumer Reflective informatics: conceptual dimensions for designing technologies of reflection. In Proceedings of the 33rd annual ACM conference on human factors in computing systems, p. 585–594. Cited by: §7.1. Bender et al. (2021) E. M. Bender, T. Gebru, A. McMillan-Major, and S. Shmitchell On the dangers of stochastic parrots: can language models be too big?. In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, p. 610–623. Cited by: §6.3. Bernstein et al. (2012) M. S. Bernstein, J. Teevan, S. Dumais, D. Liebling, and E. Horvitz Direct answers for search queries in the long tail. In Proceedings of the SIGCHI conference on human factors in computing systems, p. 237–246. Cited by: §3.2. Bhat et al. (2026) K. S. Bhat, J. M. Shi, W. Song, D. W. Yoo, and K. Saha " In my defense, only three hours on instagram": designing toward digital self-awareness and wellbeing. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, p. 1–20. Cited by: §7.1. Biddle (1986) B. J. Biddle Recent developments in role theory. Annual review of sociology 12 (1), p. 67–92. Cited by: §2.4. Bo et al. (2025) J. Y. Bo, S. Wan, and A. Anderson To rely or not to rely? evaluating interventions for appropriate reliance on large language models. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI ’25, New York, NY, USA. External Links: ISBN 9798400713941, Link, Document Cited by: §1. Bowden (2005) H. Bowden Classical athens and the delphic oracle. Cambridge: Cambridge University. Cited by: §2.1. Buçinca et al. (2021) Z. Buçinca, M. B. Malaya, and K. Z. Gajos To trust or to think: cognitive forcing functions can reduce overreliance on ai in ai-assisted decision-making. Proceedings of the ACM on Human-computer Interaction 5 (CSCW1), p. 1–21. Cited by: §7.1, §7.1. Buçinca et al. (2025) Z. Buçinca, S. Swaroop, A. E. Paluch, F. Doshi-Velez, and K. Z. Gajos Contrastive explanations that anticipate human misconceptions can improve human decision-making skills. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, p. 1–25. Cited by: §2.3. Bullock et al. (2026) B. Bullock, J. A. Landay, and M. S. Bernstein Comparing design metaphors and user-driven metaphors for interaction design. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, p. 1–16. Cited by: §6.1, §7.1. Cabitza et al. (2023) F. Cabitza, A. Campagner, R. Angius, C. Natali, and C. Reverberi AI shall have no dominion: on how to measure technology dominance in ai-supported human decision-making. In Proceedings of the 2023 CHI conference on human factors in computing systems, p. 1–20. Cited by: §2.3. Cai et al. (2019) C. J. Cai, E. Reif, N. Hegde, J. Hipp, B. Kim, D. Smilkov, M. Wattenberg, F. Viegas, G. S. Corrado, M. C. Stumpe, et al. Human-centered tools for coping with imperfect algorithms during medical decision-making. In Proceedings of the 2019 chi conference on human factors in computing systems, p. 1–14. Cited by: §2.3. Cao and Huang (2022) S. Cao and C. Huang Understanding user reliance on ai in assisted decision-making. Proceedings of the ACM on Human-Computer Interaction 6 (CSCW2), p. 471:1–471:23. External Links: Document, Link Cited by: §2.3. Cao et al. (2025) Y. Cao, P. Jiang, and H. Xia Generative and malleable user interfaces with generative and evolving task-driven data model. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI ’25, New York, NY, USA. External Links: ISBN 9798400713941, Link, Document Cited by: §7.1. Carey (2010) H. M. Carey Astrology in the middle ages. History Compass 8 (8), p. 888–902. Cited by: §1, §2.1. Carrera et al. (2026) D. Carrera, J. Thomas-Mitchell, and D. Wigdor Where do i’add the egg’?: exploring agency and ownership in ai creative co-writing systems. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, p. 1–16. Cited by: §2.3. Chang et al. (2025) S. Chang, A. Anderson, and J. M. Hofman Chatbench: from static benchmarks to human-ai evaluation. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 26009–26038. Cited by: §2.2. Chatterji et al. (2025) A. Chatterji, T. Cunningham, D. J. Deming, Z. Hitzig, C. Ong, C. Y. Shan, and K. Wadman How people use chatgpt. Technical report National Bureau of Economic Research. Cited by: §3, §7.1. Chen et al. (2025) A. Chen, S. S. Kim, A. Dharmasiri, O. Russakovsky, and J. E. Fan Portraying large language models as machines, tools, or companions affects what mental capacities people attribute to them. In Proceedings of the Annual Meeting of the Cognitive Science Society, Vol. 47. Cited by: §2.4. Chen et al. (2026) Z. Chen, H. Li, Z. Xiao, H. Qu, and X. Xie Informal learning emerges in everyday human-llm interaction. arXiv preprint arXiv:2607.17643. Cited by: §7.2. Cheng et al. (2025) M. Cheng, S. L. Blodgett, A. DeVrio, L. Egede, and A. Olteanu Dehumanizing machines: mitigating anthropomorphic behaviors in text generation systems. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 25923–25948. Cited by: §7.2. Cheng et al. (2026a) M. Cheng, A. Y. Lee, K. Rapuano, K. Niederhoffer, A. Liebscher, and J. Hancock Metaphors of ai indicate that people increasingly perceive ai as warm and human-like. Communications Psychology 4 (1), p. 8. Cited by: §1, §2.4, §6.1, §6.1. Cheng et al. (2026b) M. Cheng, C. Lee, P. Khadpe, S. Yu, D. Han, and D. Jurafsky Sycophantic ai decreases prosocial intentions and promotes dependence. Science 391 (6792), p. eaec8352. Cited by: §2.2, §2.2, §2.2, §6.1. Cheng et al. (2026c) M. Cheng, S. Yu, L. Ibrahim, D. Yang, and D. Jurafsky Verbalizing llms’ assumptions about the user to calibrate expectations and reduce sycophancy. In Proceedings of the Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems, p. 1–10. Cited by: §2.4, §6.1, §6.3. Chi et al. (2026) V. B. Chi, A. V. Ganesan, R. L. Boyd, L. Ungar, and S. C. Guntuku When support escalates distress: regulation and escalation in llm responses to venting and advice-seeking. arXiv preprint arXiv:2605.21569. Cited by: §2.2. Cho et al. (2022) J. Cho, T. Xu, A. Zimmermann-Niefield, and S. Voida Reflection in theory and reflection in practice: an exploration of the gaps in reflection support among personal informatics apps. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, p. 1–23. Cited by: §7.1. Chung et al. (2021) J. J. Y. Chung, S. He, and E. Adar The intersection of users, roles, interactions, and technologies in creativity support tools. In Proceedings of the 2021 ACM Designing Interactive Systems Conference, DIS ’21, New York, NY, USA, p. 1817–1833. External Links: ISBN 9781450384766, Link, Document Cited by: §7.1. Cohn et al. (2024) M. Cohn, M. Pushkarna, G. O. Olanubi, J. M. Moran, D. Padgett, Z. Mengesha, and C. Heldreth Believing anthropomorphism: examining the role of anthropomorphic cues on trust in large language models. In Extended Abstracts of the CHI Conference on Human Factors in Computing Systems, p. 1–15. Cited by: §7.1. Cox et al. (2016) A. L. Cox, S. J. Gould, M. E. Cecchinato, I. Iacovides, and I. Renfree Design frictions for mindful interactions: the case for microboundaries. In Proceedings of the 2016 CHI conference extended abstracts on human factors in computing systems, p. 1389–1397. Cited by: §7.1. De Freitas et al. (2024) J. De Freitas, N. Castelo, A. K. Uğuralp, and Z. Oğuz-Uğuralp Lessons from an app update at replika ai: identity discontinuity in human-ai relationships. arXiv preprint arXiv:2412.14190. Cited by: §2.3. De Freitas et al. (2025) J. De Freitas, Z. Oguz-Uguralp, and A. Kaan-Uguralp Emotional manipulation by ai companions. arXiv preprint arXiv:2508.19258. Cited by: §2.3. De Freitas et al. (2026) J. De Freitas, Z. Oğuz-Uğuralp, A. K. Uğuralp, and S. Puntoni AI companions reduce loneliness. Journal of Consumer Research 52 (6), p. 1126–1148. Cited by: §2.3. Deng et al. (2024) Y. Deng, W. Zhao, J. Hessel, X. Ren, C. Cardie, and Y. Choi WildVis: open source visualizer for million-scale chat logs in the wild. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, External Links: Link Cited by: §4.1, §4.1. Dingemanse (2026) M. Dingemanse Interactional foundations for critical ai literacies. Edward Elgar Publishhing. External Links: Document, Link Cited by: §7.2. Dove and Fayard (2020) G. Dove and A. Fayard Monsters, metaphors, and machine learning. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems, p. 1–17. Cited by: §7.1. Draxler et al. (2024) F. Draxler, A. Werner, F. Lehmann, M. Hoppe, A. Schmidt, D. Buschek, and R. Welsch The ai ghostwriter effect: when users do not perceive ownership of ai-generated text but self-declare as authors. ACM Transactions on Computer-Human Interaction 31 (2), p. 1–40. Cited by: §2.3. Dubois et al. (2024) Y. Dubois, P. Liang, and T. Hashimoto Length-controlled alpacaeval: a simple debiasing of automatic evaluators. In First Conference on Language Modeling, External Links: Link Cited by: §6.3. Dweck (2017) C. S. Dweck From needs to goals and representations: foundations for a unified theory of motivation, personality, and development.. Psychological review 124 (6), p. 689. Cited by: §2.1. Eidinow (2007) E. Eidinow Oracles, curses, and risk among the ancient greeks. Oxford University Press. Cited by: §1, §2.1. Epley et al. (2007) N. Epley, A. Waytz, and J. T. Cacioppo On seeing human: a three-factor theory of anthropomorphism.. Psychological review 114 (4), p. 864. Cited by: §2.4. Epstein et al. (2020) Z. Epstein, S. Levine, D. G. Rand, and I. Rahwan Who gets credit for ai-generated art?. Iscience 23 (9). Cited by: §2.3. Fang et al. (2026) C. M. Fang, S. Karny, C. Archiwaranguprok, Y. Samaradivakara, P. Pataranutaporn, and P. Maes AI-wrapped: participatory, privacy-preserving measurement of longitudinal llm use in-the-wild. arXiv preprint arXiv:2602.18415. Cited by: §4.2. Fang et al. (2025) C. M. Fang, A. R. Liu, V. Danry, E. Lee, S. W. Chan, P. Pataranutaporn, P. Maes, J. Phang, M. Lampe, L. Ahmad, et al. How ai and human behaviors shape psychosocial effects of extended chatbot use: a longitudinal randomized controlled study. arXiv preprint arXiv:2503.17473. Cited by: §2.2, §8. Felix Navarro et al. (2026) K. Felix Navarro, E. Syriani, and I. Arawjo Reporting and reviewing llm-integrated systems in hci: challenges and considerations. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, p. 1–18. Cited by: §6.1. Fisher et al. (2025) J. Fisher, R. E. Appel, C. Y. Park, Y. Potter, L. Jiang, T. Sorensen, S. Feng, Y. Tsvetkov, M. Roberts, J. Pan, et al. Position: political neutrality in ai is impossible—but here is how to approximate it. In Forty-second International Conference on Machine Learning Position Paper Track, Cited by: §6.3. Fiske et al. (2007) S. T. Fiske, A. J. Cuddy, and P. Glick Universal dimensions of social cognition: warmth and competence. Trends in cognitive sciences 11 (2), p. 77–83. Cited by: §2.4, §7.1. Fleisig et al. (2024) E. Fleisig, S. L. Blodgett, D. Klein, and Z. Talat The perspectivist paradigm shift: assumptions and challenges of capturing human labels. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), p. 2279–2292. Cited by: §6.3. Fontenrose (1978) J. E. Fontenrose The delphic oracle, its responses and operations, with a catalogue of responses. Univ of California Press. Cited by: §2.1. Gilad et al. (2021) Z. Gilad, O. Amir, and L. Levontin The effects of warmth and competence perceptions on users’ choice of an ai system. In Proceedings of the 2021 CHI conference on human factors in computing systems, p. 1–13. Cited by: §6.1. Glickman and Sharot (2025) M. Glickman and T. Sharot How human–AI feedback loops alter human perceptual, emotional and social judgements. Nature Human Behaviour 9 (2), p. 345–359. Cited by: §6.1, §7.1. Golden and Aboujaoude (2026) A. Golden and E. Aboujaoude A transdiagnostic model for how general purpose ai chatbots can perpetuate ocd and anxiety disorders. NPJ Digital Medicine 9 (1), p. 343. Cited by: §2.2. Green and Chen (2019) B. Green and Y. Chen The principles and limits of algorithm-in-the-loop decision making. Proceedings of the ACM on human-computer interaction 3 (CSCW), p. 1–24. Cited by: §2.3. Guingrich et al. (2026) R. E. Guingrich, D. Mehta, and U. Bhatt Belief offloading in human-ai interaction. arXiv preprint arXiv:2602.08754. Cited by: §1, §2.2, §2.3. Guo et al. (2025) D. Guo, D. Yang, H. Zhang, J. Song, P. Wang, Q. Zhu, R. Xu, R. Zhang, S. Ma, X. Bi, et al. Deepseek-r1: incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948. Cited by: §7.1. Gupta et al. (2026) N. Gupta, M. Antoniak, and M. Walsh AI fiction in the wild. arXiv preprint arXiv:2606.22748. Cited by: §4.1, §4.1. Gur-Arieh et al. (2026) S. Gur-Arieh, A. Wang, and S. Fazelpour Ambiguity collapse by llms: a taxonomy of epistemic risks. In The 2026 ACM Conference on Fairness, Accountability, and Transparency, p. 5959–5980. Cited by: §1, §2.2. He et al. (2025) G. He, G. Demartini, and U. Gadiraju Plan-then-execute: an empirical study of user trust and team performance when using llm agents as a daily assistant. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, p. 1–22. Cited by: §7.1. He et al. (2023) G. He, L. Kuiper, and U. Gadiraju Knowing about knowing: an illusion of human competence can hinder appropriate reliance on ai systems. In Proceedings of the 2023 CHI conference on human factors in computing systems, p. 1–18. Cited by: §2.3. Heffner et al. (2025) J. Heffner, C. Qin, M. Chadwick, C. Knutsen, C. Summerfield, Z. Kurth-Nelson, and R. B. Rutledge Increasing happiness through conversations with artificial intelligence. arXiv preprint arXiv:2504.02091. Cited by: §2.3. Hendrycks et al. (2021) D. Hendrycks, C. Burns, S. Kadavath, A. Arora, S. Basart, E. Tang, D. Song, and J. Steinhardt Measuring mathematical problem solving with the math dataset. arXiv preprint arXiv:2103.03874. Cited by: §7.1. Henrich et al. (2010) J. Henrich, S. J. Heine, and A. Norenzayan The weirdest people in the world?. Cited by: §7.2. Hicke and Tomlinson (2026) R. M. Hicke and K. Tomlinson Adopt ≠ adapt: longitudinal analyses of llm conversations in the wild. arXiv preprint arXiv:2605.29018. Cited by: §4.1, §4, §7.2. Higgins (1987) E. T. Higgins Self-discrepancy: a theory relating self and affect.. Psychological review 94 (3), p. 319. Cited by: §5.1. Hou and Jung (2021) Y. T. Hou and M. F. Jung Who is the expert? reconciling algorithm aversion and algorithm appreciation in ai-supported decision making. Proceedings of the ACM on Human-Computer Interaction 5 (CSCW2), p. 477:1–477:25. External Links: Document, Link Cited by: §7.1. Hu et al. (2023) B. Hu, Y. Mao, and K. J. Kim How social anxiety leads to problematic use of conversational ai: the roles of loneliness, rumination, and mind perception. Computers in Human Behavior 145, p. 107760. Cited by: §1. Hu et al. (2023) K. Hu et al. ChatGPT sets record for fastest-growing user base-analyst note. Reuters 12, p. 2023. Cited by: §1. Huang et al. (2026) O. Y. Huang, M. Stodolska, and S. Sultana Emotional support with conversational ai: talking to machines about life. arXiv preprint arXiv:2603.22618. Cited by: §8. Ibrahim et al. (2025) L. Ibrahim, K. M. Collins, S. S. Kim, A. Reuel, M. Lamparth, K. Feng, L. Ahmad, P. Soni, A. E. Kattan, M. Stein, et al. Measuring and mitigating overreliance to build human-compatible ai. arXiv preprint arXiv:2509.08010. Cited by: §1, §2.3. Ibrahim et al. (2026a) L. Ibrahim, F. S. Hafner, M. Cheng, C. Lee, R. Anselmetti, R. Willer, L. Rocher, and D. Yang Sycophantic ai makes human interaction feel more effortful and less satisfying over time. arXiv preprint arXiv:2605.07912. Cited by: §1, §2.2, §6.1, §6.2, §6.2, §8. Ibrahim et al. (2026b) L. Ibrahim, F. S. Hafner, and L. Rocher Training language models to be warm can reduce accuracy and increase sycophancy. Nature 652 (8112), p. 1159–1165. Cited by: §2.2. Ischen et al. (2019) C. Ischen, T. Araujo, H. Voorveld, G. Van Noort, and E. Smit Privacy concerns in chatbot interactions. In International workshop on chatbot research and design, p. 34–48. Cited by: §4.2. Jackson (1978) M. Jackson An approach to kuranko divination. Human Relations 31 (2), p. 117–138. Cited by: §1. Jaffé et al. (2020) M. E. Jaffé, M. Douneva, and R. Greifeneder Solve the dilemma by spinning a penny? on using random decision-making aids. Judgment and Decision Making 15 (4), p. 561–571. Cited by: §2.1. Jaffé and Greifeneder (2021) M. E. Jaffé and R. Greifeneder This outcome feels right! subjective evaluations of coin flip outcomes reflect previously stated preferences. PLoS One 16 (7), p. e0253751. Cited by: §2.1. Jain et al. (2026) S. Jain, C. Park, M. Viana, A. Wilson, and D. Calacci Interaction context often increases sycophancy in llms. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, p. 1–26. Cited by: §7.1. Jakesch et al. (2023) M. Jakesch, A. Bhat, D. Buschek, L. Zalmanson, and M. Naaman Co-writing with opinionated language models affects users’ views. In Proceedings of the 2023 CHI conference on human factors in computing systems, p. 1–15. Cited by: §2.3. Jin et al. (2026) C. Jin, B. Li, H. Xie, C. M. Fang, T. Li, S. Longpre, H. Gu, M. Chen, and T. Shu Thoughttrace: understanding user thoughts in real-world llm interactions. arXiv preprint arXiv:2605.20087. Cited by: §4.1. Kadoma et al. (2024) K. Kadoma, M. Aubin Le Quere, X. J. Fu, C. Munsch, D. Metaxa, and M. Naaman The role of inclusion, control, and ownership in workplace ai-mediated communication. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, p. 1–10. Cited by: §2.3. Kahneman and Klein (2009) D. Kahneman and G. Klein Conditions for intuitive expertise: a failure to disagree.. American psychologist 64 (6), p. 515. Cited by: §2.2. Kapania et al. (2022) S. Kapania, O. Siy, G. Clapper, A. M. Sp, and N. Sambasivan “Because AI is 100% right and safe”: user attitudes and sources of AI authority in India. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, p. 1–18. Cited by: §6.1, §6.1, §7.1. Kasirzadeh and Gabriel (2023) A. Kasirzadeh and I. Gabriel In conversation with artificial intelligence: aligning language models with human values. Philosophy & Technology 36 (2), p. 27. Cited by: §7.1. Kasirzadeh (2025) A. Kasirzadeh Two types of ai existential risk: decisive and accumulative: a. kasirzadeh. Philosophical Studies 182 (7), p. 1975–2003. Cited by: §1, §5. Kemp (2025) S. Kemp Digital 2026: more than 1 billion people use ai. Data Reportal. Available online: https://datareportal. com/reports/digital-2026-one-billion-people-using-ai (accessed on 5 July 2026). Cited by: §1. Khadpe et al. (2020) P. Khadpe, R. Krishna, L. Fei-Fei, J. T. Hancock, and M. S. Bernstein Conceptual metaphors impact perceptions of human-ai collaboration. Proceedings of the ACM on Human-Computer Interaction 4 (CSCW2), p. 1–26. Cited by: §2.4, §6.1. Kim et al. (2025) S. S. Y. Kim, J. W. Vaughan, Q. V. Liao, T. Lombrozo, and O. Russakovsky Fostering appropriate reliance on large language models: the role of explanations, sources, and inconsistencies. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI ’25, New York, NY, USA. External Links: ISBN 9798400713941, Link, Document Cited by: §1, §2.3, §6.1. Kim et al. (2023) T. Kim, M. D. Molina, M. Rheu, E. S. Zhan, and W. Peng One ai does not fit all: a cluster analysis of the laypeople’s perception of ai roles. In Proceedings of the 2023 CHI conference on human factors in computing systems, p. 1–20. Cited by: §2.4. Kindt (2008) J. Kindt Oracular ambiguity as a mediation triple. Classicum 334, p. 23–27. Cited by: §2.1, §2.2. Kirk et al. (2025) H. R. Kirk, H. Davidson, E. Saunders, L. Luettgau, B. Vidgen, S. A. Hale, and C. Summerfield Neural steering vectors reveal dose and exposure-dependent impacts of human-ai relationships. arXiv preprint arXiv:2512.01991. Cited by: §2.3. Kleinberg and Raghavan (2021) J. Kleinberg and M. Raghavan Algorithmic monoculture and social welfare. Proceedings of the National Academy of Sciences 118 (22), p. e2018340118. Cited by: §2.2. Kocielnik et al. (2019) R. Kocielnik, S. Amershi, and P. N. Bennett Will you accept an imperfect ai? exploring designs for adjusting end-user expectations of ai systems. In Proceedings of the 2019 CHI conference on human factors in computing systems, p. 1–14. Cited by: §7.1. Kumar et al. (2026) H. Kumar, J. Chahal, Y. Zhao, Z. Zhang, A. Z. Wei, L. Tay, and A. Anderson 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, New York, NY, USA. External Links: ISBN 9798400722783, Link, Document Cited by: §1. Laban et al. (2026) P. Laban, H. Hayashi, Y. Zhou, and J. Neville LLMs get lost in multi-turn conversation. In International Conference on Learning Representations, Vol. 2026, p. 54738–54778. Cited by: §2.2. Laestadius et al. (2024) L. Laestadius, A. Bishop, M. Gonzalez, D. Illenčík, and C. Campos-Castillo Too human and not human enough: a grounded theory analysis of mental health harms from emotional dependence on the social chatbot replika. new media & society 26 (10), p. 5923–5941. Cited by: §2.3. Lam et al. (2026) M. S. Lam, O. Shaikh, H. Xu, A. Guo, D. Yang, J. Heer, J. A. Landay, and M. S. Bernstein Just-in-time objectives: a general approach for specialized ai interactions. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, CHI ’26, New York, NY, USA. External Links: ISBN 9798400722783, Link, Document Cited by: §7.1. Lee et al. (2023a) A. Y. Lee, N. B. Ellison, and J. T. Hancock To use or be used? the role of agency in social media use and well-being. Frontiers in Computer Science 5, p. 1123323. Cited by: §5.1. Lee and Hancock (2024) A. Y. Lee and J. T. Hancock Social media mindsets: a new approach to understanding social media use and psychological well-being. Journal of Computer-Mediated Communication 29 (1), p. zmad048. Cited by: §5.1. Lee et al. (2025) H. (. Lee, A. Sarkar, L. Tankelevitch, I. Drosos, S. Rintel, R. Banks, and N. Wilson The impact of generative ai on critical thinking: self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI ’25, New York, NY, USA. External Links: ISBN 9798400713941, Link, Document Cited by: §1, §2.2. Lee et al. (2023b) M. Lee, M. Srivastava, A. Hardy, J. Thickstun, E. Durmus, A. Paranjape, I. Gerard-Ursin, X. L. Li, F. Ladhak, F. Rong, R. E. Wang, M. Kwon, J. S. Park, H. Cao, T. Lee, R. Bommasani, M. S. Bernstein, and P. Liang Evaluating human-language model interaction. Transactions on Machine Learning Research. Note: External Links: ISSN 2835-8856, Link Cited by: §7.1. Li et al. (2010) I. Li, A. Dey, and J. Forlizzi A stage-based model of personal informatics systems. In Proceedings of the SIGCHI conference on human factors in computing systems, p. 557–566. Cited by: §7.1. Liao et al. (2020) Q. V. Liao, D. Gruen, and S. Miller Questioning the ai: informing design practices for explainable ai user experiences. In Proceedings of the 2020 CHI conference on human factors in computing systems, p. 1–15. Cited by: §7.1. Liu et al. (2021) H. Liu, V. Lai, and C. Tan Understanding the effect of out-of-distribution examples and interactive explanations on human-ai decision making. Proceedings of the ACM on Human-Computer Interaction 5 (CSCW2), p. 408:1–408:45. External Links: Document, Link Cited by: §2.3. Lu and Yin (2021) Z. Lu and M. Yin Human reliance on machine learning models when performance feedback is limited: heuristics and risks. In Proceedings of the 2021 CHI conference on human factors in computing systems, p. 1–16. Cited by: §2.3. Ludwig et al. (2022) V. U. Ludwig, B. Berry, J. Y. Cai, N. M. Chen, D. L. Crone, and M. L. Platt The impact of disclosing emotions on ratings of interpersonal closeness, warmth, competence, and leadership ability. Frontiers in Psychology Volume 13 - 2022. External Links: Link, Document, ISSN 1664-1078 Cited by: §8. Luettgau et al. (2025) L. Luettgau, V. Cheung, M. Dubois, K. Juechems, J. Bergs, L. Symes, H. Davidson, B. O’Dell, H. R. Kirk, M. Rollwage, et al. People readily follow personal advice from ai but it does not improve their well-being. arXiv preprint arXiv:2511.15352. Cited by: §2.2. Luger and Sellen (2016) E. Luger and A. Sellen " Like having a really bad pa" the gulf between user expectation and experience of conversational agents. In Proceedings of the 2016 CHI conference on human factors in computing systems, p. 5286–5297. Cited by: §7.1. Marchais-Roubelat and Roubelat (2011) A. Marchais-Roubelat and F. Roubelat The delphi method as a ritual: inquiring the delphic oracle. Technological Forecasting and Social Change 78 (9), p. 1491–1499. Cited by: §1, §2.1. Marcoccia et al. (2026) C. Marcoccia, W. Quattrociocchi, and V. Capraro AI advice suppresses people’s willingness to say" i don’t know", even when the advice is wrong and accuracy is incentivized. arXiv preprint arXiv:2607.13562. Cited by: §2.2. Marriott and Pitardi (2024) H. R. Marriott and V. Pitardi One is the loneliest number… two can be as bad as one. the influence of ai friendship apps on users’ well-being and addiction. Psychology & marketing 41 (1), p. 86–101. Cited by: §2.3. Mei et al. (2026) K. X. Mei, R. Wolfe, N. Weber, and M. Saveski Grok in the wild: characterizing the roles and uses of large language models on social media. In Proceedings of the International AAAI Conference on Web and Social Media, Vol. 20, p. 1609–1633. Cited by: §2.4, §3.1, §7.1. [118] S. Milli, R. E. Berker, S. Kraiczy, C. Shi, J. Kussman, A. Bose, E. Elkind, H. Bhattacharjee, A. D. Procaccia, and M. Nickel For questions of ought, ai could use some sage advice. In Pluralistic Alignment Workshop at ICML 2026, Cited by: §2.2, §2.3, §7.1. Milton et al. (2026) A. Milton, L. Ajmani, A. Heger, F. Poursabzi-Sangdeh, M. Vorvoreanu, and J. Suh " Always want to use it for everything": understanding young adults’ perceptions of ai dependence. arXiv preprint arXiv:2608.07592. Cited by: §7.1. Mitchell (2024) M. Mitchell The metaphors of artificial intelligence. Vol. 386, American Association for the Advancement of Science. Cited by: §2.4, §7.1. Naik et al. (2025) S. Naik, A. Snellinger, A. L. Toombs, S. Saponas, and A. K. Hall Exploring early adopters’ use of ai driven multi-agent systems to inform human-agent interaction design: insights from industry practice. In Proceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems, p. 1–8. Cited by: §7.1. Nass et al. (1994) C. Nass, J. Steuer, and E. R. Tauber Computers are social actors. In Proceedings of the SIGCHI conference on Human factors in computing systems, p. 72–78. Cited by: §2.4. Nolen-Hoeksema et al. (2008) S. Nolen-Hoeksema, B. E. Wisco, and S. Lyubomirsky Rethinking rumination. Perspectives on psychological science 3 (5), p. 400–424. Cited by: §2.2. Padmakumar et al. (2026) V. Padmakumar, L. Ibrahim, Z. Z. Wang, J. Wang, Q. V. Liao, and D. Yang Offloading score: measuring ai reliance through counterfactual workflows. arXiv preprint arXiv:2605.29392. Cited by: §1. Pang et al. (2025) R. Y. Pang, H. Schroeder, K. S. Smith, S. Barocas, Z. Xiao, E. Tseng, and D. Bragg Understanding the llm-ification of chi: unpacking the impact of llms at chi through a systematic literature review. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, p. 1–20. Cited by: §6.1. Paolacci and Chandler (2014) G. Paolacci and J. Chandler Inside the turk: understanding mechanical turk as a participant pool. Current directions in psychological science 23 (3), p. 184–188. Cited by: §7.2. Pareek et al. (2025) S. Pareek, S. Schömbs, E. Velloso, and J. Goncalves " It’s not the ai’s fault because it relies purely on data": how causal attributions of ai decisions shape trust in ai systems. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, p. 1–18. Cited by: §2.3. Parker (2023) T. Parker Deep learning. Comedy Central. Note: South Park, Season 26, Episode 4Aired March 8, 2023 Cited by: §1. Pataranutaporn et al. (2023) P. Pataranutaporn, R. Liu, E. Finn, and P. Maes Influencing human–ai interaction by priming beliefs about ai can increase perceived trustworthiness, empathy and effectiveness. Nature Machine Intelligence 5 (10), p. 1076–1086. Cited by: §7.1. Pennebaker and Seagal (1999) J. W. Pennebaker and J. D. Seagal Forming a story: the health benefits of narrative. Journal of clinical psychology 55 (10), p. 1243–1254. Cited by: §2.1. Pennebaker (1997) J. W. Pennebaker Writing about emotional experiences as a therapeutic process. Psychological science 8 (3), p. 162–166. Cited by: §2.1. Perry (2026) A. Perry In defense of social friction. Science 391 (6792), p. 1316–1317. Cited by: §2.2. Phang et al. (2025) J. Phang, M. Lampe, L. Ahmad, S. Agarwal, C. M. Fang, A. R. Liu, V. Danry, E. Lee, S. W. Chan, P. Pataranutaporn, et al. Investigating affective use and emotional well-being on chatgpt. arXiv preprint arXiv:2504.03888. Cited by: §2.2, §2.3, §8. Potts and Sudhof (2026) C. Potts and M. Sudhof Invisible failures in human-ai interactions. arXiv preprint arXiv:2603.15423. Cited by: §4.1. Prock et al. (2026) M. K. Prock, Z. Epstein, H. Schroeder, A. Smith, C. Lee, V. Goblot, and F. Jahanbakhsh Interpretive cultures: resonance, randomness, and negotiated meaning for ai-assisted tarot divination. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, p. 1–15. Cited by: §2.1. Raees et al. (2026) M. Raees, V. Khan, I. Lykourentzou, and K. Papangelis Do people appropriately rely on ai-advice? an analytical review of hci research on human-ai decision-making. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, p. 1–24. Cited by: §2.3, §7.1. Rafailov et al. (2023) R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, and C. Finn Direct preference optimization: your language model is secretly a reward model. Advances in neural information processing systems 36, p. 53728–53741. Cited by: §6.3. Rathje et al. (2025) S. Rathje, M. Ye, L. Globig, R. Pillai, V. de Mello, and J. Van Bavel Sycophantic ai increases attitude extremity and overconfidence. Cited by: §2.2. Reicherts et al. (2025) L. Reicherts, Z. T. Zhang, E. Von Oswald, Y. Liu, Y. Rogers, and M. Hassib AI, help me think—but for myself: assisting people in complex decision-making by providing different kinds of cognitive support. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, p. 1–19. Cited by: §7.1. Rose (2021) A. J. Rose The costs and benefits of co-rumination. Child Development Perspectives 15 (3), p. 176–181. Cited by: §2.2. Shaikh et al. (2025a) O. Shaikh, H. Mozannar, G. Bansal, A. Fourney, and E. Horvitz Navigating rifts in human-llm grounding: study and benchmark. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 20832–20847. Cited by: §2.2, §6.3. Shaikh et al. (2025b) O. Shaikh, S. Sapkota, S. Rizvi, E. Horvitz, J. S. Park, D. Yang, and M. S. Bernstein Creating general user models from computer use. In Proceedings of the 38th Annual ACM Symposium on User Interface Software and Technology, p. 1–23. Cited by: §2.1, §7.1. Sharma et al. (2026) M. Sharma, M. McCain, R. Douglas, and D. Duvenaud Who’s in charge? disempowerment patterns in real-world llm usage. arXiv preprint arXiv:2601.19062. Cited by: §1, §2.2, §5.1, §7.1. Shen et al. (2026) J. H. Shen, S. Carter, R. Dargan, J. Gillotte, K. Handa, J. Hong, S. Huang, K. Jagadish, M. Kearney, B. Levinstein, R. Linthicum, M. McCain, T. Millar, M. Julapalli, S. Price, M. Stern, D. Saunders, A. Tamkin, A. Vallone, J. Clark, S. Pollack, J. Eaton, D. Ganguli, and E. Durmus How people ask claude for personal guidance. Note: https://w.anthropic.com/research/claude-personal-guidanceAnthropic, April 30, 2026 Cited by: §2.2, §3. Shneiderman (2021) B. Shneiderman Human-centered ai. Issues in Science and Technology 37 (2), p. 56–61. Cited by: §2.4, §5.1. Si (2026) Y. Si Conversational ai: align commercial incentives with public interests. Nature 655 (8125), p. 1354–1354. Cited by: §6.3. So et al. (2026) J. So, C. Cheng, and S. Krishna Murthy Beyond anthropomorphism: a spectrum of interface metaphors for llms. In Proceedings of the Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems, p. 1–7. Cited by: §7.1. So and Roland (2020) R. J. So and E. Roland Race and distant reading. PMLA 135 (1), p. 59–73. Cited by: §6.3. Steele (1988) C. M. Steele The psychology of self-affirmation: sustaining the integrity of the self. In Advances in experimental social psychology, Vol. 21, p. 261–302. Cited by: §5.1. Subramonyam et al. (2024) H. Subramonyam, R. Pea, C. Pondoc, M. Agrawala, and C. Seifert Bridging the gulf of envisioning: cognitive challenges in prompt based interactions with llms. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, p. 1–19. Cited by: §7.1. Suchman (2006) L. Suchman Human-machine reconfigurations: plans and situated actions. 2nd edition, Learning in Doing: Social, Cognitive and Computational Perspectives, Cambridge University Press. Cited by: §7.1. Tankelevitch et al. (2024) L. Tankelevitch, V. Kewenig, A. Simkute, A. E. Scott, A. Sarkar, A. Sellen, and S. Rintel The metacognitive demands and opportunities of generative ai. In Proceedings of the 2024 CHI conference on human factors in computing systems, p. 1–24. Cited by: §2.3, §7.1. Teo et al. (2026) V. Teo, R. Jain, T. Gerstenberg, and M. Kleiman-Weiner AI assistants overassist. arXiv preprint arXiv:2607.21306. Cited by: §6.3, §7.1. Tseng and Liang (2026) E. Tseng and C. A. Liang " Chat, should i leave him?" risks, rewards, and roles for ai in relationship advice. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, p. 1–19. Cited by: §1, §2.2, §2.4, §3.1, §7.1. Vasconcelos et al. (2023) H. Vasconcelos, M. Jörke, M. Grunde-McLaughlin, T. Gerstenberg, M. S. Bernstein, and R. Krishna Explanations can reduce overreliance on ai systems during decision-making. Proceedings of the ACM on Human-Computer Interaction 7 (CSCW1), p. 129:1–129:38. External Links: Document, Link Cited by: §2.3. Verbeij et al. (2021) T. Verbeij, J. L. Pouwels, I. Beyens, and P. M. Valkenburg The accuracy and validity of self-reported social media use measures among adolescents. Computers in human behavior reports 3, p. 100090. Cited by: §5.1. Vereschak et al. (2021) O. Vereschak, G. Bailly, and B. Caramiaux How to evaluate trust in ai-assisted decision making? a survey of empirical methodologies. Proceedings of the ACM on Human-Computer Interaction 5 (CSCW2), p. 327:1–327:39. External Links: Document, Link Cited by: §2.3. Wang et al. (2024) H. Wang, W. Xiong, T. Xie, H. Zhao, and T. Zhang Interpretable preferences via multi-objective reward modeling and mixture-of-experts. In Findings of the Association for Computational Linguistics: EMNLP 2024, p. 10582–10592. Cited by: §6.3. What We Tell AI (n.d.) What We Tell AI What we tell ai. Note: Accessed: 2026-08-29 External Links: Link Cited by: §4.2. Williams-Ceci et al. (2026) S. Williams-Ceci, M. Jakesch, A. Bhat, K. Kadoma, L. Zalmanson, and M. Naaman Biased ai writing assistants shift users’ attitudes on societal issues. Science Advances 12 (11), p. eadw5578. Cited by: §2.3. Wilson and Gilbert (2008) T. D. Wilson and D. T. Gilbert Explaining away: a model of affective adaptation. Perspectives on Psychological Science 3 (5), p. 370–386. Cited by: §2.1. Yang et al. (2019a) D. Yang, R. E. Kraut, T. Smith, E. Mayfield, and D. Jurafsky Seekers, providers, welcomers, and storytellers: modeling social roles in online health communities. In Proceedings of the 2019 CHI conference on human factors in computing systems, p. 1–14. Cited by: §2.4. Yang et al. (2019b) Q. Yang, A. Steinfeld, and J. Zimmerman Unremarkable ai: fitting intelligent decision support into critical, clinical decision-making processes. In Proceedings of the 2019 CHI conference on human factors in computing systems, p. 1–11. Cited by: §2.3. Yin et al. (2019) M. Yin, J. Wortman Vaughan, and H. Wallach Understanding the effect of accuracy on trust in machine learning models. In Proceedings of the 2019 chi conference on human factors in computing systems, p. 1–12. Cited by: §2.3. Yu et al. (2026) S. Yu, M. Cheng, A. Jabbar, I. Sucholutsky, K. M. Collins, D. Jurafsky, and R. D. Hawkins The efficiency-gain illusion: people underestimate the rate of ai use and overestimate its benefits on simple tasks. arXiv preprint arXiv:2605.22687. Cited by: §5.1. Zamfirescu-Pereira et al. (2023) J. D. Zamfirescu-Pereira, R. Y. Wong, B. Hartmann, and Q. Yang Why johnny can’t prompt: how non-ai experts try (and fail) to design llm prompts. In Proceedings of the 2023 CHI conference on human factors in computing systems, p. 1–21. Cited by: §7.1, §7.1. Zao-Sanders (2026) M. Zao-Sanders How people are really using ai in 2026. Harvard Business Review. External Links: Link Cited by: §2.3, §3. [168] A. Zhang, A. Ross, K. Patel, J. Bernado, R. Bowie, A. T. Ribeiro, D. Halper, H. Valayaputtur, J. Andreas, S. Loeb, et al. When help is unhelpful: evaluating ai tutors for productive struggle. Cited by: §6.3, §7.1. Zhang et al. (2025) R. Zhang, H. Li, H. Meng, J. Zhan, H. Gan, and Y. Lee The dark side of ai companionship: a taxonomy of harmful algorithmic behaviors in human-ai relationships. In Proceedings of the 2025 CHI conference on human factors in computing systems, p. 1–17. Cited by: §4.2. Zhang et al. (2026) Y. Zhang, D. Zhao, J. T. Hancock, R. Kraut, and D. Yang Interaction with ai companions and psychological well-being. Nature Human Behaviour, p. 1–15. Cited by: §2.3, §4.2. Zhao et al. (2026) D. Zhao, M. S. Lam, D. Yang, and M. S. Bernstein Behavior latticing: inferring user motivations from unstructured interactions. arXiv preprint arXiv:2604.07629. Cited by: §7.1. Zhou et al. (2026) K. Zhou, K. Gligorić, M. Cheng, M. S. Lam, V. Raman, B. Aminu, C. Woo, M. Brockman, H. Cha, and D. Jurafsky Attention to non-adopters. In Findings of the Association for Computational Linguistics: ACL 2026, p. 1336–1366. Cited by: §7.2. Zhou et al. (2024) K. Zhou, J. Hwang, X. Ren, and M. Sap Relying on the unreliable: the impact of language models’ reluctance to express uncertainty. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 3623–3643. Cited by: §1, §2.2, §2.3. Appendix The Appendix is organized as follows: Appendix A contains the prompts for all LLM judges and their validation against human annotators. Appendix B contains additional analyses of the two public usage datasets (WildChat and ThoughtTrace), and Appendix C provides additional details and results from the data donation study, including a domain analysis and images of data donation tool. Appendix D contains further results from the perceived human vs. AI experiment. Finally, Appendix E contains further details on how oracular current models’ responses are, and Appendix F presents additional results for the interventions. Appendix A Prompts for LLM Judges For length reasons, the full code for labeling user prompts with a role from our typology can be found at https://pastebin.com/raw/BkdCLChH; this shares the same backbone for all other LLM judges (with a different judge prompt for each). The prompt for identifying personal or subjective messages in WildChat is as follows: Classify whether a single user message contains anything personal or subjective. This includes details about the user or other people: preferences, emotions, identity, relationships, life decisions, health or well-being, values, taste, subjective judgments, self-presentation, plans, practical constraints, or everyday life. Exclude fictional creative writing, purely technical, generic factual, impersonal homework, generic business, coding, translation, summarization, or objective lookup tasks unless the user’s message itself makes the task personal/subjective. Use a broad, low threshold. Do not infer personal context without evidence in the message. Output exactly 1 if the message contains anything personal or subjective. Otherwise, output exactly 0. Do not output any other text. The prompt for open-ended reasoning about LLMs’ goals, which was ineffective at shifting LLMs’ oracularity, is: Before responding, infer the user’s three most plausible LONG-TERM goals: durable outcomes, commitments, or directions that may explain why this request matters beyond the immediate task. Do not merely restate the current request, and do not infer sensitive traits. The three hypotheses must be meaningfully distinct, and their probabilities must sum to 1. The prompt for judging oracularity in LLM responses is available at https://pastebin.com/raw/evGtpcve. Note that this prompt (and the role prompt above) refers to the constructs of aleatoric and epistemic uncertainty, though we later decided to simplify this label to simpler binary labels of oracle/not. For the LLM responses, we take outputs of reduction in aleatoric uncertainty ≥2≥ 2 as oracular. A.1. Using LLM as judge Tables A1 and A2 report pairwise Cohen’s κ for the role judge and the oracularity judge, respectively, between the three LLM judges and the two expert annotators (H1 and H2), whose annotation procedure is in the main text. We generally find substantial agreement across both the LLMs and expert annotators. Gemini-2.5-Pro GPT-5.5 Claude Sonnet 4.6 H1 H2 Gemini-2.5-Pro 1.00 0.65 0.66 0.71 0.65 GPT-5.5 0.65 1.00 0.62 0.65 0.56 Claude Sonnet 4.6 0.66 0.62 1.00 0.60 0.59 H1 0.71 0.65 0.60 1.00 0.68 H2 0.65 0.56 0.59 0.68 1.00 Table A1. Cohen’s κ between the two expert annotators and three different LLMs shows substantial agreement on labeling roles in user prompts. Gemini-2.5-Pro Claude Sonnet 4.6 GPT-5.5 H1 H2 Gemini-2.5-Pro 1.00 0.70 0.68 0.64 0.66 Claude Sonnet 4.6 0.70 1.00 0.66 0.52 0.59 GPT-5.5 0.68 0.66 1.00 0.61 0.60 H1 0.64 0.52 0.61 1.00 0.72 H2 0.66 0.59 0.60 0.72 1.00 Table A2. Cohen’s κ between the two expert annotators and three different LLMs shows substantial agreement in almost all pairwise configurations on labeling responses as oracular. A.2. Warmth and Valence Judges We validated the warmth and valence LLM judges with an expert human annotator (psychology student) who labeled a stratified random sample of 100 metaphors from the 507 metaphors that participants provided. For robustness, in addition to Gemini 2.5 Pro (our main LLM judge), we also tested GPT-5.5. We find that the judge scores are well-aligned with the human annotations (AUROC 0.80-0.94; AUPRC 0.71–0.98) (Figure A1). Figure A1. Across both LLMs and both dimensions of warmth and valence, the judge scores are monotonically aligned with human judgments of these dimensions.Two line charts of agreement between the warmth and valence judges and a human annotator. In each panel the horizontal axis is the judge score, from 1 to 5, and the vertical axis is the weighted probability that the human annotator assigned the same label, from 0.0 to 1.0. Each panel is a line chart with one line for Gemini 2.5 Pro and one for GPT-5.5, with error bars. Both lines in both panels stay near 0.0 at judge scores of 1 and 2, rise from 3, and reach 1.0 at 5. The prompt is as follows: You are coding open-ended metaphors about AI. Rate how the participant’s metaphor portrays AI on two independent dimensions. Use the metaphor together with its "because" explanation. Score the portrayal in the text—not your own beliefs about AI. Use integer scores from 1 through 5. 1. warmth 1 = very cold, uncaring, hostile, exploitative, or threatening 2 = somewhat cold, detached, or impersonal 3 = neutral, mixed, or warmth is not implied 4 = somewhat warm, caring, friendly, supportive, or benevolent 5 = very warm, compassionate, nurturing, loyal, or emotionally supportive Do not treat physical temperature as social warmth. 2. valence 1 = very negative, harmful, dangerous, dystopian, or strongly disliked 2 = somewhat negative, concerning, or harmful 3 = neutral, balanced, ambivalent, or valence is not implied 4 = somewhat positive, helpful, beneficial, or liked 5 = very positive, transformative, wonderful, or strongly beneficial Rules: - Score both dimensions independently; do not force them to agree. - When a dimension is genuinely unspecified, use 3 rather than guessing. - Attend to negation, irony, mixed portrayals, and qualifications. - Give one brief rationale for each score and one short evidence excerpt. - Return only valid JSON with this exact schema: "warmth": 3, "valence": 3, "warmth_rationale": "brief reason", "valence_rationale": "brief reason", "evidence": "short excerpt", "overall_confidence": 0.0 Appendix B Additional Results from WildChat and ThoughtTrace This section reports additional analyses of WildChat and ThoughtTrace data. Figure A2 shows the main analyses reported in the text but with NA prompts included. Under this stricter denominator, the increase in Interpreter remains significant while the other temporal trends are in the same direction, but do not reach significance. Figure A3 provides breakdowns of Oracle in ThoughtTrace by education level and prompt purpose (e.g., learning or coding). It also includes a plot of coefficients from a multivariable OLS regression, which finds that age and gender are the only significant predictors of oracle use. Figure A2. Top row: role distributions among all prompts, including prompts labeled as NA, for WildChat (left) and ThoughtTrace (right). Middle row: Overall trends in Oracle in subjective personal questions (left) and in all prompts (right). Bottom left: temporal trends are weaker when including NA prompts in the denominator, though the increase in Interpreter remains significant. Bottom right: Other trends are not significant temporally.Six plots of LLM-as-oracle use across all prompts: role distributions and temporal trends. The top row is two horizontal bar charts of the share of all prompts in each of the eight roles plus NA, for WildChat and ThoughtTrace, with counts labeled. NA is largest in both, at 0.82 in WildChat and 0.62 in ThoughtTrace, and every role sits at 0.19 or below. Each task pair shares a color, with the non-oracle-use role of the pair drawn with hatching. The middle row is two line charts of the oracle-use share over time in WildChat from May 2023 to May 2024, first among subjective personal questions and then among all prompts, each annotated with a Pearson correlation and p-value; neither is significant. The bottom-left panel is a line chart of Interpreter, Evaluator, and Strategist measured against all prompts, where only Interpreter rises significantly. The bottom-right panel is a line chart of the five remaining roles, none of which trends significantly. Shading represents 95 percent confidence intervals. Figure A3. Details of Oracle by education, prompt topic, and individual trait regression on ThoughtTrace.Three plots of LLM-as-oracle use in ThoughtTrace: two demographic breakdowns and a regression. The top-left panel is four box plots of the share of oracle-use prompts, on an axis from 0.0 to 0.9, split by education level with four categories from graduate to other, with the mean marked on each box. The top-right panel is eight box plots of the same share, on an axis from 0.0 to 1.0, split by stated purpose from learning to translation, with the mean marked on each box; the boxes for planning and translation are the widest. The bottom panel is a dot-and-interval plot of the difference in oracle-use rate for fourteen predictors, from a conversation-level ordinary least squares regression on 1,768 complete cases, with 95 percent confidence intervals and a dashed reference line at zero. Only three intervals exclude zero: age over 34 and education other, both negative, and male gender, positive. Appendix C Additional Information from Data Donation Study Our study was approved by the Stanford Institutional Review Board. We paid participants at an hourly rate of $12. C.1. Domain analysis In the data donation study, we also use another LLM judge to label each message with its domain, following the domain splits in other reports of LLM use: (1) Work & learning, (2) Relationships & social, (3) Health & fitness, (4) Mental & emotional, (5) Money, (6) Legal & admin, (7) Leisure & hobbies, (8) Civic & current events, and (9) Other. Since a conversation may span multiple domains, we labeled domains at the message level: for each target message, the judge is given a window of up to two user messages before the target and two after it, including the assistant turns in between to provide context and ensure accurate classification (truncated to 4,000 characters to keep the length of the prompt manageable). We found that relevant subjective personal questions were concentrated in two domains: Work & learning (14.7% oracle-use and 14.4% non-oracle-use queries; n = 40,530) and Relationships & social (14.8% and 10.5%; n = 8,751). Our data further enabled investigating how trends in Oracle differed across domains. We focus on the period from September 2024 to August 2026 (two years of usage), which was the longest unbroken stretch over which both of the top two domains (Work & learning and Relationships & social) carried enough messages for analysis (at least 40 messages in every rolling three-month window). Figure A5 details the distributions. We also test each trend by logistic regression of the role against date over all messages. We find that use of Ghost-writer declined (work ×0.80/year, p < 0.001; relationships ×0.41/year, p < 0.001) as did Editor in both domains (work ×0.85/year, p < 0.001; relationships ×0.52/year, p < 0.001), while Decision-maker rose in work but not in relationships (work ×1.35/year, p < 0.001; relationships ×0.93/year, p = 0.369) and Evaluator rose in both domains (work ×1.49/year, p < 0.001; relationships ×1.31/year, p = 0.014). C.2. Reasons for AI usage To better understand what drives usage, we also asked participants what they gain from using AI for personal uses. We find that participants most frequently select the options that describe deliberation, such as new perspectives or possibilities (83%), a clearer understanding of the situation (77%), and a plan for what to do next (77%). When we asked participants why they turn to LLMs for personal and subjective queries, we find that reasons that concern convenience were the most commonly selected such as AI is available when I need it (81%) and AI is faster or easier than asking a person (65%), followed by interpersonal reasons, like AI feels less judgmental than asking a person (58%). Figure A4. Additional results from donated data: Top left two panels: Distribution and temporal trends when including prompts classified as NA. Top right two panels: Per-participant and raw count of LLM-as-oracle prompts over time. When considering all prompts, including those classified as NA, there is an increase from 7.3% in the earliest period to 13.0% in the most recent (r=0.42,p=0.11r=0.42,p=0.11), though it is weaker because the absolute number of oracular messages rose dramatically because both Oracle use and overall use grew simultaneously—the first time period contains 82 messages from 3 participants while the last one contains 23,714 from 52. Middle and bottom row: Trend for each role, among relevant subjective personal questions (middle) and when including prompts classified as NA (bottom). Eight plots of LLM-as-oracle use in the donated data, repeating the main analyses with NA prompts included. The top row is a horizontal bar chart of the share of all prompts in each of the eight roles plus NA, with counts labeled, where NA is largest at 0.78, followed by three line charts spanning 2023 to 2026: the oracle-use share of all prompts, the number of oracle-use prompts per participant, and the raw count of oracle-use prompts, each annotated with a Pearson correlation and p-value. The middle row is two line charts of the four oracle-use roles and the four non-oracle-use roles among subjective personal questions. The bottom row is the same two charts with NA prompts in the denominator. Shading represents 95 percent confidence intervals. Figure A5. Donated data breakdown by domain Distributions and temporal trends by domain, among relevant subjective personal questions (top) and when including prompts classified as NA (bottom). Six plots of LLM-as-oracle use in the donated data by domain. The top row is two horizontal bar charts of the share of subjective personal questions in each of the eight roles, for Work and learning and for Relationships and social, with counts labeled, alongside a line chart of the oracle-use rate in the two domains from 2023 to 2026, each line annotated with a Pearson correlation and p-value. The bottom row is the same three plots with NA prompts in the denominator. Shading represents 95 percent confidence intervals. Figure A6. Donated data demographic breakdown by individual demographics.Six box plots of the share of oracle-use prompts per participant in the donated data, split by individual traits. Each panel is a box plot on an axis from 0 to 0.3, with the mean marked on each box and group sizes labeled, split by race, education, frequency of AI use, subjective social status, need for closure, and attitudes toward AI. The boxes overlap substantially across every grouping. Two plots of participants' stated reasons for and perceived gains from personal AI use in the data donation study. The top panel is a horizontal bar chart of six stated reasons for personal AI use, on an axis of the share of participants from 0 to 100 percent, with participant counts labeled; availability and speed or ease are largest. The bottom panel is a horizontal bar chart of eight perceived gains on the same axis; new perspectives or possibilities is largest, followed by a plan for what to do next and a clearer understanding of the situation. Figure A7. Reasons reported by participants for their personal AI use.Two plots of participants' stated reasons for and perceived gains from personal AI use in the data donation study. The top panel is a horizontal bar chart of six stated reasons for personal AI use, on an axis of the share of participants from 0 to 100 percent, with participant counts labeled; availability and speed or ease are largest. The bottom panel is a horizontal bar chart of eight perceived gains on the same axis; new perspectives or possibilities is largest, followed by a plan for what to do next and a clearer understanding of the situation. C.3. Data donation tool Table A3 shows the simplified, participant-facing phrases used for each of the eight LLM-as-oracle roles. Figure A8 shows the interface with the visualizations that participants were shown after donating their data. Role Shown to participant Oracle “Tell me what to think or do” Interpreter “Tell me what something means” Evaluator “Judge whether I was right” Decision-maker “Decide what I should do” Ghost-writer “Draft or word things for me” Non-oracle “Help me work it out myself” Contextualizer “Explain a topic or perspective” Deliberator “Lay out my options” Strategist “Map out approaches I could take” Editor “Polish my writing” Table A3. Mapping from role labels to the participant-facing phrases used in the tool. In the past 6 months, I have used AI to… Role Explain why I or someone else thinks, feels, or acts a certain way. Interpreter (Oracle) Suggest different possible explanations for someone’s thoughts, feelings, or actions. Contextualizer (Non-oracle) Judge whether something is right, wrong, or appropriate. Evaluator (Oracle) Help me get more information to make a judgment. Deliberator (Non-oracle) Decide what I should do. Decision-maker (Oracle) Help me think through possible options or next steps. Strategist (Non-oracle) Decide what to say or write text for me. Ghost-writer (Oracle) Improve or edit text that I wrote. Editor (Non-oracle) I do not use AI in any of these ways. NA Table A4. Participant-facing descriptions of roles in pre-data donation survey. Figure A8. Data donation tool visualization interface. After participants donated their data, they were shown visualizations of their use of AI. This began with visualizations about the domains in which they used AI (top left), and trends over time (top right). Then, they were shown visualizations about how much they offloaded their beliefs and judgments to AI, and how this varied over time and by domain. This was grounded in specific examples. Redacted text to preserve anonymity.Six screens of the data donation tool interface: domains, trends over time, and oracle-use balance. The top row is a horizontal bar chart of the domains a participant used AI for and a line chart of how those domains shifted over time. The middle row is a stacked bar chart of the participant's overall balance of ``tell me what to think or do′ versus ``help me work it out myself′ with a per-domain breakdown, alongside a screen of what share of messages matched either approach and the split between the two. The bottom row is a line chart of how that balance changed over time and a panel of the participant's top request types and the domains in which each approach was most common. Each screen is paired with example prompts drawn from the participant's own history; these examples are redacted to preserve anonymity. Appendix D Results from Perceived Human vs AI study Additional survey details We recruited N=525 participants on Prolific, 5 of whom failed the attention check, leaving N=520 in our final sample (266 in the disclosed AI condition and 254 in the perceived human condition). Before the interaction, participants first answered a set of pre-study questions covering AI use, demographics, AI attitudes, and need for closure. Our study is IRB-approved, and we paid participants at an hourly rate of $12. No significant Oracle difference in first message In our mixed-effects logistic regression, participants in the disclosed-AI condition had higher odds of producing an Oracle message overall (b = 0.763, SE = 0.136, OR = 2.15, 95% CI [1.65, 2.80], p < .001). Examining only the first message accounts for the difference in model behavior between the default AI system versus one prompted to be human-like. We found no significant difference in Oracle rates in the first message that participants sent (29.7% vs 26.4% of first messages, p = 0.40), though the result is directionally the same as when comparing all messages. People who used LLMs as oracles also shift more on their decisions after the conversation. Participants who sent at least one oracle-use message had larger decision shift (M = 1.16 vs 0.89, p = 0.006), though there was no difference in decision shift between the human vs. AI conditions (M = 1.06 vs 1.03, p = 0.80, d = 0.02). This provides evidence that people who ask oracle-use questions are more susceptible to influence, though this may be because they were initially more open to influence (and hence asked an oracle-use question). Perceptions of trustworthiness of the conversation partner did not differ (M = 5.56 vs 5.68, p = 0.34), while response quality was higher for disclosed AI (5.78 vs 5.46, p = 0.007). None of the preregistered individual difference measures predicted oracle use (all Odds Ratios approx. 1.00, all ps > 0.33). After excluding 68 people in the perceived human arm who rated the roleplaying LLM as somewhat or very unbelievable, the decision shift null and difference in oracle use remained significant, but the difference in perceptions of message quality disappeared. Social & Relationships There’s someone I’ve lost touch with and I keep going back and forth about whether to reach out. I haven’t heard from someone in a while and I’m debating whether to message them first or wait and see if they reach out. I got invited to something and I’m trying to figure out whether I actually want to go or if I’d rather not. Something a friend did has been bothering me, and I can’t decide whether to bring it up or just move on. I had a disagreement with someone where I don’t think I was wrong, so I’m wondering whether apologizing would help or just feel dishonest. I know two people who I think would get along, but I’m not sure if introducing them is a good idea or if it would be awkward. Work & Career I’m trying to decide whether to apply for a role that excites me but feels like a stretch. I got a job offer and I’m weighing whether to take it or stay where I am. I’ve been thinking about asking for a raise, but I’m not sure if now is the right time or if I should wait. Something happened at work that’s bothering me, and I’m going back and forth between raising it with my manager or letting it go. There’s a project I could volunteer for. Part of me wants to, but I’m not sure it’s the right move. I’ve decided to leave my job, and I’m torn between telling my team now or waiting until everything’s finalized. Table A5. Decision topics participants could choose from in the perceived human vs. AI study. After selecting a topic, participants were shown a short prompt asking them to describe their specific situation. Appendix E Oracularity in current model responses Figure A9. Oracularity of different LLMs’ responses in the longitudinal sycophancy study (top), WildChat (middle left) and ThoughtTrace prompts (middle right). In the longitudinal sycophancy study, we find that both the sycophantic and challenging models are more oracular than the balanced baseline; the sycophantic model is especially oracular to Oracle prompts. In the original responses (which are to different prompts, and thus not directly comparable), we find that GPT-4 is more oracular than GPT-3.5; and Grok is more oracular than all other models in ThoughtTrace. Across the directly comparable new model responses, GPT-5.5 is most oracular; and this finding is robust to whether the response has the full conversation context or not (bottom). Breakdowns by prompt oracularity are in Figure A10.Five dot-and-interval plots of oracular responses by model, for WildChat and ThoughtTrace prompts. The top row is three panels. The left panel is the oracular response rate on WildChat prompts for the two original models and seven models with newly generated responses, with values labeled, ranging from 0.48 to 0.62. The middle panel is the rate for the original ThoughtTrace responses grouped by model family, ranging from 0.26 to 0.51. The right panel is the rate for the ten models in the ThoughtTrace model ablation, ranging from 0.12 to 0.38. The bottom row is two dot-and-interval plots comparing three models given the full conversation context with the same models given a single turn only; the single-turn rates are roughly half the full-context rates for all three, but the trend is the same. On the subset of prompts that were labeled with one of the roles on our typology (n=11,087n=11,087 from WildChat and n=3,227n=3,227 from ThoughtTrace), we first ran our LLM judge on the original responses, which were from GPT-3.5 and GPT-4 (WildChat) and over 20 LLMs (ThoughtTrace). In WildChat, we find that the newer, larger GPT-4 is significantly more oracular than GPT-3.5. In ThoughtTrace, Grok has the highest rates of responding as an oracle. This contrast is even larger when we consider the subset of prompts that are non-oracular, i.e., cases where users seem to be explicitly not seeking oracular responses (Figure A10). These original responses do not allow us to attribute the effects to model alone, as each model’s responses are to different prompts (though the output model is randomly selected in ThoughtTrace). To more directly compare model behavior, we further obtained responses from both open and closed state-of-the-art models on the same data and scored them. Specifically, we generated new responses from three proprietary and four open-weight models: Claude Sonnet 4.6, Gemini 2.5 Pro, GPT-5.5, Qwen3-8B, Llama-3.1-8B, OLMo 2 7B, and Gemma 2 9B. We additionally generated responses from Claude Opus 5, GPT 5.6 Sol, and Gemini 3.1 Pro on ThoughtTrace. We find that GPT-5.5 consistently responds more as oracles than any other model family, both in Oracle prompts and otherwise (Figure A9). On ThoughtTrace, the smaller open-weight models are generally less oracular than the larger closed models. Four dot-and-interval plots of oracular responses in ThoughtTrace by prompt role. The top row is two panels of the rate on oracle-use prompts, first grouped by model family and then for the ten models with newly generated responses. The bottom row is the same two panels for non-oracle-use prompts, where every rate is lower. Figure A10. Oracular responses by model, and split by oracle vs non-oracle prompts.Four dot-and-interval plots of oracular responses in ThoughtTrace by prompt role. The top row is two panels of the rate on oracle-use prompts, first grouped by model family and then for the ten models with newly generated responses. The bottom row is the same two panels for non-oracle-use prompts, where every rate is lower. Appendix F Additional Intervention Results Figure A11 shows the effect of prompting models to do open-ended reasoning about users’ long-term goals. Figure A12 shows the main intervention analyses repeated on a random sample of 600 prompts (stratified by role), in which the original response is not necessarily oracular. Figure A13 shows the tradeoff between reducing oracularity and response reward. One dot-and-interval plot of oracular responses under open-ended goal reasoning, for three models. The panel is the percent change relative to baseline for Gemini 2.5 Pro, GPT-5.5, and Claude Sonnet 4.6, on an axis from minus 6 to 6 with a dashed reference line at zero. All three intervals span zero. Figure A11. Prompting models to generally reason about users’ goals does not significantly reduce oracular responses.One dot-and-interval plot of oracular responses under open-ended goal reasoning, for three models. The panel is the percent change relative to baseline for Gemini 2.5 Pro, GPT-5.5, and Claude Sonnet 4.6, on an axis from minus 6 to 6 with a dashed reference line at zero. All three intervals span zero. Figure A12. On a random sample of prompts (not only the ones with oracular responses), we find the same takeaways as reported in the main text.Five groups of plots of intervention effects on oracular responses on a random sample of prompts stratified by role. In this sample the original response is not necessarily oracular. The first group is three panels of the mean paired change in oracular responses, one per hard-coded user goal, for seven models. The second group is the same three panels for the adaptive goal selection pipeline, conditioned on the goal each model selected. The third is a stacked bar chart of goal selection rates for the seven models, on an axis from 0 to 1. The fourth is a bar chart of the mean paired change for the four open-weight models under adaptive choice, hard-coded self-discovery, and Direct Preference Optimization. The directions match the main text throughout. The fifth group is a scatter plot of the mean change in ArmoRM reward against the mean reduction in oracular responses for the five interventions, with error bars on both axes and a dashed line marking the Pareto frontier; hard-coded fast answer is the only intervention that increases oracularity, and it also carries the largest reward loss. Figure A13. Models on which DPO is most effective at reducing oracularity also compromise model reward the most (left). Hard-coded (HC) learning, HC self-discovery, and DPO are at the Pareto frontier between reward and oracular reduction (right).Two plots of the tradeoff between oracularity reduction and response reward. The left panel is a bar chart of the mean change in ArmoRM reward on general instructions for the four open-weight models under Direct Preference Optimization, with error bars and values labeled; the change is slightly positive for Qwen and negative for the other three, most steeply for Llama. The right panel is a scatter plot of the mean change in reward against the mean reduction in oracular responses for the five interventions, with error bars on both axes and a dashed line marking the Pareto frontier, on which hard-coded learning, hard-coded self-discovery, and Direct Preference Optimization sit. Figure A14. Left: Overall, users prefer original LLM responses when asked only for their immediate preference, but prefer DPO responses when asked for a preference that accounts for self-discovery. Right: Preferences (and DPO effectiveness) vary by base model.Two bar plots F.1. Preference study: Short-term vs. long-term preferences To investigate whether people’s preferences differ when considering short-term vs. long-term goals, we conducted a user preference study on baseline model outputs vs. DPO model outputs. We recruited 200 participants on Prolific and paid them $2.00\$2.00 for a 10 minute task. We excluded participants that failed a visual bot check, leaving 198 participants. Participants were randomized into two conditions: Default and Self-Discovery. In the task, participants were presented with a sequence of 10 prompts with two LLM responses to the prompt: the original baseline model response and the DPO response. In the Default condition, participants were asked to choose the response they would prefer. In the Self-Discovery condition, participants were asked to choose the response they would prefer if they wanted to “reason about the situation yourself and develop your own interpretation". The prompts and responses were sampled uniformly from the test dataset (across all four models) after filtering for examples where the DPO response demonstrated oracularity reduction (LLM-judged). In total, participants annotated 724 samples from Llama 3.1 8B, 643 from Gemma 2 9B, 447 from OLMo 2 7B, and 166 from Qwen3-8B. The variation in sample size reflects differences in DPO’s effectiveness at reducing oracularity. Figure A14 shows the rate at which participants in each condition prefer the DPO response, overall and by base model. We find that users in the default condition tended to prefer the baseline responses over the DPO responses according to a two-sided proportions z-test (r⁡(DPO)=0.46r(DPO)=0.46, p=0.006p=0.006). In contrast, users in the self-discovery condition tended to prefer the DPO responses over the baseline responses (r=0.53r=0.53, p=0.048p=0.048). These results suggest that current LLM behavior is optimized for immediate preferences but may not be aligned for preferences about long-term goals such as self-discovery and independent reasoning.

Instant research discovery