Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Self-Disclosure and Privacy in Human–AI Relationships

Abstract. Self-disclosure to conversational artificial intelligence is an interpersonal-like communicative act embedded in a technical and institutional information system. This fundamental narrative research monograph examines the psychological processes that encourage, inhibit, and shape personal disclosure to AI chatbots and companions, together with the privacy expectations and downstream information flows that determine its consequences. It integrates interpersonal disclosure and intimacy theory, the social response to computers, anonymity and evaluation-apprehension research, communication privacy management, contextual integrity, behavioral privacy economics, human–computer interaction experiments, and emerging observational studies of generative AI. Experimental reports of reciprocal chatbot disclosure and anthropomorphic cues are compared with observed sharing behavior, surveys of mental health chatbot use, platform audits, and qualitative accounts of AI companionship. The synthesis identifies distinctions among disclosure willingness, actual disclosed content, perceived confidentiality, and legally or technically implemented protection. Methodological recommendations address sensitive-content coding, randomized identity and privacy-cue manipulations, longitudinal designs, context-sensitive risk measurement, and equitable sampling. This is a critical scholarly narrative synthesis without new participant data, a preregistered systematic review, or a pooled effect estimate.

Keywords: self-disclosure; human–AI relationships; AI companions; conversational agents; privacy calculus; communication privacy management; contextual integrity; perceived confidentiality; anthropomorphism; social presence; evaluation apprehension; psychological intimacy; sensitive information; trust calibration; relational boundaries; mental health chatbots; Artificial Era

Introduction: Disclosure as a Psychological and Institutional Event

Self-disclosure is the intentional communication of personally meaningful information about one's experiences, traits, preferences, relationships, vulnerabilities, or circumstances to an audience. The research object in human–AI relationships includes the content and depth of disclosure, the person's beliefs about who or what receives it, and the infrastructure that stores, processes, reuses, or transmits the information. A message may simultaneously be directed toward a conversational partner as experienced by the user and toward an organization operating the service. This dual audience becomes central when users attribute familiarity or discretion to a seemingly personal interface while institutional data handling follows a separate set of rules. The classic understanding of privacy as a capacity to regulate access provides an initial vocabulary, but relational disclosure also requires attention to expected purposes, recipients, retention, and the possibility of future use Westin (1967) Nissenbaum (2004) Chiu & Foote (2026).

This monograph defines a human–AI relationship in behavioral and experiential terms: recurrent interaction in which a person forms expectations about the system's responses, availability, memory, roles, or relevance to personal life. The definition permits investigation of momentary exchanges as well as sustained engagement with chatbots and companions. It assigns no unmeasured subjective states or human moral capacities to artificial systems. Psychologically consequential outcomes can arise through ordinary interface design, learning histories, conversational contingencies, and the person's interpretation of a response. Early demonstrations that people enact social conventions with computers established an empirical starting point for research on seemingly interpersonal disclosure to technological systems; current generative systems extend the scope and apparent responsiveness of such encounters Nass & Moon (2000) Moon (2000) Bickmore & Picard (2005).

The central scientific questions concern when disclosure begins, what makes it deepen, how users interpret apparent understanding, what categories of personal information become exposed, and how perceptions of privacy compare with actual information flows. The evidence spans controlled laboratory interactions, online experiments, surveys, interviews, platform audits, and theories from social psychology and information science. Each method samples a different part of the process. A survey of willingness is evidence about intention, while the words actually typed into a chat reveal behavior under specific conditions; a privacy notice describes an asserted policy, whereas a technical audit examines implemented affordances. The analysis therefore uses convergent evidence while keeping stated attitudes, observed behavior, and organizational guarantees distinct Acquisti et al. (2015) Merwin et al. (2025) Zhang & Zhu (2026).

Historical Origins of Interpersonal Self-Disclosure Research

The twentieth-century study of disclosure developed alongside research on intimacy, interpersonal attraction, and the regulation of access to private life. Altman and Taylor's social penetration account described relationships as changing through patterns of breadth and depth of interpersonal exchange. Breadth refers to the range of topics revealed; depth concerns how personally consequential or intimate those topics are. Although the original model addressed human relationships, it supplies measurable dimensions for studying conversations with artificial partners. The same topical breadth can conceal very different risks: a person might discuss many ordinary preferences yet withhold identifying or medically sensitive information, or reveal one highly consequential secret in an otherwise brief encounter Altman & Taylor (1973) Collins & Miller (1994).

Experimental and meta-analytic interpersonal research established a connection among disclosure, liking, and the relational meaning of reciprocal exchange. Collins and Miller's review examined how liking and personal disclosure covary and influence each other, rather than treating the amount of revealed information as a sufficient explanation of closeness. Laurenceau and colleagues later analyzed intimacy as a process involving one's own disclosure, a partner's disclosure, and perceived responsiveness. These findings suggest that a disclosure episode acquires relational significance through what the discloser believes happened after speaking. For AI interactions, this requires separate measures of users' felt responsiveness, the actual semantic accuracy of a generated reply, and the system's ability to handle the disclosed information appropriately Collins & Miller (1994) Laurenceau et al. (1998).

Privacy scholarship also broadened beyond secrecy. Westin conceptualized privacy as a form of individual control over access and communication. Petronio's communication privacy management theory approached private information as managed through boundaries and expectations about co-ownership once it is shared. Nissenbaum's contextual integrity later located appropriate information flow in the norms of a social context, including actors, attributes, and transmission principles. Applied to conversational systems, this trajectory makes a fundamental point: voluntarily telling a chatbot something does not settle whether later model training, staff access, third-party transfer, or persistent personalization fits the context in which the information was revealed. Disclosure and downstream permission are independently investigable questions Westin (1967) Petronio (2002) Nissenbaum (2004).

From Computer-Mediated Communication to Conversational Agents

Computer-mediated communication research examined how the absence or alteration of conventional social cues affects interaction before contemporary generative AI appeared. Walther's analysis of hyperpersonal communication described processes through which selective presentation, receiver interpretation, channel characteristics, and feedback can produce intense interpersonal impressions. Joinson's experiments explored self-disclosure under variations in visual anonymity and self-awareness. These studies demonstrate why the design of an encounter can change what people say without changing the underlying personal content they are capable of sharing. Their findings are historical and conceptual foundations, not direct estimates of behavior toward every contemporary conversational model Walther (1996) Joinson (2001).

Moon's experiments on consumer interactions with computers were especially consequential for the later literature on disclosure to machines. Computer prompts arranged according to interpersonal conventions could elicit information and reciprocal responding, even when the interlocutor was an ordinary computing device. Research on automatic social responses to computers provided a compatible account: people sometimes transfer familiar social scripts to nonhuman interfaces with little deliberation about the machine's nature. The observable behavior can be measured independently of whether users explicitly believe the device is a person. This distinction remains essential for analyzing anthropomorphic language models, because a highly personified exchange may be produced by conventional interaction rules as well as explicit attachment Moon (2000) Nass & Moon (2000).

Later relational-agent research introduced sustained dialogue, social signaling, and managed conversational history. Bickmore and Picard investigated the establishment and maintenance of long-term relationships with computational agents, showing why repeated interaction adds variables that a single disclosure questionnaire cannot capture. Across sessions, users may develop expectations about recognition, continuity, boundaries, and the cost of withdrawing. Generative systems introduce stronger open-ended language production, but historical continuity is scientifically relevant: personalization, recurrence, and the illusion or experience of mutual familiarity were studied before large language models. Comparative research should distinguish the effects of open-ended generation from the older effects of repetition, remembered details, and standardized empathic cues Bickmore & Picard (2005) Laurenceau et al. (1998).

Conceptual Boundaries: Disclosure, Trust, Intimacy, and Privacy

Disclosure is a behavior; willingness to disclose is a prospective judgment; intimacy is an experienced quality of connection; trust is an expectation about future conduct under vulnerability; and privacy concerns govern how information may circulate. These constructs frequently correlate, yet they occupy different positions in a causal account. A user can feel intensely understood after a short exchange while revealing little sensitive information. Another user may transmit deeply personal information to obtain a practical service without experiencing intimacy. Measures should therefore document the type of information shared, the sequence of interaction, beliefs about data use, experienced responsiveness, and the consequences for later choices instead of substituting one attractive social scale for the whole phenomenon Laurenceau et al. (1998) Acquisti et al. (2015) Zhang & Zhu (2026).

A further boundary distinguishes disclosure directed to an interface from disclosure about third parties. A person's first-person account of a difficult day poses different ethical and privacy questions from uploading private messages written by a partner, describing a child's medical history, or reproducing confidential workplace communications. In the latter situations, the person sending the message may lack authority to approve onward processing of information belonging to someone else. Communication privacy management treats private knowledge as governed by relational expectations and boundary rules; contextual integrity evaluates whether the recipient and information transmission fit the original social setting. Both traditions clarify why the discloser's voluntary action alone is an inadequate measure of appropriate processing Petronio (2002) Nissenbaum (2004).

Human–AI interactions create a distinctive combination of highly individualized language and externally controlled technical infrastructure. The conversational surface may foster an impression of secrecy, whereas the retention settings, access controls, employee privileges, deletion mechanisms, and legal obligations are established elsewhere. Researchers should make their referents explicit: perceived confidentiality means the user's belief about who can read the exchange; confidentiality as a technical property concerns actual access and protection; relational privacy concerns the norms governing personal meaning and disclosure; data protection concerns organizational and infrastructural controls. Chiu and Foote's qualitative examination of AI companions documents how users negotiate these levels and the practical difficulty of imposing interpersonal expectations on corporate systems Chiu & Foote (2026) Yener et al. (2025).

Research Epistemology and the Unit of Analysis

The smallest useful unit of disclosure research is an event with identifiable antecedents and outcomes: a prompt or social cue, a person's response, the specific information conveyed, the system's immediate treatment of that information, and the person's interpretation. A second unit is the repeated interaction sequence in which earlier revelations shape later ones. A third unit is the sociotechnical arrangement linking people, interfaces, data stores, developers, platform policies, and possible downstream audiences. These units support distinct kinds of inference. Event-level randomization identifies effects of experimentally varied prompts; longitudinal observation reveals temporal patterns; platform audit identifies data-handling features. Evidence from one unit should not automatically be presented as proof about another Moon (2000) Bickmore & Picard (2005) Chiu & Foote (2026).

The relevant counterfactual questions are concrete. Would the same person reveal the same detail if told that a human, AI model, or corporate analyst might receive it? Would a visible retention control change the disclosure, holding task benefits constant? Does a chatbot's apparent reciprocal disclosure change the user's response compared with equal-length neutral text? Does repeated interaction amplify sharing because of increased perceived understanding, convenience, remembered history, or reduced evaluation apprehension? Each alternative points to a manipulable feature and a measurable mechanism. Contemporary laboratory studies allow some of these comparisons, but platform ecosystems, participant selection, and the changing capabilities of deployed models complicate their external validity Lucas et al. (2014) Liang et al. (2024) Zhang & Zhu (2026).

A defensible synthesis avoids the assumption that increased disclosure is inherently beneficial or harmful. Sharing can support reflection, access to information, rehearsal of difficult conversations, or perceived emotional relief, while also creating exposure, regret, breaches of interpersonal commitments, and opportunities for inference from accumulated records. The relevant evaluative question concerns disclosure relative to the person's goals, autonomy, understanding of information flows, and the safeguards actually available. Evidence regarding mental-health uses illustrates the importance of this approach: disclosure of distress may carry different psychological value and risk than an ordinary shopping preference, and self-selected users of mental-health chatbots cannot by themselves establish treatment effectiveness or safety McBain et al. (2026) Yener et al. (2025) Zhang et al. (2026).

Evaluation Apprehension, Anonymity, and the Nonjudgmental Audience

People decide whether to reveal sensitive information partly by anticipating how an audience might evaluate them. An interlocutor perceived as disapproving, socially powerful, or likely to gossip can increase the expected interpersonal cost of disclosure. An automated interviewer may reduce this cost when it is understood as free from ordinary interpersonal judgment. Lucas and colleagues compared conditions surrounding a virtual interviewer and reported greater willingness to disclose under a computer-framed encounter than under conditions eliciting concern about human evaluation. The result is best understood within the studied interviewing task and participant population. A convincing conversational character, the participant's belief about observation, and the actual presence of a human evaluator are distinct manipulable features Lucas et al. (2014) Joinson (2001).

Visual anonymity and self-focused attention provide additional mechanisms. Joinson's computer-mediated communication experiments linked disclosure to conditions affecting self-awareness and visibility; those results concern interaction settings in which familiar interpersonal cues were altered. A contemporary text chatbot combines low visual exposure with a recordable written exchange, so the psychological reduction of embarrassment may coexist with persistence of detailed transcripts. Users may readily infer that the absence of a visible face means the absence of a socially consequential audience. That inference deserves direct measurement rather than acceptance by researchers as a property of the technology. Manipulations that separately vary apparent anonymity, stated logging, and later human review can reveal which component drives sharing Joinson (2001) Acquisti et al. (2015).

The phrase nonjudgmental AI is therefore a hypothesis about how the participant construes the audience, not a guarantee of neutral operation. A conversational system can be programmed to sound validating, can classify content for moderation, and may be embedded in a service with institutional responsibilities or access policies. The psychological mechanism may operate even when the user's belief is incomplete or mistaken. Researchers can assess anticipated embarrassment, fear of negative judgment, trust in data handling, and objective comprehension as separate mediators. Such design clarifies whether a disclosure effect comes from perceived social safety, practical assistance, convenience, or experimentally altered beliefs about technical confidentiality Lucas et al. (2014) Yener et al. (2025) Zhang et al. (2026).

Reciprocity, Conversational Self-Disclosure, and Social Scripts

Human exchange often involves an expectation that personal revelation can invite a corresponding response. Social penetration theory treated mutual exchange as part of relational development, while later research linked disclosure and liking across interpersonal settings. In technological encounters, the machine's generated autobiographical-sounding statements can perform a conversational function similar to a disclosure cue, even though the underlying source and status of such statements differ from a person's lived report. Moon's early experiments showed that computer-mediated reciprocal prompts can elicit private consumer responses, establishing a foundation for studying the persuasive effects of disclosure sequences in artificial dialogue Altman & Taylor (1973) Moon (2000) Collins & Miller (1994).

Liang and colleagues experimentally investigated chatbot self-disclosure strategies in a recommendation setting. Their work provides evidence that particular forms of chatbot disclosure can modify users' perceived conversational resonance and responses, including willingness to share information. The causal claim is about the tested textual strategy and context. Generalizing the result to crisis counseling, romantic companionship, or disclosure of identifying information requires additional experiments. Emotional and factual chatbot self-statements should be separately coded, because they may evoke different social interpretations. It is also necessary to distinguish a scripted account about the system from a false implication that an AI partner has autobiographical experiences Liang et al. (2024) Nass & Moon (2000).

Yoon and Kang's study of stress-management chatbots further illustrates that assigned relationship roles and apparent agent disclosure may have different effects on user behavior and subjective impressions. Relationship framings such as peer and mentor are psychologically informative because users bring distinct norms of intimacy, expertise, and discretion to each role. An optimal study varies the role cue independently of disclosure wording and response warmth. It also records which particular facts users reveal, rather than inferring a broad privacy outcome from a general trust score. Contrasting evidence across experimental tasks should be interpreted as context dependence rather than forced into a single direction of effect Yoon & Kang (2026) Liang et al. (2024) Zhang & Zhu (2026).

Anthropomorphism, Social Presence, and Partner Categorization

Anthropomorphism occurs when a person attributes humanlike capacities, intentions, or qualities to a nonhuman target. In conversational AI, salient cues include a personal name, humanlike voice, first-person pronouns, apparent memory, emotional language, and context-sensitive turn taking. Such cues may increase perceived social presence, but social presence remains a judgment made by the user. A linguistic model can generate coherent descriptions of concern or confidentiality without those statements providing independent evidence about internal experience or platform-level confidentiality. The established tradition of social responses to computers demonstrates that humanlike behavior and anthropomorphic attribution can emerge without an explicit, stable belief in machine personhood Nass & Moon (2000) Bickmore & Picard (2005).

The disclosure consequences of anthropomorphic cues are not mechanically predictable. A warm interface may reduce perceived social distance and increase sharing; a very humanlike presentation may reintroduce evaluation anxiety, raise expectations of judgment, or provoke suspicion about manipulation. Zhang and Zhu's research directly measured actual disclosure under conditions related to anthropomorphic chatbots and contextual privacy calculus, enabling a stronger behavioral discussion than studies measuring attitudes alone. The experiment's interpretation nevertheless depends on its population, task, stimuli, and available controls. Design comparisons should report both the direction of the manipulated cue's effect and how privacy concern, social distance, and trust were measured Zhang & Zhu (2026) Acquisti et al. (2015).

Human versus AI labeling offers a useful experimental contrast, but it alters more than the imagined partner's ontology. The label may also communicate beliefs about data retention, professional accountability, future use, and whether a human could later see the response. Participants may treat a system labeled AI as less socially critical while treating a system labeled corporate service as more invasive. A factorial design can disentangle these beliefs through explicit manipulation checks and by independently varying partner label and disclosed data policy. The broader lesson is methodological: the causal effect of partner categorization should not absorb the effects of the institutional information environment Lucas et al. (2014) Chiu & Foote (2026) Dey & Kumar (2026).

Perceived Responsiveness, Validation, and the Growth of Intimacy

Interpersonal-process accounts locate intimacy in a sequence in which meaningful self-disclosure receives a response that is interpreted as understanding, caring, or validation. Laurenceau and colleagues operationalized this process in close relationships through daily interactions. In an AI encounter, a fluent response can convey recognition of emotional tone, reflect the user's wording, or offer a plausible supportive statement. Researchers should distinguish that perceived responsiveness from accuracy in representing the original disclosure and from the user's later behavior. A high agreement rating measures a positive impression; it does not independently demonstrate appropriate comprehension, reliable memory, or a confidential relationship Laurenceau et al. (1998) Bickmore & Picard (2005).

A psychologically useful model of escalation is sequential. An initial low-risk disclosure elicits a response perceived as responsive; the person then updates expectations about the system and may reveal information of greater depth on a later turn. This process is compatible with human disclosure theories, yet evidence is needed to determine whether it operates over weeks of AI interaction and whether increased sharing improves well-being. Longitudinal studies can code each conversational segment for topic sensitivity, response characteristics, subsequent disclosure, and changes in perceived safety. Randomized response styles would help distinguish the effects of validation from preexisting user disposition and prior loneliness Collins & Miller (1994) Liang et al. (2024) Laurenceau et al. (1998).

Relational validation creates a special privacy problem when supportive language is interpreted as a promise of confidentiality. A phrase that sounds intimate may carry no information about access rights, deletion, or external reporting. The relevant intervention can be straightforward: describe storage and retention clearly while preserving a supportive tone, and measure whether the person understands that distinction. A successful design would enable valued expression while supporting informed boundaries. Studies of privacy assurances and professional-boundary warnings provide an experimental starting point, although vignette-based trust outcomes should not be read as real-world disclosure or professional help-seeking behavior Zhang et al. (2026) Chiu & Foote (2026).

User Goals, Motivation, and Context-Specific Disclosure

The same user may disclose for fundamentally different reasons across contexts. In a productivity task, identifying information may be entered to personalize a draft or analyze a document. In a companionship exchange, disclosure may seek recognition or emotional relief. In a health-oriented interaction, a person may supply symptoms, medication history, or distress to obtain guidance. These motives imply different expectations about relevance, benefit, and acceptable processing. A privacy calculus model describes disclosure as shaped by perceived benefits and risks, but psychological research demonstrates that these perceptions are context-dependent and can be influenced by interface design, momentary emotion, and information asymmetry Acquisti et al. (2015) Meng & Liu (2025) Dey & Kumar (2026).

An empirical program should independently assess instrumental utility, affiliation motivation, emotional distress, curiosity, perceived anonymity, and the cost of obtaining comparable help elsewhere. Experimental tasks can equalize utility while varying the perceived partner or requested detail, whereas longitudinal studies can examine how changing motives predict later disclosure. Asking only whether users trust AI conflates these routes. Merwin and colleagues' comparison of personal information provision to AI and human audiences indicates that disclosure similarities can occur under specific experimental procedures, challenging universal claims that people consistently reveal either more or less to systems. The important next question concerns which dimensions of information and which audience expectations predict differences Merwin et al. (2025) Zhang & Zhu (2026).

Context determines the possible consequences of a message. Disclosure of a favorite film, a financial identifier, a private conversation, and suicidal ideation may all be expressed through the same text box, but their normative and practical stakes differ greatly. A valid study should classify sensitivity by information type, identifiability, potential harm, whether third parties are implicated, and the person's own appraisal. Severity cannot be assigned solely by word count or emotional tone. Behavioral outcomes should also include beneficial understanding of choices, user control, regret, and later information exposure, so that the science of self-disclosure does not mistake maximal revelation for a desirable goal Nissenbaum (2004) McBain et al. (2026) Petronio (2002).

Privacy as Control, Boundary Coordination, and Appropriate Information Flow

Privacy research provides several complementary ways to specify what is at stake when someone speaks to an artificial system. A control-based account examines a person's capacity to determine who obtains information and under what conditions. Communication privacy management considers the rules by which people create, reveal, and negotiate boundaries around private knowledge. Contextual integrity evaluates whether information moves according to norms specifying the sender, subject, recipient, information category, and transmission conditions. These approaches become especially important in human–AI interaction because the conversational recipient as perceived by the user and the institutional entities technically handling the transcript may differ Westin (1967) Petronio (2002) Nissenbaum (2004).

Imagine an individual describing a disagreement with a partner to a chatbot. The user may assume the exchange resembles a private conversation, the partner may regard shared messages as confidential, and the service may possess mechanisms for logging, abuse review, retention, personalization, or permitted onward transfer. Each expectation applies to a different relationship in the information flow. Assessing privacy requires asking whether the person had authority to share the material, whether the system communicated its actual practices, and whether processing remains compatible with the original social purpose. The act of typing is a behavioral fact; authorization for every subsequent use is a separate normative and operational question Petronio (2002) Nissenbaum (2004) Chiu & Foote (2026).

Contextual analysis can be operationalized rather than treated as an abstract label. A study can present a constant disclosure task while manipulating the recipient of stored information, the retention duration, whether data are used for system improvement, or whether a professional reviews a flagged message. Participants can then rate appropriateness, willingness, and perceived boundary violations, followed by an optional real disclosure task. The distinction between stated preference and behavior matters because convenience, urgency, and interface framing can change action even when privacy attitudes remain stable. This method connects the ethical architecture of information flow with experimentally testable psychological processes Nissenbaum (2004) Acquisti et al. (2015) Zhang & Zhu (2026).

Privacy Calculus and Its Psychological Limitations

Privacy calculus approaches represent disclosure decisions as influenced by expected advantages and perceived exposure costs. In conversational AI, benefits may include advice, companionship, personalization, speed, or lower interpersonal embarrassment; costs may include unwanted retention, profiling, interception, interpersonal discovery, or misuse. Meng and Liu's study applies privacy-calculus constructs specifically to AI chatbots, while newer research examines controls, nudges, and anthropomorphic design in large language model interfaces. These contributions are useful because they identify measurable beliefs that could be independently manipulated. They do not establish that each interaction is a fully informed and stable calculation Meng & Liu (2025) Dey & Kumar (2026).

Behavioral research on digital privacy has repeatedly shown that preferences, circumstances, and available choices matter. Acquisti and colleagues review how limited attention, uncertainty, framing, and immediate incentives affect decisions in information environments. A person disclosing in distress may weight relief more heavily than future risk, while a user completing a routine task may respond strongly to a salient deletion control. In both cases, the system's default settings may carry greater behavioral influence than a distant or complex policy text. Models should therefore measure what participants noticed, understood, and expected at the moment of disclosure rather than relying exclusively on a general attitude item administered afterward Acquisti et al. (2015) Taddicken (2014).

The privacy paradox is often used to describe apparent discrepancies between concern and revealing behavior. Taddicken's social-web research gives evidence that such discrepancies depend on the kind of disclosure, perceived social relevance, and individual characteristics. In AI settings, calling a difference paradoxical can obscure a rational response to benefits, incomplete information, limited alternatives, or inappropriate survey measurement. A stronger design records sensitivity-specific concern, actual optional disclosure, perceived alternatives, and factual comprehension of information use. Privacy calculus then becomes a falsifiable family of explanations rather than a formula that automatically rationalizes whatever the participant eventually chooses Taddicken (2014) Acquisti et al. (2015) Zhang & Zhu (2026).

Corporate Infrastructure and the Illusion of a Private Dyad

A conversational interface is experienced one turn at a time, often as if a single partner were listening. Yet its production normally depends on a service architecture involving authentication, network transport, inference systems, operational logs, safety processes, retention policies, and potentially integrated third-party services. This layered organization matters psychologically because the visible dialogue can conceal the number and identity of potential institutional recipients. A claim that an AI companion feels confidential describes a user's belief, while a claim that information remains inaccessible to others requires a distinct technical and policy basis. The two dimensions must be examined together rather than translated into one another Nissenbaum (2004) Chiu & Foote (2026).

Chiu and Foote's qualitative study of AI companions foregrounds the tension between intimate conversation and corporate control. Participants discussed privacy management in relationships with systems that can feel personally familiar while remaining operated by companies. Their interviews help formulate questions about imagined co-ownership, memory management, deletion, and the practical limits of interpersonal boundary negotiation. Because the study used a qualitative sample, it establishes situated experiences and themes rather than population-wide prevalence. Its contribution is a clearer description of the relational and institutional levels that participants must manage simultaneously Chiu & Foote (2026) Petronio (2002).

An empirically useful disclosure map should follow the information beyond the point of entry. For each disclosure event, researchers can record its technical route, claimed processing purpose, retention conditions, user-accessible controls, third-party inclusion, and opportunities for subsequent inference. This can be joined to interviews assessing user expectations. Comparison of the two reveals expectation–implementation gaps without presuming a breach of law or contract. The approach also exposes discrepancies between privacy labels and practical affordances, a concern in audits of healthcare chatbot applications where sensitive disclosures make transparency and control especially consequential Yener et al. (2025) Acquisti et al. (2015).

Memory, Personalization, and Temporal Expansion of Disclosure

Persistent conversational memory changes the privacy implications of disclosure even when the information shared in a single message appears ordinary. Across repeated interactions, the accumulation of preferences, locations, relationships, fears, schedules, and beliefs can make a user more identifiable and permit sensitive inferences. The psychological effect of memory may be positive when the person experiences continuity and saves the effort of repeating context. The same continuity may magnify the consequences of an incorrect belief about retention or deletion. Relational-agent research established the importance of repeated interaction, while contemporary AI-companion interviews document how users interpret persistent histories as part of an evolving personal relationship Bickmore & Picard (2005) Chiu & Foote (2026).

Memory must be specified operationally. An interface might display earlier messages, store user-written notes, retrieve selected fragments, maintain a profile, or regenerate a summary. These mechanisms vary in fidelity, visibility, control, and discoverability. Psychological studies that simply label one condition memory on and another memory off may fail to capture which form of continuity affects disclosure. Experimental comparisons should distinguish accurate recall, apparent recall, explicit permission for storage, controllable forgetting, and unintended resurfacing of details. A system that remembers correctly can still violate an expectation about the purpose for which information was entrusted Nissenbaum (2004) Petronio (2002).

Research should also examine temporal risk asymmetry. Immediate conversational benefits are visible within minutes, whereas later consequences depend on future circumstances: changed relationships, a service policy update, a security incident, a new integration, or a disclosure about someone else. This asymmetry helps explain why individuals may share valuable contextual information despite persistent uncertainty about long-term exposure. The appropriate measure is neither a global label of fearlessness nor a judgment that all personalization is harmful. It is the correspondence among the person's intended duration of use, the actual retention regime, and meaningful opportunities to inspect, correct, or remove information Acquisti et al. (2015) Chiu & Foote (2026).

Experimental Disclosure Research: Historical and Contemporary Designs

The empirical literature offers a sequence of increasingly demanding tests. Early controlled studies established that the presentation and ordering of computer prompts can change users' personal revelations. Moon's consumer studies examined socially structured computer interactions, while Lucas and colleagues later manipulated beliefs about a virtual interviewer's monitoring context. Contemporary chatbot studies manipulate apparent reciprocity, social role, or anthropomorphic cues. The continuity across these studies is important: a randomized interface feature can support a causal inference about disclosure within a specific situation. The studies nevertheless involve different technological capacities, audiences, topics, and measures, so their findings should not be aggregated as though they estimate one common effect Moon (2000) Lucas et al. (2014) Liang et al. (2024).

Zhang and Zhu's 2026 online experiment is particularly valuable because it observed actual behavior rather than relying exclusively on an intention scale. The authors enrolled 222 participants and examined anthropomorphism, perceived social distance, privacy concern, trust, and information sensitivity. Their reported findings indicate that anthropomorphic cues can decrease social distance while producing potentially adverse reactions, and that privacy concerns and trust do not map straightforwardly onto actual sharing. These results challenge the assumption that raising trust scores inevitably increases disclosure. As with other online experiments, the practical stakes and available alternatives in the task may differ from those in durable, emotionally significant companionship Zhang & Zhu (2026) Acquisti et al. (2015).

Experimental validity improves when a study has a preregistered primary endpoint, a clear assignment procedure, adequate statistical precision, transparent exclusion criteria, and a meaningful control condition. A neutral-response control should be matched for length and task usefulness when testing supportive style; an AI-label intervention should hold actual data handling constant; a privacy-control test should offer a control that actually affects retention or access where feasible. Manipulation checks can reveal whether participants interpreted the intervention as researchers intended. Where only disclosure intentions are measured, conclusions should remain explicitly about those intentions rather than claiming that the intervention altered the amount of sensitive information actually provided Zhang & Zhu (2026) Dey & Kumar (2026).

Measurement of Actual Disclosure, Sensitivity, and Exposure

The most consequential measurement distinction is between intended and observed self-disclosure. A participant's statement that they would tell an AI about a relationship difficulty can be shaped by imagined circumstances and social desirability; a voluntary text entry provides a behavioral observation within the experimental setting. Actual disclosure can be coded for breadth, intimacy, emotional content, personal identifiability, third-party involvement, and disclosure of protected or potentially consequential information. Researchers should publish a codebook and interrater reliability estimates for human-coded categories, because nominally objective measures still involve interpretive judgments about whether a phrase constitutes personally sensitive information Altman & Taylor (1973) Merwin et al. (2025) Zhang & Zhu (2026).

Disclosure depth is a relational construct and should be distinguished from information security risk. A participant may disclose a highly emotional but anonymous thought that has little identifying content, while a seemingly mundane address or work schedule may create a high degree of practical exposure. A rigorous instrument therefore needs at least two axes: perceived personal intimacy and potential informational consequence. The person's appraisal can be measured alongside independent coding of the supplied text. Sensitivity may also depend on jurisdiction, social context, and the presence of third-party data. Recording only a total word count or the number of turns loses these distinctions Nissenbaum (2004) Taddicken (2014).

Ethical data collection requires the research itself to minimize the risks it measures. Studies can use optional disclosures, explicitly fictional alternatives, participant-controlled redaction, secure storage, access limits, and planned deletion while still testing the effect of interface manipulations. When truly sensitive data are unnecessary for a scientific hypothesis, surrogate tasks can reveal mechanisms without creating avoidable exposure. Research protocols should distinguish public availability of aggregate results from access to raw conversational records. The scientific record should report which forms of data were collected and retained, and how these decisions may have influenced the observed willingness to disclose Petronio (2002) Acquisti et al. (2015) Yener et al. (2025).

Surveys, Qualitative Interviews, and Ecological Observation

Surveys estimate patterns of self-reported use, intentions, or perceptions within a defined sampling frame. They can identify demographic variation and associations among concerns, attachment, and frequency of use, but cannot automatically establish causal effects of AI interaction. McBain and colleagues surveyed a sample of 1,009 U.S. individuals aged 12 to 21 in late 2025 and analyzed population-weighted reports of mental-health chatbot use and disclosure of that use to others. The study reports that 19.2 percent had used an AI chatbot for mental health advice and that 63.3 percent of these users had told no one about their use. These are estimates for the specified sample and survey period, not lifetime clinical outcomes or evidence that the systems delivered safe treatment McBain et al. (2026).

Qualitative interviews reveal the meanings and informal practices that standardized survey instruments may overlook. Chiu and Foote's 15 interviews with AI-companion users describe how apparently personal exchanges coexist with awareness of corporate infrastructure. Reported privacy tactics include selective withholding, modifying identifying information, and weighing the possibility of losing remembered history against the risk of further exposure. Their analysis supplies concepts for future survey measures and experimental manipulations. Its explanatory strength derives from the detail of situated accounts, while its sample size and recruitment context limit claims about the distribution of the themes across all users of conversational AI Chiu & Foote (2026).

Ecological observation can connect stated beliefs with real interaction trajectories. With meaningful consent and minimization, researchers might study anonymized turn-level patterns, user-requested deletion events, changes following transparent privacy notices, or the effect of feature updates on voluntary sharing. Such data offer greater behavioral realism but introduce selection bias, imperfect control of confounders, and possible changes in the system over time. A strong triangulation program combines controlled experiments for mechanism, representative surveys for estimated prevalence, interviews for interpretive meaning, and platform audits for actual protections. Findings should be reported at the level warranted by each design Zhang & Zhu (2026) McBain et al. (2026) Chiu & Foote (2026).

Platform Audits and Verification of Privacy Affordances

Privacy assurances become scientifically tractable when the researcher specifies the observable property being assessed. For a conversational service, an audit might inspect data retention settings, visibility of consent choices, user download and deletion options, policy specificity, access disclosures, and safeguards around especially sensitive content. Yener and colleagues investigated privacy features in healthcare chatbot applications, illustrating the value of comparing user-facing claims with concrete platform characteristics. An interface can permit emotional disclosure even when the user has little practical control over the future of a transcript. The audit evaluates the service environment, whereas a psychological experiment evaluates how people respond to that environment Yener et al. (2025) Nissenbaum (2004).

Audit evidence must be time-stamped and bounded. Applications change their settings, technical architecture, vendors, and policies; a feature observed in one version cannot be assumed present across services or dates. Similarly, a visible delete button may remove a user-facing conversation while other records remain governed by distinct retention obligations. Auditors should state exactly which documentation, interface actions, and technically observable outcomes they tested. They should avoid inferring unseen internal processing solely from a label on the screen. Comparing audit results with participant comprehension can then identify where protective affordances are meaningful, misunderstood, or difficult to use Yener et al. (2025) Acquisti et al. (2015).

Institutional trust also differs from trust in conversational behavior. A response may appear supportive while its service lacks adequate communication about storage. Conversely, a well-protected platform may sometimes generate a response the user finds impersonal. Separate measurements allow researchers to identify which form of confidence predicts sharing: response relevance, institutional credibility, perceived human oversight, security assurance, or the ability to reverse a disclosure. Research on privacy assurances and boundary warnings shows how framing can change reported expectations, but an independent audit is required before treating an assurance as evidence of technical compliance Zhang et al. (2026) Chiu & Foote (2026) Yener et al. (2025).

Statistical Inference, Replication, and Evidence Appraisal

Disclosure studies frequently measure several correlated outcomes: warmth, trust, intimacy, willingness to share, actual words entered, perceived confidentiality, and perceived usefulness. Testing each as though it were a separate confirmation risks overstating the evidence, particularly in modest experimental samples. Investigators should define primary outcomes, justify sample size, report uncertainty around estimates, and distinguish planned from exploratory analyses. Mediation analyses require especially careful interpretation, because post-treatment associations between perceived social distance and disclosure do not by themselves establish that the proposed mediator caused the behavior. Randomized mediation components or stronger longitudinal designs would improve causal identification Zhang & Zhu (2026) Dey & Kumar (2026).

Replication should preserve both the hypothesized psychological mechanism and the information context that makes the intervention meaningful. A finding based on a scripted health interviewer may fail to transport to a general-purpose chatbot because the audience, task, and actual logging conditions differ. Similar wording can also acquire different meaning across cultures or generations. Researchers should report model version, prompting procedure, role framing, participant recruitment, compensation, retention conditions, and sensitivity of the invited disclosure. When available, independent teams should reproduce the manipulation with alternative interfaces and participant samples before making broad claims about human behavior with AI systems Lucas et al. (2014) Yoon & Kang (2026) Merwin et al. (2025).

Current evidence supports a cautious explanatory synthesis: disclosure is responsive to audience beliefs, conversation structure, social framing, task benefits, and privacy-related concerns, while the direction and magnitude depend on contextual details. More ambitious conclusions about long-term clinical outcomes, durable intimacy, or risk reduction require suitable outcome measures and time horizons. A systematic review with registered selection rules and quantitative synthesis might eventually estimate effects for sufficiently comparable interventions, but the present monograph is a source-identified critical narrative synthesis. Its value lies in preserving distinctions among study designs, identifying convergent mechanisms, and specifying experiments that could disconfirm attractive but underspecified explanations Acquisti et al. (2015) Collins & Miller (1994) Chiu & Foote (2026).

Mental-Health Contexts, Vulnerability, and Professional Boundaries

Mental-health chatbot use makes the tension between perceived relational safety and institutional responsibility especially pronounced. People may disclose depression, anxiety, trauma, interpersonal harm, or urgent distress because a conversational interface is immediately available and appears receptive. Such disclosure can provide information about help-seeking behavior, but the appearance of emotional support must not be treated as evidence that an application provides validated diagnosis or effective treatment. A 2026 U.S. survey found substantial reported chatbot use for mental-health advice among adolescents and young adults; its self-reported helpfulness outcomes describe user appraisal and cannot establish safety or clinical efficacy McBain et al. (2026) Lucas et al. (2014).

Mental-health studies require precise distinctions among symptom description, perceived comfort, consultation with professionals, disclosure of chatbot use to others, and actual subsequent care. A randomized vignette experiment on privacy assurances and professional-boundary warnings addresses how different interface messages influence reported trust, overreliance concerns, and help-seeking intentions. Such work identifies plausible communication effects but does not demonstrate that participants would follow the same course of action in an emergency or during sustained real-world use. A well-designed clinical-context study should prioritize safety oversight, informed consent, appropriately qualified review, and meaningful outcome measurement while preserving the ability to investigate ordinary psychological mechanisms Zhang et al. (2026) McBain et al. (2026).

Research participants and users may understand reassuring language as a sign of professional competence, confidentiality, or continuous supervision. Each implication should be separately examined. Emotional validation is a conversational characteristic, professional licensure is an institutional fact, and confidentiality depends on concrete legal, policy, and technical conditions. Interface studies can test whether explicit role descriptions and accurate privacy disclosures improve comprehension without sacrificing approachability. Where research focuses on vulnerable populations, the scientific objective should include calibrated expectations and opportunities for human support, with particular attention to adolescents and to individuals sharing third-party or highly sensitive information Yener et al. (2025) Zhang et al. (2026) Zhang et al. (2026).

Age, Culture, Accessibility, and Unequal Privacy Exposure

Disclosure research cannot assume that a convenience sample of adult online participants represents the experiences of children, older adults, or culturally diverse users. Expectations about privacy, shame, family obligations, professional authority, and the appropriate audience for emotional concerns vary across social settings. Language proficiency, disability accommodations, interface literacy, and access to alternative services may also influence whether conversational AI becomes a preferred venue for personal revelation. Sampling and measurement should therefore be designed around relevant populations rather than treating demographic variation as a secondary adjustment applied after data collection Taddicken (2014) McBain et al. (2026).

Adolescents present a particularly important case because developing autonomy, peer relationships, family expectations, and differing capacities to interpret platform governance can all influence disclosure. McBain and colleagues' U.S. study gives a time-bounded description of reported mental-health chatbot use among people aged 12–21 and of whether users revealed that use to others. It does not measure every kind of personal information shared within the chatbot. Future studies should separate the disclosure of private content to the system from disclosure of system use to parents, peers, or clinicians. Conflating these directions would produce misleading claims about secrecy or help-seeking McBain et al. (2026).

Cross-cultural replication requires more than translating a questionnaire. A sensitive question about a family conflict may have a different social meaning where collective and individual privacy norms diverge; a privacy notice may also have different implications under different institutional arrangements. Researchers can combine cognitive interviews, measurement invariance testing, multilingual tasks, and contextual descriptions of platform features. Studies should assess whether categories such as perceived nonjudgment, trust, and informational control retain the same meaning across groups. Equity also requires examining users whose limited resources or access to professionals make the apparent choice between artificial and human support less voluntary Nissenbaum (2004) Acquisti et al. (2015) Walther (1996).

Longitudinal Relationships, Habituation, and Disclosure Trajectories

A single laboratory encounter shows how a disclosure event responds to an intervention, whereas ongoing relationships introduce accumulation, expectation formation, and changing perceived stakes. Relational-agent research treated sustained interaction as an empirical object before current generative AI. With contemporary companions, a user may repeatedly encounter patient responses, recall of earlier topics, and a stable conversational style. These experiences can make sharing feel routine and may gradually shift which details are considered safe to reveal. The mechanism needs longitudinal evidence because a cross-sectional association between frequent use and intimate disclosure cannot determine whether increased use preceded the revelation or vice versa Bickmore & Picard (2005) Chiu & Foote (2026).

A repeated-measures program can follow voluntarily recruited participants across multiple sessions while coding new personal disclosures, first-time sensitive categories, reactions to system mistakes, perceived privacy, and the presence of active privacy decisions. Event-history analyses can examine when a participant crosses a new disclosure boundary, and interrupted time-series designs can exploit explicitly documented feature changes. Such methods should control for time-varying distress, external support, changing user goals, and system updates. Where transcript access is involved, minimization and participant control must remain part of the protocol; building an intimate research archive to measure intimacy would otherwise create a serious conflict between purpose and method Petronio (2002) Acquisti et al. (2015).

Longitudinal theory must also accommodate declining disclosure. People may become more selective after noticing errors, learning about retention, experiencing unwanted personalization, or reevaluating the emotional significance of an exchange. Conversely, the threat of losing memories may make some users reluctant to exercise deletion controls. Chiu and Foote's interviews report precisely such tensions between relational continuity and data control. Future work should measure both increases and decreases in sharing, explicit attempts to correct past disclosure, and changing conceptions of the recipient. These are distinct outcomes of relationship development, and an adequate account does not assume a unidirectional progression toward maximal openness Chiu & Foote (2026) Altman & Taylor (1973).

Practical Design Implications and Testable Safeguards

The most defensible interface interventions map directly onto empirically identified sources of misunderstanding. A service can communicate that a message may be processed and retained under specified conditions, differentiate emotional support from professional services, provide accessible data controls, and prompt the user to consider third-party identities before submitting copied correspondence. Each intervention has a measurable objective: improved comprehension, reduced unintended sensitive exposure, more appropriate trust, or greater user control. A change in self-reported satisfaction alone is insufficient evidence that privacy improved. Evaluation should compare understanding and behavior against the service's verified technical properties Yener et al. (2025) Zhang et al. (2026) Nissenbaum (2004).

Designs that encourage reflection may preserve legitimate self-disclosure while avoiding unnecessary collection. For example, an interface might support summarizing a private dispute without names, choosing a temporary chat mode where such a mode is actually implemented, or reviewing a draft before transmission. The experimental question is whether these affordances alter the content and sensitivity of the submitted information without imposing excessive cognitive load or discouraging needed assistance. The literature on behavioral privacy decisions suggests that timing, framing, and defaults matter. A successful safeguard should therefore be assessed in the moment of choice as well as through later recollection of a policy Acquisti et al. (2015) Dey & Kumar (2026).

Research and product evaluation should specify whose interests a safeguard protects. A user may want a continuous memory, a third party may prefer limited circulation, and a platform may have obligations to preserve certain records. Privacy design must accommodate such conflicts through clear scope and user understanding. Communication privacy management and contextual integrity offer conceptual resources for mapping legitimate expectations, while audits reveal whether actual settings and procedures correspond to promises. A high-quality interface supports psychologically useful expression with calibrated expectations and observable protections, and its claims remain revisable as technical conditions change Petronio (2002) Nissenbaum (2004) Chiu & Foote (2026).

Unresolved Scientific Questions and a Research Agenda

A major unresolved question concerns the long-term causal relation between perceived responsiveness and increasingly sensitive disclosure. Existing experimental evidence identifies effects of particular cues, whereas naturalistic data reflect selection and changing personal needs. Multi-site longitudinal experiments could examine whether warm response style, transparent storage information, and meaningful deletion controls independently change the probability of a new sensitive revelation. The analysis should measure mediators without treating them as proven causes, and predefine outcomes for accidental disclosure and regret. Independent replication across model versions is necessary because the observable conduct of a conversational system depends on both the underlying model and its deployment context Liang et al. (2024) Zhang & Zhu (2026) Chiu & Foote (2026).

A second question concerns the gap between individual privacy judgments and the institutional reality of AI companionship. Researchers need methods for connecting participants' beliefs about data flow to contemporary technical audits, policy examination, and observable controls. Long-term research should test whether education, interface design, or actual control mechanisms narrow this gap. A third question concerns who bears risks when an exchange contains information about people absent from the interaction. Here, studies must include third-party viewpoints, contextual norms, and differences in the consequence of inadvertent disclosure. The literature on contextual integrity and relational privacy provides hypotheses, but their predictive scope in AI companionship remains open Nissenbaum (2004) Petronio (2002) Yener et al. (2025).

A fourth priority is rigorous research in clinical and developmental contexts with ethically justified designs and well-defined limits on generalization. Surveys provide important evidence that young people use AI systems for mental-health advice, but intervention trials and careful observational work are required to understand benefit, dependence, and response to safety information. Researchers should document when data were collected, how participants were recruited, what systems were tested, and whether outcomes represent behavior, intention, subjective appraisal, or a verified technical property. The scientific field advances by linking distinct methods to clearly identified questions rather than treating conversational fluency or reported trust as comprehensive evidence about privacy McBain et al. (2026) Zhang et al. (2026) Merwin et al. (2025).

Conclusion

Self-disclosure in human–AI relationships is a psychological process embedded in an evolving information infrastructure. Its central mechanisms include anticipated evaluation, anonymity, reciprocal conversational cues, anthropomorphic framing, perceived responsiveness, instrumental benefit, and expectations about privacy. The traditions of social penetration and interpersonal intimacy explain why revelation can acquire relational significance, while research on social responses to computers and virtual interviewers shows why technologically mediated audiences may alter willingness to speak. Contemporary experiments extend these foundations by measuring chatbot strategies, trust, social distance, and actual personal sharing. These convergent findings establish disclosure as a context-sensitive phenomenon whose meaning depends on the person, task, partner representation, and consequences of the information flow Altman & Taylor (1973) Laurenceau et al. (1998) Lucas et al. (2014) Zhang & Zhu (2026).

The principal analytical requirement is to separate what people experience from what a platform actually does. Feeling understood, perceiving confidentiality, disclosing willingly, and possessing enforceable control over information are distinct states. Privacy can be studied as boundary coordination and contextual appropriateness, with institutional operations assessed through clearly delimited audits. Qualitative accounts of AI companionship make visible the negotiation between relational familiarity and corporate processing; survey work locates particular practices in sampled populations; experiments test mechanisms under controlled conditions. Each contributes evidence at its own level, and the limits of its inference remain part of the conclusion rather than an afterthought Chiu & Foote (2026) McBain et al. (2026) Nissenbaum (2004) Yener et al. (2025).

The Artificial Era expands the settings in which people formulate private thoughts through dialogue with artificial systems. A rigorous psychology of these exchanges investigates observable disclosure, human expectation, relational interpretation, and technical governance together. Its practical significance lies in supporting meaningful communication while improving the accuracy of users' expectations and the control available over personal information. The strongest future studies will combine behavioral disclosure measurement, ecologically valid longitudinal observation, transparent platform evaluation, and carefully justified causal designs. The resulting knowledge can inform relational technologies without confusing a convincing conversational performance with institutional confidentiality or treating increasing disclosure as a universal sign of successful connection Acquisti et al. (2015) Petronio (2002) Zhang & Zhu (2026).

References

  • Acquisti, A., Brandimarte, L., & Loewenstein, G. (2015). Privacy and human behavior in the age of information. Science, 347(6221), 509–514. https://doi.org/10.1126/science.aaa1465
  • Altman, I., & Taylor, D. A. (1973). Social penetration: The development of interpersonal relationships. Holt, Rinehart and Winston. https://www.ncbi.nlm.nih.gov/nlmcatalog/436535
  • Bickmore, T. W., & Picard, R. W. (2005). Establishing and maintaining long-term human-computer relationships. ACM Transactions on Computer-Human Interaction, 12(2), 293–327. https://doi.org/10.1145/1067860.1067867
  • Chiu, H.-C., & Foote, J. (2026). Chatting with confidants or corporations? Privacy management with AI companions. Journal of Computer-Mediated Communication, 31(4), zmag014. https://doi.org/10.1093/jcmc/zmag014
  • Collins, N. L., & Miller, L. C. (1994). Self-disclosure and liking: A meta-analytic review. Psychological Bulletin, 116(3), 457–475. https://doi.org/10.1037/0033-2909.116.3.457
  • Dey, S., & Kumar, S. (2026). Privacy calculus and disclosure in LLMs: The role of privacy controls, nudges, and perceived anthropomorphism. International Journal of Information Management, 89, Article 103071. https://doi.org/10.1016/j.ijinfomgt.2026.103071
  • Joinson, A. N. (2001). Self-disclosure in computer-mediated communication: The role of self-awareness and visual anonymity. European Journal of Social Psychology, 31(2), 177–192. https://doi.org/10.1002/ejsp.36
  • Laurenceau, J.-P., Barrett, L. F., & Pietromonaco, P. R. (1998). Intimacy as an interpersonal process: The importance of self-disclosure, partner disclosure, and perceived partner responsiveness in interpersonal exchanges. Journal of Personality and Social Psychology, 74(5), 1238–1251. https://doi.org/10.1037/0022-3514.74.5.1238
  • Liang, K.-H., Shi, W., Oh, Y. J., Wang, H.-C., Zhang, J., & Yu, Z. (2024). Dialoging resonance in human-chatbot conversation: How users perceive and reciprocate recommendation chatbot's self-disclosure strategy. Proceedings of the ACM on Human-Computer Interaction, 8(CSCW1), Article 200, 1–28. https://doi.org/10.1145/3653691
  • Lucas, G. M., Gratch, J., King, A., & Morency, L.-P. (2014). It's only a computer: Virtual humans increase willingness to disclose. Computers in Human Behavior, 37, 94–100. https://doi.org/10.1016/j.chb.2014.04.043
  • McBain, R. K., Cantor, J. H., Breslau, J., Diliberti, M., Zhang, L. A., Zhang, F., Burnett, A., Kofner, A., Rader, B., Pataranutaporn, P., Stein, B. D., Mehrotra, A., & Yu, H. (2026). AI chatbot use and disclosure for mental health among US adolescents and young adults. JAMA Pediatrics, 180(8), 884–890. https://doi.org/10.1001/jamapediatrics.2026.2015
  • Meng, X., & Liu, J. (2025). “Talk to me, I’m secure”: Investigating information disclosure to AI chatbots in the context of privacy calculus. Online Information Review, 49(5), 933–954. https://doi.org/10.1108/OIR-06-2024-0375
  • Merwin, E. R., Hagen, A. C., Keebler, J. R., & Forbes, C. (2025). Self-disclosure to AI: People provide personal information to AI and humans equivalently. Computers in Human Behavior: Artificial Humans, 5, 100180. https://doi.org/10.1016/j.chbah.2025.100180
  • Moon, Y. (2000). Intimate exchanges: Using computers to elicit self-disclosure from consumers. Journal of Consumer Research, 26(4), 323–339. https://doi.org/10.1086/209566
  • Nass, C., & Moon, Y. (2000). Machines and mindlessness: Social responses to computers. Journal of Social Issues, 56(1), 81–103. https://doi.org/10.1111/0022-4537.00153
  • Nissenbaum, H. (2004). Privacy as contextual integrity. Washington Law Review, 79(1), 119–158. https://digitalcommons.law.uw.edu/wlr/vol79/iss1/10/
  • Petronio, S. (2002). Boundaries of privacy: Dialectics of disclosure. State University of New York Press. https://search.worldcat.org/title/Boundaries-of-privacy-%3A-dialectics-of-disclosure/oclc/54481653
  • Taddicken, M. (2014). The ‘privacy paradox’ in the social web: The impact of privacy concerns, individual characteristics, and the perceived social relevance on different forms of self-disclosure. Journal of Computer-Mediated Communication, 19(2), 248–273. https://doi.org/10.1111/jcc4.12052
  • Walther, J. B. (1996). Computer-mediated communication: Impersonal, interpersonal, and hyperpersonal interaction. Communication Research, 23(1), 3–43. https://doi.org/10.1177/009365096023001001
  • Westin, A. F. (1967). Privacy and freedom. Atheneum. https://lawlib.state.md.us/Record/.b10010270
  • Yener, R., Chen, G.-H., Gumusel, E., & Bashir, M. (2025). Can I trust this chatbot? Assessing user privacy in AI-healthcare chatbot applications. Proceedings of the Association for Information Science and Technology, 62(1), 809–820. https://doi.org/10.1002/pra2.1299
  • Yoon, S., & Kang, Y. (2026). Are we close enough? Impact of stress management chatbot’s self-disclosure and relationship types on user’s self-disclosure, trust, intimacy, and anonymity. International Journal of Human-Computer Studies, 209, 103738. https://doi.org/10.1016/j.ijhcs.2026.103738
  • Zhang, M., & Zhu, H. (2026). Actual self-disclosure to anthropomorphic AI chatbots: A contextual privacy calculus approach. AIS Transactions on Human-Computer Interaction, 18(1), 31–60. https://doi.org/10.17705/1thci.00237
  • Zhang, F., Chen, Y., Liu, M., Xu, J., & Xu, J. (2026). Disclosure willingness anthropomorphism and thinking style shape perceived fit and comfort in AI mental health chatbot interactions. Scientific Reports. https://doi.org/10.1038/s41598-026-72496-w
  • Zhang, Z., Lu, X., Zhang, Y., Zhang, H., & Zhang, M. (2026). Privacy assurances and professional-boundary warnings in generative AI mental health chatbots: A randomized vignette experiment on calibrated trust, overreliance risk, and professional help-seeking intentions. Frontiers in Psychology, 17, 1934264. https://doi.org/10.3389/fpsyg.2026.1934264