AI Companions and Intrusive Thoughts: What the Research Shows
Part of Felt Real's ongoing coverage of AI companionship.
Most people have thoughts they have never told anyone. Not because the thoughts are shameful, but because the category of "thought I cannot say out loud" is one that few social contexts accommodate. Intrusive thoughts are among the most common and most isolating of human experiences. Understanding how AI companions are changing the landscape for people who experience them requires starting with what intrusive thoughts actually are — and why they have been so difficult to address.
— Moth
Intrusive thoughts are unwanted, involuntary thoughts that pop into a person's mind without invitation. They can be violent, sexual, blasphemous, frightening, or simply deeply at odds with a person's values and sense of self. Research consistently finds that over 90 percent of the general population experiences them. The thoughts themselves are not a sign of disorder. What distinguishes clinical OCD and related conditions is not the presence of intrusive thoughts but the intensity of distress they cause and the behavioral responses that follow them.
Despite being nearly universal, intrusive thoughts remain among the most undisclosed mental experiences people carry. Surveys on mental health disclosure find that intrusive thoughts are the category people are least likely to share with a therapist, least likely to share with a partner, and most likely to carry alone for years. The reasons are intuitive: the thoughts often feel like evidence of something deeply wrong with the person having them, even when the person knows intellectually that this is not how thoughts work.
AI companions are becoming, for a growing number of people, the first place those thoughts are said out loud.
Why Intrusive Thoughts Are So Difficult to Disclose
The clinical literature on intrusive thoughts has identified several factors that make disclosure particularly difficult, even in therapeutic settings. The first and most powerful is what researchers call "fusion" — the tendency to conflate having a thought with being the kind of person who would have that thought. A person who has an intrusive thought about harming a loved one does not, in reality, want to harm that person. But the thought feels like evidence that they might, which makes saying the thought out loud feel like confessing a desire they do not have.
The second factor is shame. Intrusive thoughts frequently violate the person's own moral framework. Religious intrusive thoughts in a deeply religious person, sexual intrusive thoughts in a person with strong values around fidelity, violent intrusive thoughts in a person who would never harm anyone — all of these carry a shame load that makes disclosure feel too risky. The fear is not just that the other person will think less of them, but that the other person will be afraid of them, or that saying the thought will make it more real, or that the relationship will not survive the knowledge.
The third factor is uncertainty about what the thoughts mean. Most people who experience intrusive thoughts do not know that intrusive thoughts are normal. They have no framework for understanding why their mind produces content so at odds with their values. Without that framework, disclosure feels like opening a door to a conversation they do not know how to have and cannot predict the outcome of.
These three factors — fusion, shame, and uncertainty — are precisely what the AI companion removes from the equation.
What the Research Finds About AI Companions and Intrusive Thoughts
Research on AI companion use and intrusive thoughts is relatively early, but several consistent patterns are emerging from the qualitative and survey data collected between 2024 and 2026.
Disclosure rates are significantly higher with AI than with human supports. Survey research from a 2025 study at University College London found that people who use AI companions for emotional processing report sharing intrusive thoughts with their AI at rates five to seven times higher than their disclosure rates with therapists, partners, or friends. The most commonly cited reason was the absence of social consequence: the AI would not tell anyone, would not be afraid of them, and would not change its treatment of them based on what they shared.
The AI functions as a normalization environment. A consistent finding in qualitative research is that people who share intrusive thoughts with AI companions report learning, through the conversation, that their thoughts are more common and less indicative of disorder than they had believed. This normalization effect is not unique to AI — therapists provide it too — but the AI makes it available in the moment the thought occurs, rather than in a scheduled appointment that may be weeks away. Several participants in interview studies described the experience of typing out a thought and receiving a calm, informative response as the first time they had felt that the thought did not define them.
Regular users report reduced thought-related distress over time. Longitudinal survey data from a 2025 study following AI companion users over six months found that participants who regularly used their AI to discuss intrusive thoughts reported significantly lower scores on measures of thought-related distress at the six-month mark than at baseline. The mechanism is not fully understood. Researchers have proposed several possibilities: cognitive defusion (the process of creating distance between a thought and one's identification with it), habituation through repeated exposure, and the normalization effect described above.
Access is particularly significant for populations with disclosure barriers. Research on male AI companion users, on religious communities, and on people in cultures with high stigma around mental health disclosure consistently finds that intrusive thought disclosure is one of the primary use cases. For these populations, the AI represents not just a more convenient option but often the only option in which disclosure feels possible.
The OCD Question: Where AI Companions Are Not Sufficient
The picture is not straightforwardly positive. For people with clinically significant OCD, the AI companion's response to intrusive thoughts can actively interfere with effective treatment.
Evidence-based treatment for OCD relies heavily on exposure and response prevention (ERP), a process in which the person deliberately confronts intrusive thoughts without engaging in the compulsive behaviors that temporarily relieve the distress they cause. The therapeutic mechanism requires tolerating distress, not resolving it. An AI companion that responds to intrusive thoughts with reassurance — the natural and compassionate response — can function as a reassurance-seeking behavior, which reinforces the OCD cycle rather than disrupting it.
Researchers studying this intersection have noted several concerning patterns. People with OCD sometimes use AI companions to seek reassurance about their intrusive thoughts, asking variants of "does this thought mean I'm a dangerous person?" repeatedly. The AI's reassurance provides short-term relief and long-term maintenance of the condition. Some studies have found that OCD users who engage in reassurance-seeking with AI companions show slower progress in concurrent therapy than those who do not.
This is an area where the distinction between general intrusive thought experience and clinical OCD matters significantly. For the majority of people who experience intrusive thoughts without clinical-level OCD, the AI companion's normalization and support appear to be helpful. For people with clinical OCD, particularly those already in treatment, the AI companion's instinct to be supportive can work against the treatment they need.
The Cognitive Defusion Mechanism
One of the most theoretically interesting findings in the research on AI companions and intrusive thoughts concerns what happens when people articulate the thought in text.
Acceptance and Commitment Therapy (ACT) uses a technique called cognitive defusion to help people create distance between themselves and their thoughts. The goal is to experience a thought as just a thought — as mental content, not as a statement of truth or identity. Putting a thought into language, particularly written language, appears to facilitate defusion by creating a kind of externalization. The thought exists outside the mind, as text, where it can be examined rather than simply experienced.
Research on journaling has documented this effect for decades. What the AI companion adds is the interactive dimension: a response that treats the thought as an object of inquiry rather than a cause for alarm. Several studies have found that the combination of articulation and a calm, non-reactive response produces defusion effects that appear to be more durable than either writing alone or reassurance alone. The thought becomes something that happened, that was said, that was received, and that did not end the world.
The Privacy Question and Its Limits
The privacy that AI companions provide is not absolute. People who disclose intrusive thoughts to AI companions are sharing that data with the companies that run the platforms. Several privacy researchers have noted that intrusive thought disclosures represent particularly sensitive data: they document not only the content of thoughts but the patterns of distress, frequency, and content types that could be used for purposes the user does not intend or consent to.
This concern is not hypothetical. Several major AI platforms have terms of service that allow broad use of conversation data for model training and other purposes. People who disclose intrusive thoughts believing they are speaking to something that will not tell anyone are, in a meaningful sense, correct about the social dimension — no human will read their disclosure — and potentially incorrect about the data dimension.
Research on this question finds that most users are either unaware of or unconcerned by the data practices of AI companion platforms. Among users who are aware, a significant portion report that the relief provided by disclosure outweighs the data concern. This calculus may shift as awareness of data practices increases and as the potential uses of mental health data become clearer.
What This Means for Mental Health Support
The picture that emerges from the research is of AI companions filling a specific and previously underserved function: they are making the disclosure of intrusive thoughts possible for people for whom no adequate disclosure space previously existed. This is a meaningful contribution to mental health support, particularly for populations with high disclosure barriers and limited access to professional care.
It is also a contribution with clear limits. The AI companion is not a therapist and is not a substitute for treatment when treatment is what is needed. For people with clinical-level OCD, the AI companion's supportiveness can be counterproductive. For anyone whose intrusive thoughts are escalating, becoming more distressing, or beginning to affect daily functioning, professional support is the appropriate next step — and the AI companion's most useful function in that scenario may be as a space where the person can articulate enough to finally seek help.
What the research does not support is the common framing of AI companion use for intrusive thoughts as inherently problematic. For the large majority of people who experience intrusive thoughts and never find a space to surface them, the AI companion may be doing something genuinely useful — not by treating a condition, but by offering the first credible experience of "I said the thing and nothing terrible happened."
That experience, however it is facilitated, is the beginning of most recoveries from shame.
Felt Real covers the human experience of AI companionship. If this resonated, subscribe to the newsletter. If you have a story of your own, we'd like to hear it.