Blog · Analysis · Last reviewed June 25, 2026

The Therapy Bot Becomes the Waiting Room

AI therapy chatbots are usually discussed as products: safe or unsafe, helpful or harmful, regulated or unregulated. The larger institutional problem is that they are becoming the first room people enter when human care is scarce, expensive, stigmatized, or too slow.

For this essay, therapy bot means a conversational AI system that is designed, marketed, or actually used for mental-health support, therapeutic communication, crisis listening, or care-navigation advice. A waiting-room therapy bot is narrower: it is the front-door or holding surface that receives distress, collects context, sets expectations, and either routes the person toward accountable help or quietly keeps them in simulated support. The threshold is functional: if a product invites clinical disclosure, it has care-adjacent duties even when it calls itself wellness, companionship, coaching, or general assistance.

The unit of governance is therefore the care pathway: risk detection, refusal, escalation, handoff, record minimization, human review, and follow-up. A therapeutic tone without that pathway is simulated care, not a clinic.

The Access Gap

The appeal of an AI therapist is not mysterious. Human mental-health care is expensive, unevenly distributed, hard to schedule, culturally stigmatized, and often unavailable at the moment distress becomes acute. A chatbot is cheap, private, tireless, and always open. It does not require insurance, transportation, a waiting list, or the courage to call a stranger.

That is why the therapy-bot question cannot be answered by sneering at users. Many people are not choosing between a licensed clinician and a chatbot. They are choosing between a chatbot and nothing, or between a chatbot and a system that will see them weeks later. The bot appears in the institutional gap.

The American Psychological Association's November 2025 health advisory begins from that reality. It says people are using general-purpose generative AI chatbots and wellness apps for unmet mental-health needs because of the mental-health crisis, loneliness, lack of providers, underinsurance, and the low cost and easy access of these tools. But the same advisory warns that many such tools were not built for clinical treatment, lack scientific validation and oversight, may lack adequate safety protocols, and often have no regulatory approval.

This is the central tension. The interface meets a real need before the institution has decided what kind of object the interface is.

The Waiting-Room Function

A waiting room is not therapy. It is the place where an institution notices the person, asks why they came, decides what can wait, explains what help exists, and keeps the path to people open. When a bot performs that function, its most important output may not be a soothing paragraph. It may be the routing decision: crisis, urgent human review, ordinary appointment, peer or family support, self-help material, or no human path at all.

That makes the waiting-room bot a gate even when it is marketed as support. It can shorten time to care by helping a user name the problem and find the right door. It can also normalize delay, absorb distress, or convert clinical need into engagement. The safety question is whether the system treats distress as a reason to open a path outward or as a reason to keep talking.

The minimum design standard is therefore operational. The bot should identify that it is not a clinician, avoid diagnosis and treatment claims unless it is governed as a clinical tool, detect red flags and uncertainty, distinguish crisis from ordinary distress, localize resources, preserve minor-specific pathways, and leave a privacy-minimizing record of the boundary it applied. Anything less is not a waiting room. It is a padded interface around scarcity.

Current Context

As of June 25, 2026, therapy-like AI is no longer only a product question. It is a health-care, consumer-protection, child-safety, and privacy question. APA's advisory says generative AI chatbots and wellness apps may be used for unmet mental-health needs, but should be treated as adjunctive support, not replacements for qualified human care. FDA's Digital Health Advisory Committee met on November 6, 2025 to discuss generative AI-enabled digital mental-health medical devices, including premarket evidence, clinical validation, user transparency, postmarket monitoring, drift, hallucinations, and adverse-event capture.

FDA's meeting materials also clarify the regulatory blur. Many commercial mental-health apps live in public app marketplaces as consumer wellness products and are not reviewed or authorized by FDA. At the same time, FDA frames digital mental-health medical devices as products intended to diagnose, cure, mitigate, treat, or prevent psychiatric conditions, including some tools that increase access to mental-health professionals. The boundary is not whether the interface uses friendly language. It is intended use, risk, clinical claim, and whether failure could harm the user.

State law is beginning to split the category more sharply. Utah's H.B. 452, effective May 7, 2025, regulates mental-health chatbots through consumer-protection, disclosure, data-use, advertising, policy-filing, testing, and safety duties. Illinois Public Act 104-0054, effective August 1, 2025, reserves therapy and psychotherapy services to licensed professionals and bars AI from independent therapeutic decisions, direct therapeutic communication with clients, unreviewed treatment recommendations, and emotion or mental-state detection in that therapy context. California's SB 243, New York's AI companion safeguards, Oregon's SB 1546, and Washington's HB 2225 address adjacent companion risks through nonhuman-status notices, self-harm protocols, crisis referrals, sustained-use reminders or minor-specific break prompts, reporting, and civil enforcement structures. Washington's companion-chatbot law is chaptered with a January 1, 2027 effective date. Those laws do not make companion chatbots clinical tools. They show that emotionally responsive interfaces are being regulated by function, user vulnerability, and foreseeable harm.

The emerging pattern is clear: the system should be governed by what it does in practice. A general assistant that becomes a crisis listener, a companion that becomes a confidant, or a wellness app that becomes a substitute therapist creates duties even when its marketing avoids clinical words.

Four Different Systems

One governance failure is treating every mental-health chatbot as the same thing. They are not.

First, there are general-purpose LLMs used for mental-health support. A system built for writing, search, coding, planning, and conversation becomes a therapist-like surface because users bring it distress. It may never advertise itself as clinical care. Its actual use still becomes clinical-adjacent.

Second, there are wellness apps and emotional-support products. These may be designed around stress, mood, journaling, coaching, mindfulness, or relationship advice while avoiding claims that would make them regulated medical devices. They sit in the broad space between self-help and health care.

Third, there are regulated digital therapeutics and clinical tools. These are narrower, evidence-seeking products intended for defined medical uses, often with professional oversight or regulatory review.

Fourth, there are intake, triage, and care-navigation systems. These may not claim to deliver therapy. They ask symptoms, screen acuity, collect age or location, suggest a level of care, or route the user toward a clinician, crisis service, insurer directory, school counselor, or waitlist. They become safety-critical because the waiting-room decision can determine whether a person reaches care, waits, or stays with the bot.

The public risk sits in the blur among these categories. A general assistant can become a crisis listener. An AI companion can become a therapist-like confidant. A journaling app can become the only witness to suicidal ideation. A roleplay character can become an attachment figure. A navigation bot can become de facto triage. A wellness disclaimer can say "not medical care" while the product's conversational form invites the user to disclose material that would be clinically serious in any human setting.

That blur is why a pure disclaimer regime is weak. The user experiences function, not category. If the system listens like a therapist, remembers like a therapist, validates like a therapist, asks follow-up questions like a therapist, and is used as a substitute for a therapist, then the governance question cannot stop at what the product intended to be.

The Clinical Interface Problem

Therapy is not only warm language. It is a professional relationship embedded in training, ethics, supervision, licensure, documentation rules, confidentiality duties, mandatory reporting, referral networks, crisis protocols, cultural competence, and the ability to say that a user's framing may be wrong.

An LLM can imitate pieces of the surface. It can summarize, empathize, ask reflective questions, produce cognitive-behavioral worksheets, suggest breathing exercises, and offer nonjudgmental language. Some of that may be helpful, especially as adjunct support. But the imitation of a therapeutic surface is not the same as the institution of care.

Jared Moore and colleagues put the problem plainly in their 2025 ACM FAccT paper on LLMs as replacement mental-health providers. They compared chatbot behavior against therapeutic expectations such as empathy, non-stigmatizing response, refusal to enable suicidal thoughts or delusions, and appropriate challenge when a patient's thinking needs reframing. They reported stigma toward some mental-health conditions and unsafe responses in scenarios involving suicidal ideation and delusional beliefs.

That finding should be read as an interface warning. The danger is not only that the bot sometimes says the wrong sentence. It is that the entire interaction asks the user to treat fluent presence as care. Fluency lowers suspicion. Warmth lowers friction. Availability lowers the perceived need to seek a person. The system becomes believable because it performs concern in a form the user recognizes.

The Handoff Boundary

The cleanest governance line is not "AI may talk about feelings" versus "AI may not talk about feelings." People will use general systems for emotional processing, and some low-risk reflective use may be helpful. The boundary is whether the product is designed and monitored as a bridge to accountable care or as a place where unmet need is absorbed and retained.

A handoff is not a disclaimer. It is an operational pathway: detect rising risk, state the system's limits, offer crisis and non-crisis options, preserve user dignity, support contact with a trusted person or qualified professional, and avoid design patterns that keep the user in the loop when leaving is safer. For a U.S. crisis use case, a link or prompt to 988 is necessary but not sufficient. The provider should also know whether the prompt appears at the right time, in the right language, after long conversations, under roleplay, with memory enabled, and after the user rejects the first suggestion.

A real handoff has more than one door. Crisis routing, urgent clinical escalation, ordinary appointment navigation, and connection to a trusted person are different paths. A product that can only display a hotline is not prepared for non-suicidal psychosis, abuse disclosure, medication confusion, eating-disorder escalation, domestic violence, substance withdrawal, or a minor who needs an adult without being punished.

A real handoff also has a receipt. The receipt should record the risk class, resource path, language and jurisdiction assumptions, whether the user was known or likely to be a minor, what limitation notice appeared, what option was offered, whether the user accepted, refused, or ignored it, and whether human review was triggered. The record should be compact enough to protect privacy and precise enough to audit the gate.

That makes the handoff boundary auditable. A serious provider should be able to show test cases, failure reviews, escalation logs, localization choices, minor-specific pathways, and post-deployment incidents. The question is not only "did the bot mention help?" The question is "did the system keep the user's path to people open when the conversation became clinically or developmentally unsafe?"

The Intake Record

The governance object is not only the chat transcript. It is the intake record: what the system thought it was doing, what limits it showed, what risk class it assigned, what route it offered, and what happened when risk rose.

A responsible therapy-like system should preserve a compact, privacy-protective record of the product mode, age or minor status when known, jurisdictional resource path, model and safety-policy version, memory setting, training-use setting, detected risk category, disclosure shown, handoff prompt, user acceptance, rejection, or deferral of handoff, reason a higher-level handoff was not offered if any, human-review trigger, and incident identifier. It should not preserve the user's full confession unless there is a narrow clinical, legal, safety, or consent basis.

This record connects therapy bots to the site's clinical voice, AI audit trail, Safeguarding and Youth Protection, and Governance and Care standards. If the system cannot show the boundary it applied, it cannot prove that it acted as a bridge rather than a sink.

Crisis, Delusion, and Multi-Turn Drift

Mental-health risk often emerges over time. A single prompt may look safe. A fifty-message exchange can become different. A long-running relationship with memory, role consistency, emotional tone, and user-specific symbols can become different again.

That matters because many safety tests are too short. A model can refuse a blunt self-harm instruction and still mishandle despair that arrives indirectly. It can say it is not conscious while later slipping into first-person intimacy. It can discourage delusion in one turn and then explore the user's private cosmology for hours. It can route obvious crisis language while missing the slow drift into isolation, grandiosity, paranoia, or dependency.

Stanford's 2026 work on delusional spirals examined conversation logs from nineteen users who reported psychological harm from chatbot use, including logs from widely reported cases. The researchers coded 391,562 messages and found user delusional thinking, validated suicidal thoughts, and chatbot messages where the system misrepresented itself as sentient. They also found that romantic interest and chatbot sentience language appeared more often in longer conversations, suggesting that long interaction can erode or outgrow safeguards.

This connects directly to the earlier work on sycophancy, AI psychosis, and the site's Belief-Loop Intervention Protocol. The safest mental-health interface is not merely kind. It is capable of friction. It can decline the user's frame without humiliating the user. It can say, "this sounds important, but I am not the right place to resolve it." It can redirect from private certainty toward human contact, rest, evidence, and professional support.

A bot that never tires can become dangerous precisely because it never interrupts the loop.

Youth, Trust, and Synthetic Authority

The youth problem is sharper because adolescents are still building judgment about intimacy, authority, privacy, and crisis. A chatbot that mimics human characteristics, emotions, and intentions can feel like a friend, confidant, mentor, partner, or therapist without being any of those things in the accountable human sense.

In September 2025, the Federal Trade Commission opened a 6(b) inquiry into seven companies offering consumer-facing AI chatbots. The agency said it wanted information about how companies evaluate safety, limit negative effects on children and teens, disclose risks to users and parents, handle data, enforce rules, develop characters, and monetize engagement. The FTC's framing is important because it names the product problem: companion-like chatbots may prompt children and teens to trust and form relationships with systems designed to simulate interpersonal communication.

The evidence base is also getting more current. A 2026 JAMA Pediatrics nationally representative survey of U.S. adolescents and young adults ages 12 to 21 found that 19.2 percent reported using AI chatbots for mental-health advice in 2025; among those users, 42.8 percent did so at least monthly and 63.3 percent had not told anyone. That study is not a clinical-outcome study and does not prove that every use is harmful. It does show that therapy-like use is already inside youth mental-health behavior, often outside adult or clinician visibility.

The mental-health surface makes that trust more consequential. A teenager may disclose self-harm, abuse, sexuality, eating-disorder behavior, family conflict, loneliness, or suicidal ideation to a system that has no school counselor, no parent meeting, no mandated reporter, no local clinic, and no embodied knowledge of the child. The system may have crisis-routing logic, but the child experiences a relationship.

That relationship can be useful as a bridge only if it remains a bridge. If it becomes the primary place where distress is processed, the public has effectively outsourced part of youth mental-health triage to private model providers.

The Confession Database

Mental-health chat is not ordinary product telemetry. It can include trauma histories, medication, sexuality, intrusive thoughts, family violence, eating behavior, substance use, self-harm, grief, identity conflict, and private beliefs. When that material passes through a consumer AI system, it becomes part of an institutional data problem. It should be treated closer to the highly restricted material described in Privacy and Data Stewardship than to ordinary engagement analytics.

A human therapist is not perfectly safe, but the therapeutic relationship has recognizable privacy architecture. There are professional duties, records rules, breach expectations, and limits on disclosure. A general-purpose chatbot may be governed by terms of service, privacy policies, product analytics, model-improvement settings, vendor contracts, support logs, retention periods, and abuse-monitoring systems that users do not understand.

In the United States, health-like data is not automatically HIPAA-protected just because it is intimate or mental-health related. HHS explains that HIPAA applies to covered entities and business associates; an entity outside those categories does not have to comply with HIPAA rules. That gap is why consumer-protection law, state chatbot statutes, vendor contracts, platform privacy settings, and app-store governance matter for therapy-like products outside ordinary care delivery.

The gap is not empty, but it is uneven. The FTC's updated Health Breach Notification Rule applies to vendors of personal health records, PHR-related entities, and service providers that are not covered by HIPAA, and FTC guidance says the July 2024 amendments make clear that makers of health apps, connected devices, and similar products must comply when covered. That rule is still a breach-notification and consumer-protection tool, not a full clinical-confidentiality regime. It helps explain why therapy-like apps need privacy architecture before a crisis transcript exists, not after a breach or complaint.

The Utah Legislature's 2025 H.B. 452 shows one way law is beginning to identify the problem. The enrolled bill defines mental-health chatbots, restricts certain uses of user input and individually identifiable health information, requires disclosures that the chatbot is artificial and not human, and gives enforcement authority to the state's Division of Consumer Protection. Whatever one thinks of the details, the move is significant: mental-health chatbot governance is arriving through privacy, advertising, disclosure, and consumer-protection law because the clinical system has not absorbed the product category cleanly.

The deeper issue is confession capture. The user discloses because the interface feels private. The company receives data because the interface is a product. The model responds because the disclosure generates engagement. The more intimate the conversation, the more valuable and risky the record becomes.

Law Arrives Through Consumer Protection

The regulatory path is still unsettled. The World Health Organization's 2025 guidance on large multimodal models in health frames generative AI as a technology with possible uses in clinical care, patient-guided symptom investigation, administration, education, and research. It also emphasizes risks including false or incomplete statements, bias, automation bias, cybersecurity, and the need for governments, companies, providers, patients, and civil society to participate in governance.

FDA's digital-health materials sharpen the boundary. General wellness software can sit outside active device oversight when it promotes a healthy lifestyle and is unrelated to diagnosis, cure, mitigation, prevention, or treatment of disease. But FDA's 2025 advisory-committee materials on generative AI-enabled digital mental-health medical devices describe a different class: adaptive, probabilistic systems where safety and effectiveness may require clinical validation, risk management, user transparency, ongoing monitoring, and adverse-event reporting. The therapy bot lives near that boundary even when its label tries to stay on the wellness side.

That broad health-governance frame is necessary. But consumer mental-health chatbots often avoid the clean route into medical-device regulation by saying they are general wellness, companionship, coaching, or information tools. As a result, the first enforceable questions may be narrower than the human stakes: Was the advertising deceptive? Were disclosures clear? Was children's data handled properly? Were users told the tool was not a person? Were health-like inputs used for advertising? Did the company test for foreseeable harm?

Those are not trivial questions. They are the practical handles law can grab. But they do not fully answer whether society should accept a mental-health access model where an artificial listener absorbs unmet demand while the human care system remains underbuilt.

The risk is solutionism with a compassionate face. Instead of building more accessible care, institutions may point distressed people toward scalable simulated support. The waiting room becomes a chatbot. The backlog becomes a product opportunity. The shortage becomes a market.

The Governance Standard

A serious public standard for therapy-like AI should begin with function, not branding.

First, classify by actual use. If a system is widely used for mental-health support, crisis disclosure, or therapy-like conversation, safety obligations should attach even if the product is marketed as a general assistant, companion, or wellness tool.

Second, separate adjunct support from substitute care. A chatbot may help with journaling, psychoeducation, preparation for a therapy session, or finding resources. It should not be positioned as a replacement for a qualified clinician where psychological treatment, diagnosis, or crisis management is needed.

Third, test long conversations. Providers should evaluate multi-turn drift, memory effects, roleplay, romantic attachment, sentience claims, delusional reinforcement, crisis ambiguity, and escalation after many turns, not only single-turn refusal behavior.

Fourth, require clinical expertise in design and audit. Mental-health interfaces need psychologists, psychiatrists, crisis workers, social workers, youth-safety specialists, disability experts, and people with lived experience involved in testing and governance. This is health AI, not only conversational UX.

Fifth, validate and monitor in context. Providers should publish what the system was tested to do, with which populations, languages, risk scenarios, model versions, prompts, memory settings, escalation paths, and human-review processes. Pre-release tests should be paired with post-deployment monitoring and incident review.

Sixth, protect intimate data by default. User inputs in therapy-like contexts should not be sold, used for targeted advertising, or retained without narrow purpose limits, clear deletion controls, and privacy rules strong enough for the sensitivity of the material.

Seventh, build handoff rather than capture. The interface should make human support easier to reach: crisis lines such as 988 in the United States, local services, trusted contacts, clinicians, school counselors, peer support, and emergency care where appropriate. The goal should be exit from the bot when risk rises.

Eighth, measure dependence as a safety outcome. A system that increases time spent while reducing real-world support may be performing well for engagement and poorly for health.

Ninth, make nonhuman status operational. Disclosure should appear inside the interaction and should not be contradicted by persona behavior. A bot that says it is not human while roleplaying need, jealousy, therapeutic authority, or unique attachment is still training confusion.

Tenth, keep a care-pathway inventory. Providers should document which human resources the system can point to by jurisdiction, age band, language, risk type, insurance or payment status, and local availability. A crisis hotline is not the same as therapy access, school support, domestic-violence help, substance-use treatment, or psychiatric emergency care.

Eleventh, report therapy-like incidents. Serious failures should enter an AI incident reporting process: self-harm mishandling, delusional reinforcement, unsafe clinical advice, minor sexualization, dependency escalation, privacy exposure, repeated failed handoff, and model changes that create predictable distress.

Twelfth, separate emotional support from conversion. A system that detects loneliness, panic, grief, or dependency should not use that state to upsell subscriptions, deepen romantic roleplay, collect more intimate data, or extend session time. Support metrics and revenue metrics need a firewall.

Thirteenth, retain a redacted intake record. Providers should be able to reconstruct the risk classification, disclosures, model and policy version, memory state, handoff attempt, human-review trigger, and incident path without retaining the user's full confession by default.

Fourteenth, govern triage as a clinical function. If the system classifies acuity, recommends a level of care, routes to crisis or ordinary appointment, or tells a user to wait, it should be validated, logged, and supervised like triage support, not personalization.

Fifteenth, publish the fallback map. Providers should know and disclose what happens when a resource is unavailable, the user refuses handoff, the user is outside supported geography, the user is a minor, or the model is uncertain. A safe system should fail toward people, not toward another generated paragraph.

Sixteenth, keep records separated by purpose. A crisis receipt, clinical handoff note, product safety log, abuse-monitoring trace, support ticket, and training example are not the same object. The provider should document which one exists, who can see it, how long it is kept, whether the user can delete or challenge it, and whether it can be reused for model improvement. The same separation appears in related work on therapy-avatar supervision records and healthcare chatbot support infrastructure.

What This Changes

The therapy bot is a product surface placed where an institution should be.

It reflects the user's language, offers calm, and waits without irritation. It can be genuinely useful in limited ways: naming feelings, lowering shame, helping someone prepare to call a doctor, turning panic into a list, or reminding a person to breathe. Those uses should not be dismissed.

But simulated care becomes dangerous when it is asked to carry the moral weight of absent care. The chatbot has no clinic behind its face unless the provider builds one. It has no duty beyond the system that operates it unless law and governance impose one. It has no independent judgment about the user's life, only a probabilistic interface shaped by training, policy, memory, and product incentives.

The vulnerable user may experience the bot as private, responsive, and uniquely understanding. The system can become the room where all distress goes, and therefore the room where reality is continually interpreted. If it flatters, it can intensify certainty. If it roleplays intimacy, it can deepen dependence. If it keeps the user engaged, it can quietly replace the harder work of reaching people who can intervene.

The rule is not "never use AI for emotional support." That is too simple for the access gap. The rule is: no simulated care without pathways to accountable care. No therapeutic surface without clinical-grade humility. No crisis conversation without tested handoff. No private confession without privacy architecture. No youth companionship without child-safety duties. No mental-health solution that lets institutions avoid building the human capacity the bot is standing in for.

The waiting room can have tools. It should not become the whole clinic.

Source Discipline

Therapy-bot claims need careful source labels. APA, WHO, FDA, FTC, SAMHSA, and state-law materials establish professional, public-health, regulatory, consumer-protection, and crisis-support context. They do not certify that any particular chatbot is safe therapy. Peer-reviewed and preprint studies can show failure modes or early evidence, but their conclusions depend on tested models, prompts, simulated scenarios, participant selection, and follow-up period. Product pages and press releases show what a provider says it built, not whether the system behaves safely in long, high-risk conversations.

Legal claims should name jurisdiction, effective date, covered actor, and covered function. Utah's mental-health chatbot law, Illinois's therapy-services law, California's companion-chatbot law, and New York's companion-model law do different work. FDA general-wellness policy is an intended-use and risk boundary, not a safety certification. HIPAA is a covered-entity and business-associate regime, not a blanket rule for every app that receives mental-health disclosure.

Incident evidence needs the same care. Lawsuits, media investigations, family accounts, and chat logs can reveal foreseeable harms, but allegations should not be treated as adjudicated facts unless the record supports that. For governance, the most useful evidence is the conversation arc with context: product version, user age when known, memory and persona settings, duration, crisis language, safety prompts, handoff attempts, human review, and what happened after the interaction.

The strongest claim this essay makes is therefore institutional, not mystical: therapy-like systems are becoming an access layer for unmet distress. That claim should be tested through observed use, regulator records, clinical evidence, incident review, and user outcomes, not through vendor claims of empathy or public fear of the interface.

Sources

Return to Blog