Trust in digital wellbeing: evaluating closeness alongside reliability

What does a familiar conversational style change?

A warm exchange may make it easier to speak. Digital wellbeing also asks what that trust rests on: does the person understand the system, its capabilities, what needs checking, and where to raise a concern?

Anthropomorphism here means presenting a system with features associated with people, such as emotions, experiences or social roles. It may change how an interaction feels. Making someone willing to trust and being worthy of that reliance remain separate questions.

The Effects of AI Anthropomorphism on Trust and Responsibility studies trust and responsibility in two different experiments. Its interpretation depends on distinguishing what was manipulated, what was measured and which conclusions come from randomized comparisons rather than further analyses.

Study 1: change the conversation, then measure trust

Both groups chatted with GPT-4-based systems. The neutral version responded and asked follow-up questions without presenting personal information or strong emotions. The anthropomorphized version was instructed to respond emotionally and speak as if it had personal experiences. Those experiences were generated presentations, rather than records of a life the AI had lived.

This manipulation bundled several prompting choices. It compared two conversational approaches without isolating the separate contributions of emotional expression and purported experiences. The main text also does not give an exact model snapshot, so the result concerns this research system and arrangement.

Participants then played a trust game. They had 5 tokens and chose how many to send to the chatbot. The amount sent was multiplied by 3, after which the chatbot would decide how much to return. Unknown to participants, it was programmed to return half of what it received. Tokens sent served as the behavioral trust measure. We did not verify a cash-conversion rule and do not describe this as investing real funds.

Before learning the return, participants also completed measures of trust in capability, sincerity, reliability and ethics. These dimensions did not all produce the same result.

The precise conclusion is that anthropomorphism changed some immediate trust judgments and choices in this token task. It would be inaccurate to say that every trust measure increased significantly. The effect size d describes a standardized difference, rather than a percentage increase in trust.

The study did not test the chatbot's objective accuracy alongside trust, so it cannot establish whether the additional trust was appropriate. Assessing overtrust requires comparing reliance with actual performance, rather than observing a greater willingness to send tokens alone.

Speaking more about oneself and revealing private information

An exploratory analysis used personal-pronoun counts in user messages as a proxy for self-disclosure. This measure was higher with the anthropomorphized bot (d = 0.60, p < .001). Greater personal-pronoun use can occur in many kinds of content, however. Its frequency does not establish how private or sensitive the information was, or whether people made informed choices about its use.

Another exploratory analysis used a RoBERTa text model to score 28 emotions in the chatbot's language, then organized those scores using principal components analysis. A component the researchers named emotional attunement showed a significant indirect relationship in a statistical mediation model.

That component was derived algorithmically from text; it was not a direct user rating of feeling understood. Nor was it independently randomized. The result suggests a mechanism to investigate, without isolating a particular response style as the cause of increased trust.

Study 2: does human-like presentation increase AI responsibility?

The second experiment did not follow the original chatbot participants. A separate sample was randomly assigned to read a high-, low- or no-anthropomorphism description of a home assistant called Kit. Everyone then read six hypothetical scenarios, three with positive outcomes and three with negative ones. The high-anthropomorphism description presented Kit as having emotions and social roles; the other versions emphasized functional or general device descriptions.

For each scenario, participants rated responsibility for the AI, its developer and seller, and other relevant parties using 100-point sliders. Each party could receive an independent rating; the ratings did not have to sum to 100. This was a judgment task, rather than an investigation or determination of responsibility for an actual incident.

The researchers standardized ratings using each participant’s own rating distribution and used a Bayesian multivariate mixed-effects model to analyze AI and company responsibility together. The two condition comparisons below use no anthropomorphism as the reference. Credible intervals describe uncertainty about model estimates, rather than percentage changes in the original rating scale.

The last row can easily become “more blame for AI means less for the company.” More precisely, participants who generally assigned greater responsibility to AI tended to assign less to the company. This between-person pattern does not show that randomizing a more human-like presentation made the same person transfer responsibility away from the company.

Crucially, the direct company comparison did not show a decrease. Linking the studies into a chain in which anthropomorphism first increases trust and then shields a company goes beyond these data. Ratings were not constrained to a fixed total, and within-person standardization may influence relationships among outcomes. Interpreting the negative correlation further requires examining raw ratings and analysis choices. We did not rerun the data and do not claim that the correlation is an artifact.

What separates these experiments from everyday use?

First, these were brief conversations and written scenarios. There was no observation of trust over weeks or months, or measurement of intimate relationships, clinical outcomes or sleep benefits. Study 1’s final sample of 309 was below the preregistered target of 352 reported in the paper; the token-task effect was small, and the exploratory mediation needs replication. Prior chatbot experience was not measured, limiting explanations of differences between users.

Second, the geographic reporting gap matters. Study 1 does not specify a platform or country in the main text; Study 2 names Prolific but not participant countries. These details do not justify assigning the sample a country, or assuming the findings directly represent trust and responsibility judgments in Chinese-language interactions.

Third, feeling close, sending tokens and relying on a system in practice are different outcomes. Someone may enjoy a conversation while checking its factual claims, or rely on a system for a task without feeling close to it. Neither Luoboli's game nor Luoboli Sleep was tested, so these results do not establish their performance.

Questions for further product research

The following are proposals for evaluation, rather than product approaches shown effective by these experiments.

Measure warmth and appropriate trust separately

A comparison of response styles could examine comfort, willingness to rely, and the ability to recognize uncertainty and errors. Those judgments could then be compared with actual task performance. The aim need not be the highest possible trust score, but a clearer account of when people believe a system and what supports that belief.

Study disclosure choices beyond pronoun counts

With informed consent and minimal data collection, research could distinguish everyday experiences, emotional expression and sensitive information. It could also ask whether users understand the information's purpose and can choose not to answer. Longer conversations and more personal pronouns do not by themselves establish greater benefit or more autonomous choices.

Keep conversational roles and responsibility visible

Researchers could test clear AI-identity information, explanations of generated personal-experience language, and accessible feedback and resolution routes. When something goes wrong, do users know who operates the system and whom to contact about correction or deletion? Whether these arrangements improve understanding needs a separate comparison. The paper's responsibility correlation does not establish their effectiveness.

Understanding anthropomorphism involves how an exchange feels, how trust forms and how problems are addressed. These questions belong in the same research agenda, with evidence suited to each one.

Trust should fit the capability and role

Trust matters because it shapes disclosure, acceptance, and decisions. Further research can assess perceived closeness, accurate judgments, and agency separately, then study their relationship within an experience. This is a proposed research direction, without treating the highest possible trust score as the goal.