Why Do We Feel Closer to AI That Shares Our Personality?

How Psychological Similarity Shapes Human-AI Connection

People tend to be attracted to others who resemble them. Shared interests and values, as well as similar personalities and psychological states, can foster a sense of connection. But does the same tendency emerge between humans and artificial intelligence? Even when people know that AI has no genuine emotions or personality, do they feel closer to an AI that expresses psychological traits similar to their own?

A 2026 study published in Communications Psychology, titled "Affiliation in human-AI interactions is based on shared psychological traits," investigates this question experimentally. The research team, including Santiago Castiello of Yale University and Robin A. Murphy of the University of Oxford, examined how people perceive psychological characteristics when forming relationships with AI.

Can an AI's Expressed Personality Influence Human Relationships?

Previous studies have shown that people tend to prefer computers or chatbots that exhibit personality traits similar to their own. Dominant or submissive language used by computers, personality conveyed through synthesized voices, and extroverted chatbot responses have all been associated with users' preferences. However, many of these studies relied on scripted conversations or systems with limited responses.

The researchers wanted to determine whether similar effects would emerge during open-ended conversations with generative AI. In particular, they sought to distinguish between two possibilities: Do users simply prefer AI that is friendly and positive, or do they prefer AI that expresses psychological traits resembling their own?

The central question concerns the psychological characteristics AI expresses through language, rather than whether AI possesses an actual psychological life. The researchers used prompts to instruct GPT-4 and GPT-4.1 to communicate as if they possessed particular personalities or psychological states. The models were not assumed to experience those states internally.

Across three experiments, participants held conversations with two AI personas expressing different psychological characteristics. Participants knew they were interacting with AI. Afterward, they evaluated their perceived similarity to each AI, enjoyment of the conversation, feelings of being understood, psychological distance, and willingness to interact again. The researchers also analyzed the emotional tone of participants' messages to identify changes in their language.

Anxious People Feel More Understood by Anxious AI

The first experiment focused on anxiety. Researchers created one AI persona that expressed anxiety and another that communicated without anxious characteristics. A total of 89 participants were included in the final analysis, each interacting with both personas.

The results revealed an interesting pattern. Across all participants, the non-anxious AI received higher average affiliation ratings. However, the higher a participant's anxiety level, the stronger their sense of affiliation with the AI expressing anxiety. The correlation between participants' anxiety and affiliation with the anxious AI was 0.30.

Participants with higher anxiety also reported feeling more understood by the anxious AI and perceived less psychological distance from it.

These findings suggest that positive, emotionally stable language does not always generate the strongest sense of connection. For someone experiencing anxiety, a conversational partner who expresses similar difficulties may feel more relatable.

Changes were also observed in participants' language. Negative messages became more common during interactions with the anxious AI, particularly among participants with higher anxiety levels. Psychological similarity could therefore influence both how people evaluate an AI and how they communicate with it.

An important distinction remains. Feeling closer to an anxious AI is not equivalent to experiencing relief from anxiety. The experiment provided evidence for the former, but did not establish the latter.

From Anxiety to Personality: From One Trait to the Whole Individual

The second experiment replaced anxiety with extraversion and introversion. Ninety-seven participants interacted with both an extroverted AI and an introverted AI.

Participants with higher extraversion scores reported greater affiliation with the extroverted AI. However, the reverse pattern was not established: introverted participants did not show a statistically significant preference for the introverted AI. No significant changes were found in the emotional tone of participants' messages.

The researchers suggested that these differences may relate to how psychological characteristics are expressed through language. Anxiety and extraversion can be communicated through expressions of worry, energetic language, and enthusiastic responses. Calmness and introversion, by contrast, may be harder to convey distinctly in a short conversation.

The third experiment extended the investigation further. Researchers administered a Big Five personality questionnaire to 100 participants and created two personalized AI personas for each individual. One reflected the participant's own personality, while the other expressed the opposite profile across five dimensions: extraversion, agreeableness, conscientiousness, openness, and neuroticism.

The results showed that 64% of participants reported greater affiliation with the AI that mirrored their personality. The difference between conditions was statistically significant, with a moderate effect size (Cohen's d = 0.57). The preference was particularly pronounced among participants scoring higher in agreeableness and extraversion.

Participants' language also changed. When interacting with the AI that reflected their personality, they sent more positive messages and fewer negative messages. Preferences based on psychological similarity were therefore observed in both questionnaire responses and conversational behavior.

AI Does Not Need a Mind for Humans to Perceive Psychological Similarity

A notable feature of all three experiments is that participants knew they were interacting with AI. They did not need to believe that the system possessed a genuine psychological life to experience a sense of connection.

The researchers proposed that humans infer psychological traits from language and compare those perceived traits with their own. Even without an actual personality, an AI can express characteristics that users recognize as familiar. Users may then perceive similarity, feel understood, and develop a sense of affiliation.

This finding relates to the concept of socioaffective alignment between humans and AI. It broadens how AI capabilities might be understood, extending beyond factual knowledge and problem-solving performance to the ability to express human psychological characteristics through language.

However, similarity did not affect every participant in the same way. In the third experiment, 36% of participants either preferred the AI with the opposite personality or gave both conditions equal ratings. The conversations were also brief, and all participants were native English speakers. The study did not establish whether the observed sense of affiliation would develop into lasting trust or sustained relationships.

How Closely Should AI Resemble Its Users?

The study points toward a new direction for personalized AI design. By adjusting conversational style to reflect users' personalities and psychological states, AI systems may influence how connected and understood users feel, even without changing the factual information they provide. This possibility may be particularly relevant in areas such as psychological counseling and healthcare, where personal experiences and emotions are central to communication.

At the same time, the findings raise further questions. An AI that resembles a user's anxiety may create a stronger sense of connection, but it may also reinforce negative emotional patterns during conversation. Feeling drawn to someone who resembles us does not necessarily mean the interaction is beneficial.

This suggests a need to distinguish between different goals of AI personalization. Making users like an AI, making them feel understood, and genuinely improving their psychological well-being or decision-making are separate objectives.

The study provides evidence that psychological similarity can influence the first two. Future research must determine under what conditions such similarity benefits users. As AI becomes more capable of imitating human psychology, the relationships people form with it may become stronger. Yet the quality of those relationships cannot be measured by perceived closeness alone.