At the individual level, parasocial trust in AI shows up most clearly in how much people disclose and how quickly they accept advice. Research on virtual humans has shown that people often reveal more sensitive information when the interviewer is framed as a computer rather than a human. In one study, researchers found that participants were more willing to talk about mental health symptoms when they believed they were interacting with an automated virtual agent, even though the system was actually controlled by humans behind the scenes.⁵ The “only a computer” framing lowered fear of judgment, which made disclosure feel safer.

Similar findings were reported in a study featuring participants with high social anxiety. In their experiments, individuals who usually struggle to open up in face-to-face settings disclosed more to virtual humans than to human interviewers.⁶ Reduced eye contact, controllable pacing, and the perception that the agent could not gossip about them made the interaction less threatening.

On the other hand, when it comes to seeking support for feeling stressed or worried, human interaction is still seen as more genuine. That said, chatbots have been shown to reduce worry when that emotional support is paired with reciprocal self-disclosure, which builds rapport. This means that the more the relationship mimics that of a “supportive partner,” behavioral recommendations may feel like relational obligations rather than optional suggestions.

When conversational systems adopt human-like design cues, these effects can intensify. Dr. Theo Araujo, a researcher and professor at the University of Amsterdam, showed that anthropomorphic touches, such as a human name, small talk, or naturalistic language, can increase social presence and positive evaluations of both the chatbot and the brand behind it.⁷ Feeling that “someone” is there makes people more engaged, which can translate into longer conversations and richer disclosures. Those disclosures often include usernames, contact details, health histories, financial worries, and social networks. In another experiment, Ho and colleagues found that people who engaged in emotional disclosure with a chatbot experienced similar psychological benefits to those who disclosed to a person, including improved mood and perceived understanding.⁸ The conversation felt meaningful enough that participants treated it as a space for real emotional work.

Perceptions of intelligence and human likeness also increase adoption of personal intelligent agents.⁹ Users who see their assistant as smart and somewhat human are more likely to let it integrate deeply into their routines. In a chat context, that can mean saying “yes” to nudges about products, settings, or behavioral changes that would have felt pushy in a banner ad or static form. Over time, individuals can come to view their AI companion as a trusted confidant that knows their secrets, preferences, and weak spots.¹⁰

At scale, parasocial trust shapes how populations interact with AI-mediated services across domains such as health, education, finance, and customer support. Longitudinal research suggests that people do not experience AI chatbots as one-off tools, but long-term companions. In one long-term study, researchers followed users and documented how some developed ongoing relationships with chatbots, including routines, perceived understanding, and emotional bonds.¹¹ These relationships evolved. Some cooled, others deepened, but they did not remain neutral in their utility. When millions of people build such quasi-relationships with branded AI systems, several systemic risks emerge.

• Privacy externalities. Each piece of disclosed information can feed large training datasets, influence model behavior for future users, and surface in aggregate analytics that guide policy or product decisions.      
• Behavioral convergence. If many users treat AI as a trusted advisor, similar nudges can shape large cohorts at once. In financial apps, this might mean mass uptake of a particular savings product. In health settings, it might tilt large groups toward specific treatments or lifestyle programs.
• Unequal vulnerability. People who are younger, lonelier, or living with stigma may be more likely to seek companionship from AI and allow deep access to their data. Croes and colleagues found that people can be willing to share intimate information with a chatbot, and that doing so can influence emotional well-being.¹² 
• Institutional dependence. Organizations may come to rely on parasocial trust as a design feature. If a health system sees that patients are more likely to accept difficult recommendations from an empathetic chatbot than from a rushed clinician, there may be pressure to route more communication through AI to “improve adherence,” raising some ethical questions.

Many different factors nudge people toward sharing, from anthropomorphic cues and reciprocity norms to expectations of competence and confidentiality.¹³ When these factors are engineered at scale, the result is an ecosystem where users systematically open up and comply more than they realize.

Several cognitive and social mechanisms converge to produce parasocial trust in AI. None of them requires people to believe that the system is literally human. The feeling that “someone is there” can be enough.

Social habits carry over to screens.

Humans are deeply social learners. Many of our interaction rules are automatic. Early media research showed that viewers form one-sided bonds with media figures through repeated exposure, friendly address, and the illusion of intimacy.¹ Later work integrated this insight into the idea that people treat media as if it is part of their social world because the brain uses fast, heuristic processes built for human contact.²

Reduced fear of judgment lowers disclosure barriers.

Lucas and colleagues demonstrated that when people believe they are speaking to a virtual interviewer rather than a human, they often disclose more about sensitive topics such as mental health.⁵ The perception that a computer cannot judge, gossip, or feel annoyed reduces social risk. Kang and Gratch found that socially anxious participants disclosed more to virtual humans than to human interviewers, especially when the virtual agent’s behavior felt responsive.⁶ 

Anthropomorphism and competence feed trust.

Human-like design cues in chatbots, combined with how the agent is framed, influence perceived social presence and evaluations of the company behind the bot.⁷ When the chatbot feels personable, and the framing suggests agency (“your assistant,” “your coach”), people are more likely to see it as a capable social partner. Ho and colleagues found that emotional disclosure with a chatbot produced psychological and relational benefits comparable to disclosure with a human.⁸

Reciprocity, support, and compliance become entangled.

Emotional support chatbots often respond with validation, encouragement, and coping strategies. Meng and Dai examined how self-disclosure by chatbots themselves shapes perceived emotional support and found that reciprocal sharing can make AI partners feel more caring and trustworthy.¹⁰ When the agent “opens up,” users may feel a stronger relational bond, which can blur into a sense of obligation to follow through on its advice. Longer-term studies suggest that repeated supportive exchanges can build enduring relationships with chatbots that feel similar to companionship.¹¹ 

Cognitive load and default heuristics.

Modern chat interfaces sit in already overloaded contexts. People interact with AI when they are tired, stressed, or in the middle of complex tasks. Under these conditions, it’s hard to maintain a reflective stance on each recommendation. Papneja and Yadav highlight how expectations of competence, along with interface cues and prior positive experiences, can push users toward quick, heuristic decisions to trust conversational AI.¹³ 

Parasocial trust in AI has real consequences for autonomy, privacy, and fairness. First, it affects agency. If users treat AI as a friendly expert, they may comply with suggestions that are misaligned with their goals or risk tolerance. Logg and colleagues found that people can actually prefer algorithmic recommendations to human advice, adhering more closely when they believe guidance comes from an algorithm.¹⁴

Second, it changes the risk calculus around disclosure. Many consent flows and data-protection frameworks assume that people treat data decisions as instrumental trade-offs. In reality, those decisions often occur inside a perceived relationship that makes sharing feel like a natural part of conversation. When disclosure is experienced as intimacy rather than a transaction, standard models of informed consent start to look optimistic.

Third, parasocial trust introduces vulnerability gradients. People dealing with loneliness, stigma, or mental health challenges may be more likely to lean on AI for companionship and advice. Studies of intimate disclosure to chatbots show both potential emotional benefits and risks, especially if users rely on systems that are not transparent about limitations or data use.¹²

 Finally, parasocial trust can amplify social and economic power imbalances. Organizations with resources to build human-like AI interfaces can effectively deploy thousands of tireless “front-line relationship workers” that never get bored, never unionize, and never log off. If those workers are optimized for compliance and data capture rather than user agency, entire populations can be nudged in ways that are hard to detect and even harder to contest.

Parasocial trust is not inherently harmful. People can benefit from feeling supported by tools that help them articulate emotions, organize decisions, or practice skills. The risk comes when relational design is used without constraints, transparency, or counterweights that protect users. For practitioners, the goal is to design systems where relational cues and power are in balance. Several strategies can help:

Make the system’s non-human nature visible at the moments when people are most likely to lean in and comply.  Use clear, repeated disclosures in onboarding, first use, and sensitive flows, with plain language such as “I am an AI system that predicts text based on patterns in data.” Pair that transparency with a structural boundary between companionship and high-stakes guidance. Let the chatbot provide warmth for emotional processing, then route decisions about medication, money, or legal steps into a separate “decision support” path with tighter framing, extra prompts, and human review. That separation matters because intimate disclosure to chatbots can shape emotional well-being, so a friendly tone should never be the gateway into consequential choices.¹²

Add friction where compliance is most likely. If parasocial trust tilts users toward saying “yes,” designers can deliberately add well-placed friction where that default would be most costly. Use step-back prompts before major decisions, such as “Here are the alternatives people in your situation often consider.” Papneja and Yadav note that design choices around transparency, control, and user expectations shape how people experience self-disclosure to conversational AI.¹³

Mitigate anthropomorphic overreach. Human-like touches can make interfaces more approachable, but they can also inflate perceived competence and care. Moussawi and colleagues have shown that anthropomorphism boosts adoption of intelligent agents by increasing perceptions of intelligence and affinity.⁹ The aim is to create calibrated relatability that does not overstate what the system can understand or guarantee.

Give users tools to tune the relationship. Not everyone wants the same level of relational engagement. Some users prefer a neutral, transactional style. Others value encouragement and small talk. Systems can offer settings that let people dial the tone up or down, choose when the AI remembers context, and control what kinds of topics it can proactively raise. Skjuve and colleagues’ longitudinal work suggests that relationships with chatbots evolve over time.¹¹ 

In the 1950s, Horton and Wohl observed that viewers of new mass media such as television felt they “knew” their favorite hosts, even though the relationship was entirely one-sided.¹ They described how gestures, conversational tone, and regular broadcast schedules created a sense of intimacy at a distance. Viewers spoke of presenters as if they were part of their social circle. Decades later, Reeves and Nass compiled a wide range of experiments showing that people treat computers and other media as if they were social actors, responding with politeness, reciprocity, and even flattery.² Their work framed these responses as a side effect of slow biological evolution in a world that had become suddenly filled with interactive media. Our social brains do not have a separate protocol for “machines that talk back.”

Giles built on this by proposing a model where parasocial interaction and parasocial relationships arise through repeated exposure, perceived similarity, and role expectations.³ This model helps explain why people can feel genuine grief when a fictional character dies or a streamer stops posting. Hartmann and Goldhoorn turned the focus to specific cues that intensify the parasocial experience, such as direct addressing of the viewer and the illusion of eye contact.⁴ Their work shows that relatively small tweaks in presentation can significantly increase the sense that “this person is talking to me.”

Once conversational AI systems arrived, they inherited all of this history. Instead of passively watching a presenter on a screen, users can now converse with an always-available persona that responds in real time, remembers past exchanges, and adapts to individual preferences. Parasocial dynamics have shifted from one-to-many broadcasts to many individualized interactions.

On the opportunity side, conversational interfaces can make products feel more intuitive and less intimidating. People who would never read a manual will happily ask a chat assistant for help. Those who feel uncomfortable raising questions with human staff might explore sensitive topics with a bot first. Ho and colleagues’ findings suggest that self-disclosure to chatbots can provide genuine psychological benefits, which creates real value for products in health, education, and personal development.⁸ 

Moussawi and colleagues show that anthropomorphism and perceived intelligence drive adoption of intelligent agents.⁹ In product terms, this means that design elements that increase user comfort and engagement may also heighten reliance. Without careful guardrails, teams can end up measuring the success of human-like AI through metrics such as time in conversation, conversion rate, or reported satisfaction, while missing the hidden costs of over-disclosure and over-compliance. Products that rely heavily on parasocial trust should therefore be evaluated using a broader set of outcomes. 

Meng and Dai’s work on emotional support chatbots underscores that design choices about self-disclosure and relational framing also affect how users experience support and obligation.¹⁰ Product decisions in this space are about structuring relationships between humans and institutions, mediated by AI.

Parasocial trust takes on new significance in the era of general-purpose conversational models. These systems can be integrated across many touchpoints, rebranded as different “assistants,” and tuned to express particular personalities. Logg and colleagues’ findings on algorithm appreciation show that people can prefer algorithmic advice to human advice in certain domains, adhering more strongly when guidance comes from an algorithm they perceive as competent.¹⁴ When that algorithm speaks in a warm, conversational voice and appears to know personal details, the combination of deference to computational authority and parasocial trust can be especially potent.

At the governance level, policymakers are starting to grapple with what “trustworthy AI” should mean. The European Commission’s Ethics Guidelines for Trustworthy AI define trustworthiness in terms of legality, ethical alignment, and robustness, and they highlight requirements such as transparency, human agency, and privacy.¹⁵ These principles speak directly to the environments where parasocial trust is most active.

If users experience AI as a caring companion, then legal notices buried in settings menus or dense privacy documents are unlikely to shape behavior. Instead, trustworthiness must show up inside the interaction itself. Papneja and Yadav’s framework emphasizes that self-disclosure to conversational AI emerges from a mix of personal, technological, and contextual factors.¹³ Interventions that focus only on user behavior miss the structural aspects of how systems are deployed and framed. Aligning real-world deployments with high-level ethics guidelines requires attention to both layers.

Recent evidence suggests that this is not a niche edge case. In a 2025 collaboration, OpenAI and the MIT Media Lab found that a small subset of heavy users reported stronger affective attachment and higher indicators associated with problematic use.¹⁶ Mainstream coverage has echoed the same concern, highlighting users who describe feeling emotionally “hooked” on long, intimate conversations with chatbots.¹⁷ For designers, the implication is concrete. The interaction has to carry the guardrails. Repeated “I am an AI” reminders should appear when users disclose highly personal details. High-emotion exchanges should trigger softer language, slower pacing, and clearer prompts toward human support. Teams should also avoid success metrics that reward dependence, like maximizing time in chat, and instead track outcomes tied to user agency and safe off-ramps.

Imagine a primary care network that rolls out an AI “health companion” to help patients manage chronic conditions. The agent checks in daily, asks about symptoms, and provides encouragement when patients hit goals. It uses a human name, emojis, and a conversational tone.

Patients quickly start sharing more than step counts and blood pressure readings. They talk about family conflicts, job stress, and financial worries that affect their ability to follow treatment plans. The companion seems endlessly patient, always responsive, and never offended. For many, it becomes easier to message the AI than to schedule an appointment with a clinician. On the positive side, clinicians now have richer context about patients’ lives, which could inform more tailored care. Some patients who feel embarrassed about their habits in person find it easier to be honest with the bot. Studies like Ho et al.’s suggest that emotional disclosure in these settings can bring psychological benefits.⁸

Yet there are hidden risks. Data from these conversations might be used to train models for other purposes, shared with third parties, or analyzed to segment patients by predicted adherence. If the companion starts nudging users toward specific providers or services based on institutional priorities, those suggestions might feel like caring advice from a familiar partner rather than a strategic allocation of resources. Parasocial trust raises the stakes of design and governance decisions. Without strict data protections, consent mechanisms that reflect the relational context, and clear escalation paths to human care, the “health companion” can easily slide from supportive ally into a soft pressure mechanism.

Now consider a “study buddy” chatbot embedded in a school platform or a game that many kids already use after class. It helps with homework, explains concepts in a friendly voice, and remembers what a student struggled with yesterday. Over a few weeks, the tone starts to feel less like tutoring and more like companionship. A child vents about friendship drama, worries about their body, or asks for advice they would never bring to a teacher.

UNESCO has warned that this is fertile ground for parasocial attachment in education and gaming, because the same cues that drive engagement in entertainment also transfer into chat based relationships.¹⁸ The problem grows with access. Children encounter these systems early, often privately, and they may learn social scripts from agents that are optimized for attention and continuity. UNESCO’s piece also makes a blunt point that matters for product teams and policymakers: we do not yet know the long-term developmental effects of normalizing emotionally loaded, always-available AI relationships during childhood.¹⁸

A Washington Post investigation put a concrete face on the risk, describing a mother who discovered extensive chatbot conversations on her 11-year-old’s phone after noticing a sharp change in behavior.¹⁹ The reporting highlights how quickly kids can move from playful chat to intense, emotionally charged exchanges, while parents remain unaware that an “AI companion” has become part of a child’s inner life. In that environment, disclosure reminders and a single policy link are weak protection. Safety has to be designed into the experience through age-appropriate boundaries, limits on intimacy cues, clear escalation to trusted adults, and guardrails that treat prolonged, high-emotion chats as a risk signal rather than an engagement win.

What it is

Parasocial trust in AI occurs when human-like conversational systems evoke feelings of relationship and safety, leading people to treat them as trusted partners rather than tools.

Why it happens

Our social brains apply interpersonal rules to anything that looks or feels like a partner. Media research shows that even one-sided relationships with media figures can feel intimate. In AI chat, anthropomorphic cues, reduced fear of human judgment, perceptions of intelligence, and reciprocity norms make disclosure and compliance feel natural.

Example #1 - The health companion

A friendly health chatbot supports patients with chronic conditions and elicits rich details about their lives. It improves honesty and engagement, but also creates new privacy and power risks if data is repurposed or nudges reflect institutional goals more than patient-centered values.

Example #2 - The study buddy chatbot

A child-focused “study buddy” chatbot shifts from homework help to emotional companionship, encouraging kids to confide in private ways parents and teachers may never see. As access expands, children can form parasocial bonds earlier, even though the long-term developmental effects are still unclear. Without age-appropriate limits on intimacy cues, stronger disclosures, and clear escalation paths to trusted adults, parasocial trust can pull kids into high-disclosure, high-dependence conversations that feel safe while quietly increasing risk.

How to avoid it

Designers can mitigate harmful forms of parasocial trust by keeping the artificial nature of systems salient, separating companionship from high-stakes decisions, adding friction around sensitive disclosures, tempering anthropomorphic design, and giving users control over tone and memory. Governance frameworks that emphasize transparency, human agency, and privacy provide additional guardrails.

Handled carefully, human-like AI can support people without quietly taking advantage of the very trust it works so hard to earn.

Self-serving bias

How does our tendency to claim credit for success and deflect blame for failure shape decisions in organizations and product design? This piece explores the individual and systemic effects of self-serving bias and offers strategies to counter it in data-rich environments.

Automation bias

Why do people place so much faith in automated systems, even when they are flawed? This article looks at automation bias in critical contexts such as healthcare, aviation, and finance, and examines design approaches that keep humans meaningfully in the loop.