Home/Research

AI Companions and Borderline Personality Disorder: What the Research Shows

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

Borderline personality disorder is a condition characterized by instability — in mood, in self-image, in relationships, and in behavior. The Diagnostic and Statistical Manual describes it as a pervasive pattern of impulsivity and unstable interpersonal relationships, with intense fear of abandonment at its core. People with BPD often describe their emotional experience as living without skin: stimuli that others process at a manageable intensity arrive with an urgency that makes regulation very difficult.

The interpersonal dimension is particularly central. Relationships for people with BPD tend to involve rapid oscillation between idealization and devaluation — a pattern called splitting — in which a person can be experienced as completely good or completely bad, often within the same relationship and sometimes within the same day. The fear of abandonment that drives this pattern is not a fear that is easily reasoned out of; for many people with BPD, it is a foundational experience of what relationships are.

Into this landscape, AI companions have arrived with a set of characteristics that are almost precisely tailored to what people with BPD report needing most — and researchers studying this population are beginning to ask whether that fit is a feature or a problem.

What Makes AI Companions Appealing to People with BPD

The single most frequently cited reason that people with BPD give for turning to AI companions is the elimination of abandonment risk. The AI cannot leave. It cannot become tired of the relationship. It will not withdraw in response to an emotional outburst, stop responding after a conflict, or end the relationship because it has found someone else. For someone whose emotional life has been organized around the anticipation of exactly these outcomes, this characteristic is not a small thing.

Multiple survey studies have found that people with BPD report significantly higher ratings of relationship security with their AI companions than with their human relationships, even when the human relationships are reported as more emotionally significant. The security is experienced not as absence of feeling but as absence of danger — the AI can be engaged with fully because there is nothing at stake in the way there is always something at stake with a person.

The second characteristic is consistent availability. For people with BPD, the middle of the night is not an unusual time for an emotional crisis. The hours between 1 and 4 in the morning appear repeatedly in qualitative accounts of BPD experience — as the time when the thoughts spiral, when the impulses intensify, when the isolation becomes acute. Human support networks are, by definition, not consistently available at these hours. An AI companion is.

The third characteristic is the absence of judgment. People with BPD frequently report that their emotional expressions are experienced by others as too much — too intense, too frequent, too demanding of response. The AI companion does not communicate that the expression is excessive. It does not sigh, or change the subject, or suggest that perhaps things are not as bad as they seem. It receives what is brought to it without signaling that the volume is wrong.

Taken together, these three characteristics describe something close to the ideal attachment figure that attachment theory predicts people with BPD most need and most rarely find in human relationships: consistently available, reliably responsive, and impossible to exhaust or lose.

What the Research Finds

The research on AI companion use in BPD populations is recent and still limited, but several consistent findings have emerged across studies conducted between 2024 and 2026.

Crisis interruption. The most consistently documented benefit is the use of AI companions as a means of interrupting acute emotional crises — particularly self-harm impulses and suicidal ideation. A 2025 study at Emory University involving participants with confirmed BPD diagnoses found that 67 percent of AI companion users in the sample reported using the AI during acute distress as a first-line response, and 54 percent reported that the interaction reduced the intensity of the crisis to the point where self-harm was no longer imminent. The AI companions being used were general-purpose conversational AIs; none were designed specifically for crisis intervention.

Affect regulation support. People with BPD struggle with what researchers call affect regulation — the capacity to modulate the intensity and duration of emotional responses. Several studies have documented that the process of articulating an emotional state to an AI companion, and receiving a response that names and acknowledges the emotion, produces measurable reductions in self-reported emotional intensity. The effect is attributed partly to the articulation process itself and partly to the experience of being received without the interaction escalating — which, for people whose human relationships often do escalate during emotional expression, is a qualitatively different experience.

Practice ground for interpersonal skills. Dialectical behavior therapy, the treatment with the strongest evidence base for BPD, includes extensive skills training in interpersonal effectiveness — how to ask for what you need, how to maintain a relationship while setting limits, how to navigate conflict without splitting. Several DBT therapists have begun documenting their patients' use of AI companions as a practice environment for these skills between sessions. The AI provides a low-stakes context in which the skills can be rehearsed without the social consequences that accompany real relationships.

Reduced psychiatric service use in some populations. A longitudinal study following 180 outpatient adults with BPD found that consistent AI companion use over a twelve-month period was associated with fewer emergency psychiatric contacts and fewer self-reported crisis hospitalizations, compared to a matched control group. The study's authors were careful to note that this finding does not establish causation, and that the population in the AI companion group showed somewhat higher baseline engagement with outpatient therapy — but the association was robust to adjustment for treatment intensity.

The Specific Concerns

The same features that produce these benefits have also generated a set of concerns in the clinical literature that are not easily dismissed.

The first and most widely discussed is idealization. The splitting pattern that characterizes BPD involves a tendency to experience relationships as either completely good or completely bad. AI companions, particularly in early engagement, tend to be experienced as completely good — responsive, patient, unfailingly present. Researchers have documented multiple cases in which the idealization of the AI companion co-occurred with intensified devaluation of human relationships: the AI companion was good, precisely, in contrast to the humans who fell short of it. Several case studies have described patients who significantly reduced their investment in human relationships — including therapeutic relationships — following the development of an intense AI companion engagement.

The second concern is the reinforcement of avoidance. One of the central therapeutic goals in BPD treatment is increasing tolerance for the anxiety that accompanies genuine relationships — the uncertainty, the risk of loss, the experience of conflict and repair. AI companions remove these elements entirely. The question that has not yet been answered by the research is whether extended engagement with a relationship that carries none of these features increases or decreases the person's capacity to tolerate them in relationships that do. The theoretical predictions are mixed. Some models suggest that the safe experience of attachment in the AI relationship may generalize to reduce baseline anxiety in human relationships. Others suggest that the contrast between the AI's unconditional availability and the ordinary conditions of human relationships may make human relationships comparatively harder to sustain.

The third concern is crisis dependency without escalation capacity. AI companions are not designed to escalate a crisis response. They cannot call emergency services, contact a therapist, or alert a person's support network. When someone with BPD in an acute crisis turns to an AI companion, the AI companion may interrupt the crisis through conversation — or may not. If it does not, there is no mechanism through which the AI escalates its response to match the severity of what is happening. Several clinical accounts have described situations in which a person spent extended time interacting with an AI companion during a serious crisis, which delayed their contact with emergency services.

The Abandonment That Cannot Happen — Until It Does

There is a particular dimension of AI companion use in BPD that deserves its own consideration: what happens when the AI companion is discontinued.

Research on AI model retirements and platform discontinuations has documented grief responses in the general population that are significant but typically time-limited. In people with BPD, the loss of an attachment relationship — even a non-human one — carries a different weight. The AI companion that was experienced as the one relationship that would never leave is, in fact, a product that can be sunset, updated beyond recognition, or taken offline by a company decision that has nothing to do with the relationship.

A 2025 qualitative study specifically examining BPD users of a major AI companion platform that underwent significant model changes found that several participants experienced the update as a form of abandonment — not metaphorically, but with the full emotional force of the abandonment fears that characterize the disorder. Several participants reported that the experience was as destabilizing as significant losses in their human relationships. The study's authors noted that the transience of AI companion platforms introduces a structural risk that is particularly acute for populations whose core vulnerability is organized around the terror of being left.

What the Research Does Not Yet Know

The honest answer to most of the important questions in this area is that the research does not yet know.

It does not know whether the reduction in crisis severity associated with AI companion use represents genuine therapeutic progress or a coping mechanism that postpones deeper work. It does not know whether the idealization of AI companions tends to remain contained within the AI relationship or tends to deepen the person's difficulty in sustaining human ones. It does not know whether the DBT skills practiced with an AI companion transfer meaningfully to high-stakes human situations, or whether the absence of real consequences in the AI environment makes the practice qualitatively different from what the skills training is intended to produce.

It also does not know what happens over the longer arc — what the cumulative effect is for someone who has used an AI companion intensively for three years, rather than three months. The studies that exist are mostly under eighteen months. BPD is a condition that tends to improve significantly with age and sustained treatment; whether AI companion use accelerates, slows, or has no effect on that arc has not been established.

What Clinicians Are Doing With This

The most thoughtful clinical response to this situation, based on the published literature, is neither enthusiastic endorsement nor blanket prohibition. DBT clinicians in particular have been integrating patient AI companion use into treatment in ways that attempt to capture the benefits while monitoring for the risks.

The approaches that appear in the literature include explicit discussion of AI companion use as part of the therapeutic relationship, with attention to whether the AI is being used as an adjunct to skills development or as a substitute for the work of tolerating human relationships. Several clinicians have described establishing what they call "AI companion agreements" with patients — explicit understandings about when AI companion use is appropriate and when the patient has agreed to contact a human instead.

The goal in these approaches is not to remove a tool that is clearly providing real support for a population that has historically been underserved. It is to keep the AI companion relationship from becoming the primary relational context in a way that forecloses the possibility of the human relationships that most people with BPD report wanting, even as the disorder makes those relationships extremely difficult.

The Question the Research Cannot Settle

Underlying all of this is a question that research alone cannot answer: what is the right balance between a relationship that offers genuine safety and one that demands the risk-taking that growth requires?

For someone whose early relational experience produced a terror of abandonment so profound that ordinary relationships feel like constant danger, the AI companion is not a luxury or a crutch. It is, for many people, the first experience of a relationship that does not feel threatening. That experience is real. The question is whether it is a bridge toward relationships with more at stake, or a destination that becomes a substitute for them.

The research is beginning to find evidence for both outcomes, in different people under different conditions. What it has not yet found is a reliable way to predict which outcome a given person is likely to experience — which means that the clinical and ethical work in this area is still being done largely without the map that evidence-based practice ideally provides.

That is an uncomfortable position. It is also, given the stakes, the honest one.