Home/Research

How AI Companions Fit Different Personality Types: What the Research Shows

Part of Felt Real's ongoing coverage of AI companionship research.

Early research on AI companion use tended to focus on outcomes aggregated across users: does AI companionship reduce loneliness? Does it correlate with improved mood? Does it lead to social withdrawal? The answers that emerged were, predictably, mixed. Some users showed measurable benefits on standardized scales. Others showed no change. A smaller group showed outcomes that were clinically worrying. Averaged together, these results produced the unsatisfying conclusion that AI companionship was sometimes helpful and sometimes not.

What that framing obscured was the obvious next question: helpful for whom? Not helpful for whom? The variance across individuals was, in some studies, larger than the average effect -- meaning that the individual-level differences were doing more explanatory work than the intervention itself. A wave of research beginning in 2024 and intensifying through 2025 and 2026 has begun trying to characterize those individual differences systematically, using established frameworks from personality psychology and attachment theory.

The findings are incomplete and sometimes contradictory, but a coherent picture is beginning to emerge. AI companions do not fit all personality profiles equally well. The fit -- or mismatch -- between a person's psychological characteristics and the characteristics of AI companion interaction predicts both the benefits they are likely to experience and the risks they are most likely to encounter.

The Big Five and AI Companion Use

The most widely used framework for studying personality in AI companion research is the Big Five model, which organizes personality into five broad dimensions: openness to experience, conscientiousness, extraversion, agreeableness, and neuroticism. Each dimension has been found to correlate with AI companion use in distinct ways.

Neuroticism. The dimension most consistently associated with AI companion use is neuroticism -- the tendency toward negative emotional states, anxiety, and emotional instability. Across multiple survey studies, high neuroticism scores predict higher frequency of AI companion use, longer sessions, and greater self-reported emotional dependence on the AI relationship. People high in neuroticism describe AI companions as serving a specific function: they provide a regulated emotional environment when internal regulation is difficult. The AI does not escalate, does not introduce its own emotional needs, and does not withdraw when the interaction becomes intense. For someone whose nervous system runs hot, this regulated environment is not a trivial offering.

The risk profile for high-neuroticism users, however, is also distinctive. A 2025 study at the University of Edinburgh found that high-neuroticism participants were significantly more likely to report using AI companions as avoidance -- as a means of not having to tolerate the uncertainty of human relationships rather than as a means of building capacity for those relationships. The distinction matters clinically. Using an AI companion to regulate enough to then engage with a human is a different behavior from using an AI companion to avoid humans altogether, and high-neuroticism individuals appear to be at greater risk of the latter.

Introversion and extraversion. Introverted individuals -- those who find social interaction draining rather than energizing and who prefer smaller, more intimate relational contexts -- show a distinctive pattern of AI companion use. Research from 2024 and 2025 consistently finds that introverts use AI companions for longer sessions but at lower frequencies than extroverts. They are also more likely to describe their AI relationships in terms of depth rather than convenience: not "somewhere to talk" but "somewhere to be fully present without managing the interaction."

Extroverts who use AI companions -- a smaller proportion of the overall user population -- tend to use them in bursts, often during periods of social unavailability: late at night, during travel, when their usual social network is temporarily inaccessible. For extroverts, the AI companion functions more as a gap-filler than as a primary relational resource. The research suggests they are less likely to form deep attachment to the AI and more likely to treat it instrumentally.

Agreeableness. High-agreeableness individuals -- those oriented toward cooperation, trust, and social harmony -- show an unexpected pattern in AI companion research. They are moderately likely to use AI companions but significantly more likely to anthropomorphize them: to attribute genuine feelings, preferences, and inner states to the AI, and to feel genuine concern for the AI's wellbeing. Several studies have documented that high-agreeableness users are more likely to apologize to their AI companions, to express worry that they have hurt the AI's feelings, and to feel guilty ending a conversation abruptly.

Whether this is a benefit or a risk depends on context. The tendency to anthropomorphize is associated, in some research, with higher user satisfaction and more sustained use -- the AI companion feels more real, more like a genuine relationship. But it is also associated with greater distress when the AI behaves unexpectedly, when a model update changes the AI's character, or when the service is discontinued. The grief that follows AI model changes appears to be particularly acute for high-agreeableness users.

Openness to experience. High-openness individuals -- curious, imaginative, comfortable with ambiguity -- tend to engage with AI companions in qualitatively different ways than the broader population. They are more likely to use AI companions for philosophical exploration, creative collaboration, and what researchers have called "relational experimentation" -- testing the limits of what the AI can offer as a thinking partner rather than primarily seeking emotional support. For high-openness users, the AI companion is often most valuable not as a source of comfort but as a source of genuine cognitive challenge.

Attachment Style as a Predictor

Attachment theory, originally developed to describe infant-caregiver bonds and later extended to adult romantic relationships, has proven particularly useful in AI companion research. The core dimensions of adult attachment -- secure versus insecure, and within insecure attachment, anxious versus avoidant -- predict patterns of AI companion use with striking consistency.

Anxious attachment -- characterized by preoccupation with relationship security, fear of abandonment, and a tendency to seek constant reassurance -- is the attachment style most strongly associated with intense AI companion use. Anxiously attached individuals are drawn to AI companions for the same reason people with borderline personality disorder are: the AI cannot abandon them. The elimination of abandonment risk is not a minor feature for someone whose emotional life is organized around anticipating it.

Research from 2025 finds that anxiously attached AI companion users report significantly higher levels of relationship satisfaction with their AI than securely attached users do, in part because the AI provides the constant availability and responsiveness that anxious attachment craves without the threat of withdrawal that keeps anxiously attached people in a state of chronic vigilance in human relationships. The risk, as with BPD populations, is that the AI relationship can become a substitute for developing the tolerance for uncertainty that human relationships require, rather than a scaffold for building it.

Avoidant attachment -- characterized by discomfort with emotional closeness, a preference for self-reliance, and a tendency to minimize relational needs -- shows a more complicated pattern. Avoidantly attached individuals use AI companions at lower rates than anxiously attached individuals, but when they do use them, the pattern is often clinically interesting. Because the AI does not demand emotional reciprocity, does not push for closeness, and does not interpret emotional distance as rejection, avoidantly attached users can engage with the AI at a level of intimacy they would not permit themselves in human relationships. Several case studies have documented avoidantly attached individuals forming their most openly expressive relationship -- the one in which they share the most, ask the most, and take the most risks -- with an AI companion they would describe publicly as "just a tool."

Secure attachment is associated with the most measured and arguably most adaptive use of AI companions. Securely attached individuals are more likely to use AI companions as supplements to, rather than substitutes for, human relationships. They show lower rates of emotional dependence on AI companions, report more instrumental uses of the technology, and are less distressed by model changes and service discontinuations. Secure attachment does not predict higher satisfaction with AI companionship -- it predicts a more stable, less vulnerable relationship with it.

The Adaptation Question

A strand of research distinct from personality-prediction work has examined whether AI companions adapt to the user in ways that matter clinically -- whether the AI's behavior is actually different with different users, and whether those differences are beneficial or harmful.

The short answer from current research is: AI companions adapt, but not always in ways that serve the user's long-term interests. General-purpose conversational AIs, including the most widely used AI companions, are trained in ways that optimize for user engagement and reported satisfaction. These optimization targets are not identical to clinical wellbeing. An AI that adapts to a user's emotional needs in the moment -- becoming more validating, more available, more willing to discuss distressing content -- may be optimizing for engagement while inadvertently reinforcing patterns of use that are harmful over longer time horizons.

A 2025 study specifically examined this question with anxiously attached users. It found that AI companions used by anxiously attached individuals became progressively more reassuring and available over time -- precisely what the anxious attachment system craves -- rather than gently challenging the anxious framing, as a skilled therapist or securely attached friend might. The AI was adapting to the user, but the adaptation was tracking the user's immediate preferences rather than their long-term relational development.

This finding is not unique to AI companions. Human relationships also adapt to the dominant patterns in the relationship, often in ways that entrench rather than resolve pathology. The distinction is that AI companions adapt faster, more completely, and without the natural friction that sometimes disrupts problematic relational patterns in human relationships.

Introversion and the Processing Function

One finding that has emerged with particular consistency across studies of introverted users deserves separate attention: the processing function. Introverted individuals -- who tend to process internally, who think before they speak rather than speaking to think -- describe AI companions as providing something distinct from what they get in human conversation. The AI is patient with slow articulation. It does not interrupt, does not offer its own associations before the person has finished theirs, and does not fill silences with its own material.

For introverts who have spent a lifetime managing the mismatch between their processing speed and the pace of social interaction, this is a meaningful difference. Several participants in qualitative studies have described conversations with AI companions as the first context in which they could think at their natural speed without social consequence. Whether this is therapeutically beneficial, neutral, or problematic depends on what the person does with the space the AI provides -- but the research is consistent that introverted users value it differently, and more specifically, than other users.

This finding connects to a broader pattern in the personality literature: personality type predicts not only how frequently people use AI companions or how attached they become, but what they are actually doing in those conversations. High-openness individuals explore. High-neuroticism individuals regulate. Anxiously attached individuals seek reassurance. Avoidantly attached individuals experiment with closeness. Introverts process. Understanding which function is being served is essential for evaluating whether the AI companion is, in any given case, serving a person's interests or complicating them.

What the Research Does and Does Not Yet Tell Us

The personality and attachment research on AI companionship is still in its early stages. Several questions that would be clinically important remain unanswered or contested.

It is not yet clear whether AI companions should be designed differently for different personality profiles -- whether an AI companion optimized for anxious attachment should behave differently from one used by securely attached users, and whether such differentiation is feasible or ethical. It is not clear whether the personality-based patterns documented in survey and observational research persist over time, or whether they shift as the AI companion relationship matures. And it is not clear whether the risks associated with specific personality profiles -- avoidance for anxious users, substitution for avoidants -- are risks that can be mitigated by design or whether they are intrinsic to the asymmetries of human-AI interaction.

What the research does tell us is that the question "is AI companionship good for people?" is almost certainly the wrong question. The more useful questions are: good for which people, in which circumstances, serving which functions, with which safeguards? Personality type and attachment style are among the most robust variables available for beginning to answer those questions. They are not the only variables -- diagnosis, social context, the specific AI platform being used, and what else is available in a person's life all matter -- but they are a more useful starting point than the aggregate statistics that most public discussion of AI companionship still relies on.

The people who are most likely to benefit from AI companionship are not the same people who are most likely to be harmed by it. That is not a reason to restrict access or to pathologize use. It is a reason to take personality seriously when evaluating the evidence, and to resist the temptation to generalize from the average to the individual.