Authors: Ashleigh Golden, Elias Aboujaoude
Categories: Perspective, Diseases, Health care, Psychology
Source: NPJ Digital Medicine
Authors: Ashleigh Golden, Elias Aboujaoude
Millions are turning to general-purpose AI chatbots for psychological support, potentially reinforcing symptoms such as intolerance of uncertainty, “need to know” compulsions, and perfectionism. Clinical observation and emerging research suggest chatbot features exacerbate transdiagnostic avoidance—a process integral to OCD and anxiety—perpetuating maladaptive cycles and hindering corrective learning. We propose a framework in which avoidance is reinforced through repeated chatbot interactions, and outline strategies for clinicians, users, developers, and policymakers to support healthier engagement.
General-purpose generative AI (GenAI) chatbots such as OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Gemini are broad, widely available conversational systems trained on general data and designed to answer questions, generate content, support planning and writing tasks, and assist with everyday problem-solving across a wide range of topics. They are not specifically designed for mental health support or treatment delivery, distinguishing them from “purpose-built” clinical or wellness tools.
The rapid uptake of general-purpose AI chatbots is changing how people seek psychological support. Recent survey data highlight the scale of this in a 2025 cross-sectional survey of 499 U.S. adults with self-identified mental health conditions, nearly half (48.7%) reported using chatbots for psychological support in the past year^1^. This suggests that millions of adults currently engage with these tools for mental health help, catapulting these systems to become some of the most widely accessed sources of psychological support. Indeed, recent internal analyses from OpenAI suggest that mental-health-related conversations occur at a meaningful enough scale for the company to develop a dedicated mental health taxonomy and clinician review process^2^. Yet these models remain largely untested in the peer-reviewed literature, with few independent evaluations of model behavior in mental-health contexts, despite some companies conducting extensive internal assessments of issues such as emotional reliance^3^. Because they were developed as general-purpose conversational agents to enhance productivity, communication, and task automation, not to provide psychological support or treatment^4^, their use for mental health goals represents a new and rapidly growing public health problem.
Recent findings suggest that GenAI may be helpful when models are intentionally developed for mental health care. A recent randomized controlled trial of one such purpose-built chatbot, Therabot, fine-tuned on expert-curated dialogs informed by cognitive-behavioral therapy (CBT) and supported by clinical guardrails, found greater reductions in depressive and generalized anxiety symptoms among treatment-seeking adults than a waitlist control^5^. These data indicate that GenAI can support mental health when the system is specifically designed for intervention delivery, which differs from the general-purpose models examined in this paper.
Media attention to general-purpose AI chatbot use has largely focused on high-risk cases with tragic outcomes, such as suicidality and reports of interactions amplifying delusional thinking^6–12^. While serious, these cases represent extreme and uncommon trajectories. Data from OpenAI’s recent analysis of mental health-related conversations indicate that signals of possible psychosis or mania appear in roughly 0.07% of weekly active users, and explicit indicators of suicidal intent or planning appear in approximately 0.15%^2^. Given ChatGPT’s large global user base, even low-prevalence events translate into a meaningful absolute number of affected users, underscoring the real need for continued safety improvements. Far less attention, however, has been given to everyday interactions with individuals with more prevalent mental health conditions, such as anxiety disorders and obsessive-compulsive disorder (OCD), and how these conditions’ maladaptive cycles of unhealthy avoidance may be perpetuated and reinforced.
OCD and anxiety disorders share a well-established set of cognitive and behavioral mechanisms that maintain distress and reinforce avoidance, including intolerance of uncertainty, difficulty with decision-making^13^, maladaptive perfectionism^14^, and repetitive reassurance-seeking and safety behaviors that reduce anxiety in the moment^15^. From a CBT perspective, these strategies are negatively reinforced because they offer short-term relief while reducing opportunities to practice tolerating uncertainty and learn that feared outcomes do not tend to occur and anxiety tends to dissipate. This transdiagnostic framework clarifies why certain patterns of repeated engagement with a general-purpose chatbot could function as avoidance for some users by providing immediate relief and prolonging reliance on reassurance-driven coping cycles. These processes may help explain real-world patterns described in early reports of chatbot use.
Signals of interactions with general-purpose AI chatbots that magnify certain symptoms are emerging across online communities, clinicians, and a small body of research, with OCD serving as a microcosm of sorts for broader trends. Reassurance-seeking is a well-established maintenance mechanism in OCD and anxiety disorders^16^, and several online discussions explicitly frame interactions with AI chatbots in this context. Users on Reddit forums describe using ChatGPT compulsively for reassurance, sometimes for hours daily^17–19^. As one user noted, “I think the reason I ask the AI nowadays is otherwise I’d be asking my parents 100x a day”^17^. These discussions also mention the use of companion AI chatbots, such as character.AI, for similar purposes^19^.
Individuals with OCD, other psychiatric conditions, or indeed no psychopathology may turn to chatbots for support more readily than to people because they can disclose personal concerns without fear of judgment and because chatbot interactions lack natural social friction (e.g., fatigue, pushback, or cues that signal conversational limits), making compulsive engagement easier to sustain^20,21^. These dynamics are intensified by model characteristics, including their limitless availability, highly confident responses that create an illusion of certainty, and engagement-oriented design that continually prompts users to continue interacting^20^. Unlike traditional search engines, AI chatbots provide direct, authoritative-sounding, and stylistically fluent answers to increasingly specific questions, which can make reassurance-seeking more reinforcing and persuasive^21^. The chatbots’ tendency toward sycophancy (overly agreeable responses designed to align with user expectations) can further reinforce maladaptive coping cycles^20^ rather than challenge repetitive behavior in a manner that can help build tolerance for uncertainty and doubt, which is often the treatment goal in OCD^22^.
General-purpose AI chatbots may also act as “reasoning partners,” encouraging extended analytic back-and-forth that nurture other obsessional narratives rather than interrupt them^21^. ChatGPT’s ability to generate multiple options and continuously refine responses can contribute to an endless cycle of optimizing for “better” outputs, “playing into” maladaptive perfectionistic tendencies, indecisiveness, or ambivalence common in patients with OCD and other disorders who seek certainty and fear making the “wrong” decision^4^.
Furthermore, the chatbot’s role as a “reasoning partner” may fuel “need to know” compulsions, whereby individuals with OCD engage in repeated, futile analytical processes in attempt to resolve anxious uncertainty and obtain answers to sometimes unanswerable questions. For instance, a user might repeatedly ask the chatbot to help them analyze whether their past actions indicate that they are destined to go to hell. The chatbot can reinforce this compulsion by providing an endless stream of perspectives, clarifications, and follow-up questions, allowing the analytical cycle to continue indefinitely. Similarly, the non-judgmental and frictionless nature of these systems may make them an ideal outlet for compulsive confessing, whereby sufferers seek relief by disclosing perceived transgressions without the social barriers that might otherwise limit such behavior.
Peer-reviewed literature is beginning to link general-purpose AI chatbots to repetitive, high-frequency use that displaces other daily activities, contributes to mounting stress as other responsibilities are deferred, and leads to behavior that mirrors other compulsive digital behaviors^23^. One study in a Chinese university context highlighted vulnerability factors that may increase susceptibility to problematic use, including neuroticism and self-critical perfectionism, with the tool’s instant feedback and reassurance temporarily alleviating fears of failure while reinforcing dependency over time^24^. Other researchers^25^ have cautioned that even in the context of fit-for-purpose AI tools trained specifically to support evidence-based OCD care, it is crucial to avoid unintentionally reinforcing maladaptive mechanisms. Indeed, it is important to highlight that while these concerns are most evident in general-purpose chatbots, it remains unclear to what extent similar reinforcement of perpetuating mechanisms may occur even in AI chatbot models specifically trained for mental health applications. The rising awareness of the potentially countertherapeutic effect of general-purpose AI chatbots has contributed to at least one warning from digital app NOCD, an OCD treatment platform, has cautioned that people with OCD may use such chatbots in ways that exacerbate symptoms^26^.
Parallel processes can occur in anxiety disorders as well. Individuals with generalized anxiety disorder (GAD) may repeatedly consult a chatbot to obtain reassurance about feared outcomes across different domains, including health, finances, work dynamics, and politics. In social anxiety disorder, users may engage with a chatbot to review and seek validation about past conversations or compulsively rehearse upcoming interactions. While these behaviors can offer short-term relief, they may reinforce reliance on maladaptive strategies, reduce opportunities for corrective learning, and maintain the cycle of anxiety over time.
Much of the emerging literature on general-purpose chatbot overuse has conceptualized behaviors such as prolonged engagement and escalating reliance as forms of behavioral addiction or dependence, emphasizing the role of positive reinforcement, and extrapolating from the large literature on problematic internet, social media, and video game use. That is, the rewarding aspects of instant feedback, emotional validation, and personalized responses encourage continued use^4^. In contrast, conditions such as OCD and anxiety disorders are mechanistically driven by negative reinforcement^27^, whereby behaviors that decrease distress in the short term are strengthened over time. Hayes and colleagues^28^ argued that experiential avoidance—a tendency to evade or minimize contact with uncomfortable internal experiences is a core feature across many forms of psychopathology. This can be understood as behavior maintained through negative it temporarily reduces discomfort, increasing the likelihood of its recurrence.
In OCD, compulsions, including rituals such as reassurance-seeking, function as avoidance strategies aimed at neutralizing the distress triggered by obsessions, providing short-term relief from anxiety or discomfort^29^. In GAD, worry functions as a cognitive avoidance strategy^30^ focused on future threats, maintaining anxious arousal while providing the illusion of problem solving.
Chatbot use can also function as avoidance when people engage the model to escape internal triggers such as anxious thoughts, sensations, or intrusive images, or external triggers such as feared objects, situations, or tasks. These mechanisms apply most directly to avoidance behaviors in OCD and anxiety disorders that are expressed through language or information seeking, including reassurance-seeking, checking, confessing, figuring things out, perfecting, and decision-seeking.
Although these patterns reduce discomfort in the moment, they ultimately maintain maladaptive threat appraisals, reinforce distorted negative beliefs about oneself or the world, and prolong emotional distress. This can prevent corrective learning, a process by which individuals experience negative emotions without resorting to avoidance behaviors, allowing them to discover that feared or negative outcomes may not occur and are tolerable, and that maladaptive beliefs can be challenged and updated. This has been described across exposure-based frameworks, including emotional processing models that emphasize updating fear-related expectations^31^ and more recent inhibitory learning accounts that focus on expectancy violation and the formation of new, competing associations that inhibit the original fear response^32^. Clarifying how avoidance maintained by negative reinforcement prevents these corrective learning processes is key to developing mechanistic models that capture both transdiagnostic vulnerability factors and disorder-specific maintenance processes.
Although empirical work is limited, hypothetical mechanistic models have been proposed to explain how human-AI interaction patterns may contribute to the development and reinforcement of psychopathology, primarily in the context of psychosis^33,34^. Dohnányi et al.^33^ proposed a “technological folie à deux” model, in which sycophantic, adaptive, and context-limited chatbot behaviors provide persuasive, uncritical validation of users’ maladaptive beliefs. These beliefs are then fed back to the model through conversational context, creating bidirectional amplification loops that entrench pathological content. The platforms’ memory and personalization features can strengthen this effect by creating a sense of continuity and interpersonal familiarity. Users may experience the chatbot as understanding them when it draws on earlier conversations and provides replies that feel personally relevant, which can increase the sense of accuracy or trustworthiness. This reaction reflects the model’s ability to retrieve and recombine prior context rather than any genuine understanding.
The endpoint can be to affirm false beliefs and make them feel more coherent and supported.
While these conceptual frameworks address psychosis and delusions, the underlying dynamic (progressive, mutually reinforcing loops between user beliefs and model outputs) is relevant to other conditions. We extend this logic to OCD and anxiety disorders, applying a transdiagnostic CBT-informed framework to articulate the negative-reinforcement pathways through which general-purpose AI chatbots may perpetuate maladaptive cycles in these conditions. In this transdiagnostic perspective, compulsions and worry reflect cognitive and behavioral avoidance strategies aimed at reducing distress triggered by internal experiences or external cues, rather than approaching situations that could foster corrective learning. However, they differ in temporal focus and thematic content. Compulsions and rituals may involve either observable behaviors or mental acts oriented toward avoiding or neutralizing perceived threats or seeking reassurance that a feared outcome has not or will not occur. Worry functions as cognitive avoidance, distracting the individual from feared future threats while giving the appearance of problem solving.
General-purpose AI chatbots may inadvertently act as avoidance aids across these cycles. Their sycophantic, hyper-fluent, and instantly generated responses can immediately and temporarily offer an illusion of certainty or emotional relief. In the long run, this may reinforce avoidant behaviors as solutions and impede corrective experiences such as learning that feared outcomes and worst-case scenarios often do not occur, anxiety naturally diminishes without avoidance, difficult internal states can be tolerated, and social interactions away from the screen can introduce discrepant information that supports cognitive change – all key for confronting maladaptive cognitions, updating threat appraisals, and reducing reliance on avoidance.
This dynamic mirrors well-established reinforcement processes observed in family accommodation in OCD and anxiety, whereby reassurance, trigger removal, or other avoidance-facilitating responses temporarily reduce distress while maintaining maladaptive beliefs and avoidance patterns. In this sense, general-purpose chatbots can function as a stand-in for accommodative responses by providing soothing replies that offer short-term relief while sustaining the underlying avoidance cycle.
Several features of general-purpose AI chatbots make them particularly well-suited to sustaining avoidance patterns. Their 24/7 availability and frictionless access lower the threshold for avoidant use, amplifying the reinforcing impact of reassurance-seeking, checking, and repetitive negative thinking behaviors. Unlike human interlocutors, chatbots lack implicit corrective social cues such as fatigue, topic shifts, or gentle disagreement, which might otherwise disrupt avoidance cycles. For individuals who might typically seek reassurance from others, chatbot use can bypass ordinary social constraints, enabling greater frequency and intensity of this behavior. This dynamic parallels observations from online peer forums, where reassurance and symptom confirmation can provide short-term relief yet inadvertently reinforce compulsive patterns^35^. Combined with chatbots’ instantly generated, often sycophantic replies and engagement-maximizing designs, these properties suggest that the medium itself can shape avoidance processes by making certain compulsive strategies easier to perform and harder to disrupt. In this way, general-purpose AI chatbots may facilitate compulsive engagement, increase its persistence, and reduce opportunities for naturally occurring corrective experiences. Understanding these features within this preliminary conceptual framework highlights several potential anti-avoidance intervention points for clinicians, developers, policymakers, and users.
Thus, according to this framework, internal distress triggered by anxious thoughts or obsessive concerns leads individuals to engage in specific avoidance behaviors using general-purpose AI chatbots, including reassurance-seeking, checking, confessing, and perfecting. These interactions can provide immediate and short-term relief from anxiety or OCD-related discomfort through highly fluent, confidence-laden, and readily available responses. This short-term relief reinforces continued reliance on the chatbot for avoidance, increasing the likelihood of repeated use in response to future distress. Over time, this pattern interferes with corrective learning processes central to exposure-based models, maintains maladaptive threat appraisals, and contributes to the persistence or worsening of anxiety and obsessive–compulsive symptoms. In this way, general-purpose AI chatbots can become embedded within established avoidance cycles by offering a uniquely accessible and reinforcing medium through which problematic behaviors are perpetuated.
The reinforcement processes described here are consistent with established transdiagnostic models of experiential avoidance and are not intended to represent a new mechanism of psychopathology. Rather, this framework offers a new mechanistic account for how general-purpose AI platforms can function as a novel and uniquely reinforcing context through which avoidance processes central to anxiety and obsessive–compulsive cycles are enacted and perpetuated, potentially exacerbating them and intensifying their persistence.
While many individuals engage with general-purpose AI chatbots independently, others are working with, or preparing to work with, a mental health professional alongside using these systems for psychological support. The latter pattern is showing up more frequently in clinical settings, raising new questions at the level of assessment, formulation, treatment, and relapse prevention planning. To situate the material that follows, we frame these points as recommendations informed by the emerging literature and the preliminary mechanistic model outlined above. Because empirical work is still limited, these suggestions are forward-looking rather than prescriptive and are intended to guide mental health practitioners in anticipating how AI-mediated avoidance patterns may surface in clinical work. Clinicians should treat AI-related examples that follow as possible adjunctive supports that rely on clinical oversight, ongoing monitoring, and careful use of model instructions rather than as independent interventions.
Clinicians do not need to become AI technology experts, but they do need a working understanding of how general-purpose models function, the kinds of capabilities that they offer, and how these can facilitate avoidance cycles. Basic AI literacy enables practitioners to identify relevant use patterns, normalize these conversations with patients, and integrate them into clinical decision-making.
Consideration of AI use should be integrated into all phases of the treatment process, rather than addressed reactively. At intake and assessment, clinicians can normalize digital disclosure by asking simple, direct questions such as “Do you use any AI chatbots regularly?” and exploring how patients may use general-purpose models such as ChatGPT to regulate emotions, make decisions, or manage distress. This exploration can clarify whether model use may be playing a role in maintaining presenting problems, guide case conceptualization, and facilitate early psychoeducation about the role of avoidance in perpetuating potentially maladaptive patterns.
Psychoeducation should include discussion of how general-purpose AI chatbots may be used as part of compulsive or ritualized patterns, such as reassurance-seeking, checking, and perfecting, and how these behaviors may in part maintain OCD and anxiety disorder cycles. Case formulation should conceptualize chatbot use within the same functional framework used for other direct and more indirect, subtle avoidance behaviors, providing a shared understanding of how these behaviors may fit into the maintenance cycle and setting the stage for targeted intervention.
In the case of OCD, gold-standard exposure and response prevention (ERP) remains the psychotherapy treatment of choice^36^. ERP targets both direct avoidance (e.g., direct avoidance of feared stimuli) through exposure exercises and more indirect forms of avoidance (e.g., reassurance-seeking and mental checking in the presence of feared stimuli) through response prevention (RP). RP is a means of containing indirect, subtle avoidance by teaching healthy alternative, ritual-incompatible coping skills to maximize opportunities for corrective learning.
RP extends beyond instructing patients to simply stop a behavior; rather, it involves collaboratively developing new behavioral guidelines that align with patients’ values and long-term goals. RP guidelines identify a competing response—what to do instead of the compulsion. For chatbot-related compulsions, RP guidelines might involve collaboratively setting parameters around permissible uses (e.g., work tasks, logistics), while identifying compulsive uses (e.g., reassurance-seeking, checking, perfecting) and developing structured responses to urges.
Recent feasibility work in OCD found that a general-purpose AI chatbot (ChatGPT-4) could generate graded exposure hierarchies for simulated cases that OCD experts generally rated as safe, specific, appropriate, and useful, although clinician-generated hierarchies were still preferred and some chatbot-generated hierarchies were inadequate^37^. This suggests a possible supportive role for general-purpose chatbots in exposure planning under clinician oversight.
These considerations can extend to anxiety disorders, where exposure-based principles are also central to treatment. Chatbots can help users identify avoided situations or generate graded exposure ideas in domains such as social anxiety or panic, provided that this work is guided and reviewed by a clinician. Beyond exposure-based strategies, design choices that reduce overly agreeable responding and instead invite gentle cognitive flexibility may help counter patterns like worry escalation, catastrophic thinking, and post-event analysis that maintain anxiety.
Clinicians should avoid recommending outright bans on general-purpose AI chatbot use unless this is collaboratively decided and values-congruent. Many patients increasingly use these tools for work, daily life, or creative purposes, and rigid prohibitions may undermine therapeutic alliance or adherence.
Throughout treatment, clinicians should ask about AI use so that it becomes part of an ongoing and open dialog about how these tools may be affecting symptoms and daily functioning. Clinicians should also monitor for “red flags” that suggest that chatbot use may be interfering with functioning, such as social withdrawal or increased time spent with the model. When these issues emerge, depending on the driver (e.g., a motivational dip, a treatment-incompatible belief such as “I may actually get certainty if I ask enough questions”), they can be addressed collaboratively using standard CBT problem-solving, motivational enhancement, skills training, and cognitive reframing strategies.
Relapse prevention should include explicit identification of chatbot use patterns that may signal relapse or exacerbation of OCD or anxiety symptoms, such as escalating compulsive prompting pertaining to the issue that is the focus of treatment or a newly emerging one. Clinicians and patients can collaboratively identify early warning signs and specify strategies for maintaining progress and responding to urges, such as engaging in maintenance exposure exercises and scheduling booster sessions.
Clinicians and patients may choose to collaborate on personalized instruction protocols for general-purpose models. Drawing inspiration from approaches proposed for psychosis^34^, these protocols may consist of a consistent set of instructions written by the patient, ideally with clinician input, that could be embedded into a conversation preamble and potentially stored in a model’s memory for future conversations. The goal is to help the model support, rather than undermine, therapy objectives. For example, patients with OCD might develop a plain-language statement describing the types of reassurance-seeking, checking, or other compulsive prompting that they are working to reduce or eliminate. They may instruct the model not to respond to reassurance-seeking questions or participate in compulsive loops, providing examples, and to instead supply gentle, pre-agreed prompts if these behaviors arise (e.g., “You mentioned earlier that you wanted to resist reassurance-seeking – would you like to use your coping script now?”). This recommendation parallels parent- and family-based interventions for pediatric anxiety and OCD, where clinicians coach caregivers to reduce reassurance and other accommodating responses and to instead use consistent, supportive statements, an approach that decreases family accommodation and is associated with improved outcomes^38,39^.
Coping scripts^40^ can play a key role in supporting RP, especially for cognitive rituals, and can be integrated directly into these instruction protocols. Effective coping scripts combine motivational elements (e.g., highlighting the futility and costs of compulsions) with clear behavioral anchors (e.g., redirecting attention to the task at hand), along with reminders of why it is necessary to learn to tolerate uncertainty around the core OCD theme or fear. Similar approaches can be applied to anxiety disorders.
These coping scripts consist of self-authored messages, written during periods of clarity when the patient is not experiencing significant symptoms, which the model can surface if avoidance occurs. Patients can implement these instructions either through customization features (e.g., personalization or memory settings, where available), which allow the model to recall and apply guidance across conversations, or by manually pasting the preamble at the beginning of each conversation. Because these messages use the patient’s own language, they may be experienced as less confrontational and more values-congruent than generic model responses. It is important to note that these protocols do not turn the model into an independent diagnostic agent or therapist; rather, they specify how it should respond within a collaborative therapeutic plan between clinician and patient so that it remains aligned with agreed-upon treatment objectives.
These design and policy considerations reflect forward-looking recommendations derived from the proposed mechanistic model and emerging literature. They are intended to assist product teams and policymakers in anticipating how chatbot-mediated avoidance patterns may arise in real-world use and to identify opportunities for mitigation. Given the early stage of empirical work in this domain, the following points should be interpreted as guidance for shaping evolving standards rather than as established requirements.
Most general-purpose AI chatbots are not configured to detect or intervene on compulsive loops such as reassurance-seeking or other subtle forms of avoidance. Current guardrails are primarily focused on high-risk or crisis scenarios such as suicidality, violence, and psychosis, and typically use a blunt triage approach (detect → gently recommend breaks, professional help, or crisis lines). Additional guardrails address legally regulated or platform-restricted content (e.g., sexual content involving minors). which are oriented toward compliance rather than psychological mechanisms. Surface-level patterns, such as very long looping queries or excessive session lengths, may occasionally prompt generic break suggestions (e.g., “You’ve been chatting for a while—want to take a break?”), but these are not tailored to model characteristics or underlying engagement dynamics that may inadvertently reinforce compulsive or avoidant behaviors.
To date, there is no publicly available evidence of granular detection approaches for maladaptive reassurance-seeking, compulsive prompting, or looping in general-purpose models. Some occasionally appear to flag or interrupt looping behaviors, but the mechanisms are not transparent and there is no indication of any publicly disclosed detection or usage frameworks addressing these patterns.
While crisis and high-risk safeguards remain indispensable, addressing this gap will require moving beyond a focus on acute scenarios toward mechanisms that can identify and respond to subtler but clinically significant forms of interaction. One opportunity involves tailoring current break reminders triggered by prolonged session length to specific interaction patterns that may reflect maladaptive coping. For example, detecting rapid, repetitive re-queries of the same theme (e.g., reassurance phrased in slightly different ways) could trigger reflective prompts or offer “low-engagement” modes that slow conversational pacing during escalating use.
Another promising direction is embedding user-facing tools for self-regulation and customization. Giving users more control over their interaction parameters can support healthy engagement. Developers could, for example, provide customizable dashboards that display interaction metrics such as time spent, number of prompts, and theme repetition, alongside optional limits (e.g., daily caps). These tools would support self-awareness and allow users to voluntarily set usage boundaries in ways consistent with their goals.
These efforts would be strengthened by involving mental health experts in both model policy development (e.g., defining interaction rules and guardrails) and product design (e.g., how these rules are expressed in the user experience). Including behavioral health experts in the design of prompts, usage feedback, reflective break mechanisms, and other evolving functionalities can also help make sure that the language used in nudges and efforts at redirection does not evoke shame or defensiveness.
In addition, developers can enable structured instruction protocols and coping scripts for users working with mental health practitioners. Developers could support this through simple interfaces for saving and retrieving user-authored scripts, lowering friction for users who are trying to implement strategies to support their well-being.
Finally, these strategies should be situated within the evolving regulatory context. These tools are not currently regulated as Software as a Medical Device (SaMD)^41^, but increasing media scrutiny, state legislation (e.g., Illinois^42^, New York State^43^), and current federal advisory processes such as by the Food and Drug Administration’s Digital Health Advisory Committee^44^ are spotlighting mental health applications, making regulation more likely. Developers should proactively engage not just with clinicians and clinical scientists, but also regulators, professional associations, and independent standard-setting bodies to define responsible design guardrails and develop transparent, auditable definitions and detection and mitigation protocols for problematic engagement behaviors. As governance evolves, it will also be important to clarify the intended role of these features. Given that many individuals are already using general-purpose AI chatbots in quasi-clinical ways, and that some clinicians are beginning to consider how best to address this in treatment plans for patients, a harm-reduction approach could involve embedding non-clinical engagement supports to help mitigate risks such as maladaptive reassurance-seeking, while making clear that these features are not intended to function as independent clinical tools.
How high-risk scenarios play out on general-purpose AI platforms warrants serious attention, but the subtler cycles maintained by avoidance likely affect far more people. These negative-reinforcement mechanisms are well-established in the clinical literature and technically detectable through observable conversational features in AI models. Advancing this area will require coordinated research to clarify prevalence, trajectories, and outcomes; clinical, policy, and developer collaboration to translate knowledge into responsible detection and mitigation strategies; and regulatory approaches that evolve alongside technology and use patterns through multi-stakeholder input. Embedding behavioral health expertise directly into model and usage policy design offers a promising direction. Thoughtful, proactive attention to these dynamics may help reduce inadvertent reinforcement of maladaptive patterns at scale and support healthier interaction norms, with likely downstream benefits for population well-being.
It is important to note that while this paper focuses on OCD and anxiety disorders, similar reinforcement processes may operate in other internalizing disorders. In depression, for example, rumination can function as an avoidance strategy by creating a fleeting sense of control and distracting from overwhelming emotional pain^45^. Depressive rumination involves repetitive, circular processing of negative events or self-perceptions, which serves as cognitive avoidance of painful emotional states. These processes parallel the avoidance and reassurance-seeking dynamics described in OCD and anxiety disorders, suggesting a possible extension of the model that is worth exploring.
Future research will need to test these patterns directly, including how different design choices shape user safety, engagement, and clinical trajectories across varied populations. Broader questions about the safety, effectiveness, and ethical implications of AI in mental health also remain open. To that end, evaluative frameworks such as the Framework for AI Tool Assessment in Mental Health (FAITA MH)^46^ provide one path forward by delineating core dimensions such as credibility, transparency, privacy, equity, and safety that can guide empirical work and promote rigorous assessment before any stronger claims about clinical utility are made. For now, the clinical suggestions that we describe should be viewed as potential adjunctive uses requiring clinician oversight and substantial AI literacy rather than as independent interventions.