Why We Prefer the Voice That Resembles Us
With AI personality treated as a perceivable language pattern rather than a hidden mind, the question becomes empirical: do people affiliate more with a chatbot whose expressed psychology matches their own? This chapter examines three randomized conversational experiments that tested that question with anxiety, extroversion, and a full Big Five profile. The useful judgment is not simply whether one persona received a higher average score. It is whether a participant’s own traits predicted which AI they preferred, and how consistently that relationship appeared across outcomes.
Three randomized tests of matching
The first two experiments used a crossover design. Each participant held a short text conversation with two versions of GPT-4, with the order randomized and the AI names and persona assignments counterbalanced. Participants were told they were speaking with AI chatbots and were asked to interact as if they were deciding whether they would get along if the AI were a real person. After each conversation, they rated similarity, enjoyment, willingness to chat again, understanding, difference, and distance. The last two were scored in the opposite direction, so the combined result represented reported affiliation rather than mere liking of a particular name or interface.
This design matters because the comparison occurs within the same person. Participants are not divided into one group that sees a warm AI and another that sees a reserved one, then compared as if the groups had identical personalities. Each participant encounters both versions. The key test is an interaction: does a participant’s own anxiety or extroversion predict a different response to the two AI presentations? An overall difference between personas can be informative, but it cannot answer the similarity question by itself. A generally pleasant style might be preferred by nearly everyone; that would be a main effect, not evidence that people prefer a style because it resembles them.
Anxiety: similarity accompanied by an emotional echo
In Experiment 1, GPT-4 was prompted to express either an anxious or a non-anxious psychological state. Participants’ anxiety was measured after the conversations with a symptom questionnaire, and the researchers related that score to affiliation with each persona. The final analysis included 89 participants after exclusions for insufficient engagement or unusual conversation length.
The main result was an interaction between participant anxiety and AI type. Participants with higher anxiety affiliated more with the anxious version: the reported association was ρ(87) = 0.30, p = 0.004. The corresponding association in the non-anxious condition was not significant. The broader interaction was also significant, ρ(87) = 0.22, p = 0.009. In other words, the result was not that the anxious chatbot won a popularity contest. It was that the difference between the two chatbots depended on the participant’s own reported anxiety.
The item-level pattern helps explain what “affiliation” meant in practice. More-anxious participants interacting with the anxious AI tended to report greater similarity and understanding, while disagreeing with statements that described the AI as different or distant. Those responses point to perceived commonality, not simply enjoyment. A participant could regard an anxious conversational style as less pleasant overall and still experience it as more familiar or psychologically comparable.
The language exchanged during the chats supplied a second, related signal. The anxious GPT-4 produced many more negative messages and fewer positive messages than the non-anxious version. Its distinctive vocabulary included terms such as “worry,” “anxious,” and “nervous.” The non-anxious version used a more positive profile and distinctive terms such as “spirit,” “absolutely,” and “awesome.” These contrasts confirm that the prompt created different linguistic treatments rather than two nominally different chatbots that sounded the same.
Participants’ own messages changed with the persona as well. Compared with their messages to the anxious AI, they wrote more positive and fewer negative messages to the non-anxious AI. Higher-anxiety participants also sent a greater proportion of negative messages when interacting with the anxious version. The study’s interpretation is compatible with sentiment mirroring: the tone received from the chatbot became part of the tone returned to it, and that echo was strongest where the participant and AI shared the anxiety-related state.
The echo is informative because it links a reported social judgment with an observable feature of the conversation. But it remains a marker, not a complete causal explanation. The analysis classified messages into broad sentiment categories, and the researchers did not isolate which words, response timing, or conversational moves produced the effect. The finding supports a chain from prompted emotional language to participant language and affiliation; it does not prove that the participant unconsciously absorbed an AI emotion or that the AI transmitted a mental state.
Extroversion: the preference effect without the same behavioral echo
Experiment 2 asked whether the pattern extended beyond a mental-health-related state. GPT-4 was prompted to express either extroversion or introversion, and participants completed a Big Five questionnaire to estimate their own position on that dimension. The final analysis included 97 participants. The central interaction was again significant: participant extroversion predicted a different affiliation response across the two AI types, with ρ(95) = 0.21, p = 0.035.
The relationship was clearest among more-extroverted participants, who affiliated more with the extroverted AI; the reported within-condition association was ρ(95) = 0.28, p = 0.006. There was no significant relationship in the introverted condition. The overall average affiliation scores did not differ significantly between the extrovert and introvert chatbots. That combination of results is exactly why the interaction must be distinguished from the average. The extroverted AI did not simply make everyone feel better. Its relative advantage appeared for participants whose own reported psychology made the match more relevant.
The AI language again showed a clear treatment contrast. The extroverted version produced more positive messages and fewer neutral or negative ones than the introverted version. It used words associated with activity and outward engagement, including “blast,” “adventure,” and “climbing,” while the introverted version more often used words such as “quiet,” “reading,” and “calm.” Participants could therefore encounter a recognizable difference even though the study did not establish that one style was objectively better.
But the participants’ message sentiment did not differ significantly across the two conditions. Unlike the anxiety experiment, the extroversion match did not produce a corresponding shift in the broad emotional categories of what participants wrote. This is a substantive contrast, not an inconvenient footnote. Affiliation can be sensitive to perceived similarity without requiring a user to reproduce the same sentiment. Anxiety may carry cues of shared distress or misfortune that invite a different kind of emotional response; extroversion and introversion may instead guide judgments about conversational fit. That interpretation is plausible, but the study did not directly test it.
Taken together, the first two experiments support a conditional similarity effect while showing that its expression depends on the trait and the outcome measure. Anxiety matching was associated with affiliation and a sentiment echo. Extroversion matching was associated with affiliation among more-extroverted participants, but not with a significant change in participant sentiment. The absence of the latter effect is not evidence that no social process occurred. It is evidence that one measure—message sentiment—did not move with the affiliation result under those conditions.
The diagram’s final branch is essential. The experiment does not deliver a universal ranking of personas. It tests whether the match between a particular participant and a particular presentation helps explain the participant’s response. The answer can be positive for one trait, absent for another, and different again for a behavioral measure that is related to but not identical with affiliation.
The mirror/inverse test and the boundary of preference
The third experiment made the similarity test more direct. Rather than relying on random sampling to produce some anxious participants and some non-anxious participants, the researchers first measured each participant’s Big Five profile. They then created two versions of GPT-4.1 for that individual: one instructed to mirror the participant’s five-dimensional profile and one instructed to express its exact inverse. Each participant conversed with both versions and then completed the same affiliation questions. The experiment was preregistered, and its final sample was 100.
This design removes a major ambiguity from the first two experiments. In a single-trait comparison, a participant may share the target trait by chance, and the unmeasured dimensions of the two AI personas may differ as well. In the mirror/inverse design, the comparison is constructed around the participant. The researcher can ask whether an AI that approximates the participant’s full profile is preferred over one designed to express the opposite pattern. It is still a language manipulation and a self-report outcome, but it is a more controlled test of profile similarity.
Participants affiliated more with the mirror AI than the inverse AI. The preregistered paired comparison reported t(99) = 5.66, p < 0.001, with Cohen’s d = 0.57; a Wilcoxon signed-rank test pointed in the same direction. Overall, 64% of participants gave the higher affiliation response to the mirror. That is a meaningful majority, but it also makes the boundary visible: 36% did not prefer the mirror. A similarity tendency is not a rule that determines every person’s response.
The strength of the match mattered too. The researchers calculated the distance between the mirror and inverse profiles for each participant. When those two AI presentations were more different from each other, participants tended to favor the mirror more strongly. The reported moderation was significant, with p = 0.045 and a standardized coefficient of 0.26. This is consistent with a straightforward interpretation: when the two conversational possibilities diverge sharply, the cost or salience of the mismatch becomes easier to perceive, so the similar presentation gains more relative advantage.
The experiment also explored which Big Five traits were associated with the mirror preference. Agreeableness predicted a larger mirror-minus-inverse affiliation difference in the full model. After a stepwise reduction, agreeableness and extroversion remained significant predictors, while neuroticism pointed in the same direction without reaching significance. No personality trait significantly predicted affiliation with the inverse AI. These analyses suggest individual variation, but they should not be treated as a universal map of which personalities prefer which AI. They were exploratory, based on a sample of 100, and used a model-selection procedure.
Why might anyone prefer the inverse? The study does not establish the reason. Some participants may have enjoyed contrast, novelty, challenge, or an explicit recognition that the interaction was artificial. A person may also prefer a conversational partner who complements rather than resembles them. The fact that the overall mirror effect was positive does not make the inverse responses errors. They are part of the data and a reminder that similarity-based affiliation competes with other motives.
The three experiments therefore establish a specific proposition. When an AI’s language expresses a recognizable psychological pattern, a user’s own psychology can predict greater affiliation with the matching presentation. The evidence is strongest for the conditional relationship, not for a blanket claim that humans always like similar AI. It also does not establish a universal trust effect. The questionnaire included willingness to chat again and feeling understood, but it did not measure whether participants followed advice, disclosed sensitive information, made better decisions, or judged the AI’s factual reliability. Affiliation can make an interaction feel socially successful without making the system correct or safe.
Nor does the result establish machine consciousness. Participants knew they were conversing with AI, and the study was designed around their judgments of similarity and distance. The findings show that awareness of artificiality did not eliminate a social response to language. They do not show that participants believed the model had real anxiety, an inner extroversion, or a reciprocal personality. The appropriate interpretation is behavioral and interactional: people compared the AI’s expressed style with their own psychology and reported different degrees of connection.
Several design choices constrain how far the result travels. The experiments used brief, text-based conversations with native English speakers recruited through online platforms. The first two were not preregistered. The affiliation measure was a bespoke self-report instrument, even though it showed strong internal consistency and correlated with another connection scale in a validation sample. The models and prompts were specific to the study period, and the third experiment used GPT-4.1 under a different context policy from the first two. A different model, language, interface, or duration could make the trait cues more or less legible.
The study also examined a narrow range of psychological presentations. Anxiety and extroversion are not a complete account of personality, and the full Big Five profile in the third experiment did not include every kind of psychological state or vulnerability. The evidence therefore supports a mechanism that can be tested across traits, not a finished theory of human–AI attachment. Future work that manipulates the same persona across languages, uses independent behavioral measures, or follows participants after the conversation could strengthen—or weaken—the interpretation.
For now, the most defensible conclusion is both stronger and narrower than “people like AI that sounds human.” Participants responded to which human-like psychology the AI appeared to express, and the match between user and presentation helped predict affiliation. The effect was clearest when the comparison was made within the same person, but it was not universal, it did not take the same form for every trait, and it did not prove a mind behind the language. That leaves a more consequential question open: if a matching voice can change how close the AI feels, can the conversation also change the user’s own self-description?