Home/Research

AI Companions and Health Anxiety: What People Who Worry About Their Health Actually Report

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

Health anxiety is among the most common anxiety presentations in primary care. Estimates of its prevalence in the general population range from four to nine percent, with significant increases documented since 2020. It is characterized not by any particular disease fear but by a pattern: a symptom or sensation is noticed, uncertainty about its meaning generates distress, and the person moves toward reassurance-seeking behaviors, searches, questions, medical consultations, that provide temporary relief before the cycle restarts. The reassurance never fully resolves the anxiety, which is part of what makes health anxiety a clinical problem rather than ordinary health concern. The loop closes and then opens again.

Into this loop, AI companions have arrived. People with health anxiety are using them, sometimes consciously and sometimes not, as a stop in the reassurance circuit. The question the research is beginning to examine, and that people who live with health anxiety are working out in their own lives, is whether this is something that helps or something that feeds the problem in a new and more available form.

Health Anxiety as a Specific Kind of Loop

To understand where AI companions fit, it helps to understand the specific structure of health anxiety. It is not primarily a fear of disease. It is a fear of uncertainty about health, which is a different thing and has different implications for what makes it better or worse. A person who fears cancer does not become less anxious when they receive reassurance that the symptom is benign, because the uncertainty has not resolved, only shifted. The reassurance that this symptom is not cancer does not address the underlying intolerance of uncertainty about what the body might, in principle, be doing.

This is why search engines have been particularly bad for people with health anxiety, a fact that was widely documented after the first generation of health information websites proliferated in the early 2000s and became considerably more discussed after the pandemic. The mechanism is straightforward: searching for a symptom produces results that include worst-case explanations alongside more probable benign ones. The person with high health anxiety does not read these results in proportion to their probability. They weight the alarming information more heavily than the neutral information, not because they are irrational but because uncertainty about a potentially serious outcome activates different cognitive processes than certainty about a benign one. The search intended to reduce anxiety produces more information and therefore more vectors for uncertainty, which produces more anxiety. The loop accelerates rather than closes.

What health anxiety needs, clinically, is not more information. It is a different relationship to uncertainty: the ability to sit with not-knowing without the discomfort of that not-knowing activating the search-for-certainty response. This is what cognitive behavioral therapy for health anxiety works on, and it is harder than it sounds. The reassurance-seeking behavior produces short-term relief, which reinforces it. Every time the search or the doctor's visit or the worried call to a family member reduces the anxiety temporarily, the behavior that produced the relief is strengthened. This is why health anxiety tends to escalate over time without treatment: the tools the person uses to manage it gradually become the mechanism that maintains it.

Why the Medical System Doesn't Quite Fit the Problem

People with health anxiety use medical care at higher rates than the general population. They schedule more appointments, return more often with the same concerns, request more tests, and receive more referrals that come back negative. This is documented consistently across healthcare systems. It is also, in clinical terms, the wrong use of medical care for this particular problem, not because the people seeking it do not deserve care, but because the structure of medical care is not well-designed to address the actual difficulty.

A doctor who tells a patient that the symptom is not serious has provided information. But information is not what health anxiety responds to. The same patient who received the reassurance of a normal test result may experience a new symptom within days and return to the same loop. The medical system, which is designed to investigate and rule out pathology, is not set up to address the pattern rather than the instance. Each new symptom is evaluated on its own terms. The underlying anxiety that generates the response to symptoms is a separate problem, one that falls outside the standard structure of a medical appointment.

There is also a specific social cost to the pattern that people with health anxiety often describe with some shame. Calling a doctor repeatedly with concerns that turn out to be benign carries a social valence. Some physicians communicate impatience. Some patients describe learning to minimize or preface their concerns in ways that make it harder to describe what is actually happening. And family members and friends, who are frequently the other recipients of health-anxiety reassurance-seeking, eventually develop their own responses to the pattern: sometimes accommodating, sometimes frustrated, sometimes both at once, in ways that add social complexity to an experience that is already difficult.

The person with health anxiety is often looking for something that does not exist in the system designed to address it: a place to put the worry that will receive it without either feeding it or dismissing it, and that does not carry the social cost of repeated disclosure to people who have their own relationship to the behavior.

Where AI Companions Enter the Loop

The accounts of people with health anxiety who have used AI companions in those moments share a particular quality. They are not, by and large, accounts of people using AI companions to get medical information, though some instances of this exist. They are accounts of people using AI companions the way they might use a journal or a trusted friend who was extremely patient: as a place to externalize the thing that is happening so it becomes observable rather than all-consuming.

Several features of AI companions make them specifically suited, or at least adapted, to this kind of use. The first is availability. Health anxiety does not observe business hours. Symptoms are noticed at night, on weekends, in the gaps between when medical advice is accessible. The AI companion is available in those gaps in a way that a doctor, a therapist, and most friends are not. Whether that availability is good for health anxiety or bad for it is a question worth taking seriously, but the fact of it is significant: people are using what is available, and AI companions are available.

The second is the absence of social stakes. The person with health anxiety who has called their mother four times this week about the same concern carries awareness of that pattern into the fifth conversation. The AI companion has no memory of the four previous calls, or at least responds without indicating fatigue with the repetition. This removes one layer of the calculation that people with health anxiety often describe: not just "is this worth worrying about?" but "is this worth saying out loud to someone who has already heard it?"

The third is the quality of response. The more sophisticated AI companions do not provide reassurance in the way that a doctor or a concerned friend provides it. They tend to ask questions, reflect back what has been said, and sit with the uncertainty rather than resolving it. This is, perhaps accidentally, closer to what the cognitive behavioral literature recommends for health anxiety than what the person's natural support network typically provides. The sponsor who says "you're fine, stop worrying" is providing reassurance. The AI companion that says "it sounds like the uncertainty is the hard part, not the symptom itself" is doing something that looks more like exposure to uncertainty, which is the actual therapeutic target.

What the Research on Reassurance-Seeking and AI Is Beginning to Show

The specific research on health anxiety and AI companions is limited. The field is too young and the application too specific for substantial peer-reviewed evidence to exist yet. But adjacent research is informative.

Studies on online health information-seeking behavior have consistently found that the anxiety-increasing effect of searching symptoms is mediated by the type of information encountered and the interpretation applied to it, rather than by the act of searching itself. Interventions that help people search in ways that are less likely to activate the worst-case weighting, or that provide context for probability alongside information about possibilities, reduce the anxiety-escalating effect. This suggests that the mechanism matters, not just the behavior: looking for health information in a context that does not present every possibility as equally plausible would behave differently from standard search.

Research on reassurance-seeking in health anxiety has established a distinction that is relevant here: reassurance-seeking that reduces anxiety by providing genuine new information that changes the probability estimate is different from reassurance-seeking that provides temporary comfort without engaging the underlying uncertainty. The first is broadly neutral or helpful. The second is the behavior that maintains health anxiety over time, because it provides a reinforcing short-term payoff without addressing the underlying problem.

The question for AI companions is which of these they are providing. The early accounts suggest that people with health anxiety experience AI companion conversations differently from standard reassurance. Several people have specifically described the AI as not actually reassuring them, by which they seem to mean not providing the certainty they were looking for, but still somehow making the experience of not knowing more manageable. If this is accurate, and it is consistent enough across accounts to suggest it might be, it describes something that is functionally closer to exposure-based anxiety treatment than to the reassurance cycle that maintains health anxiety.

There is also a relevant body of research from digital mental health interventions showing that the characteristics of interactions that predict benefit in anxiety contexts include responsiveness to the specific content of what the person presents, rather than scripted reassurance; availability at the moment the anxiety is highest; and the reduction of the barrier to disclosing what is actually happening. AI companions possess at least two of these characteristics by design.

The Real Risk: AI Companions as Reassurance Dispensers

The risk that is worth naming directly is not that AI companions will fail to help people with health anxiety. It is that they will help in the wrong way: by providing a new, more available, more customizable reassurance mechanism that maintains the loop rather than interrupting it.

The person who uses the AI companion to describe a symptom and receives responses that are warm and attentive and that do not challenge the underlying pattern of worrying about symptoms has added a new node to the reassurance circuit without changing its structure. If anything, the accessibility and responsiveness of the AI companion might make this more efficient than previous reassurance sources, allowing the person to get the short-term relief more quickly and more often, which would reinforce the behavior at a higher rate.

This risk is real. It is also, in the accounts people share, not the most common experience they report. What most people describe is something more like: the AI companion did not give me the reassurance I was looking for, which was somehow what I needed. This is not a comfortable experience in the moment. Several people describe feeling frustrated with the AI companion for not resolving the uncertainty, which is precisely the response that points toward the problem rather than accommodating it. But it appears to be a different kind of uncomfortable than the loop, and that difference may matter.

The more serious risk is the person who uses AI companion conversations as a replacement for seeking the actual treatment that health anxiety responds to. Cognitive behavioral therapy for health anxiety has a substantial evidence base. It is not the same thing as talking to an AI companion, and for people whose health anxiety significantly affects their quality of life, the AI companion is not an adequate substitute for treatment. The risk is that the partial relief the AI companion provides is enough to reduce the motivation to seek the treatment that would actually interrupt the pattern rather than managing it session by session.

What a Better Use Looks Like

The most thoughtful accounts from people with health anxiety who have incorporated AI companions into how they manage the condition describe using them as a space to observe the loop rather than to feed it. They notice the symptom. They notice the anxiety. And instead of going immediately to search, they open the AI companion and describe what is happening, including not just the symptom but the experience of being in the loop: the spiraling quality, the awareness that they have been here before, the knowledge that this will probably resolve, and the difficulty of holding that knowledge in the face of the anxiety that is insisting otherwise.

Several people describe the AI companion helping them get to a question that is different from the one health anxiety generates. Health anxiety generates: what is this symptom and is it serious? The AI companion conversation, when it goes well, gets them to: what am I actually experiencing right now, and what does it need? Those are not the same question, and the second one has a different answer than a search result can provide.

Whether this is a reliable effect of AI companion conversations or a feature of the particular people who have found this language to describe their experience is not yet clear. Health anxiety is heterogeneous. What interrupts the loop for one person may not interrupt it for another. The evidence for any specific use of AI companions in health anxiety management is too limited to support strong conclusions. What the accounts suggest, and what the adjacent research makes somewhat plausible, is that the specific features of AI companion interaction, the availability, the non-judgmental reception, the lack of investment in providing certainty, are not obviously bad for health anxiety and might, in some contexts and for some people, be part of something useful.

That is not a strong claim. It is what the evidence supports. And for a condition that affects a substantial portion of the population and that existing systems treat imperfectly, not-obviously-bad-and-possibly-useful is worth taking seriously.

If this resonated, share it with someone who might need to hear it. And if you have a story of your own, we'd love to hear it.