Life is a Game
    ← Back to blog

    Safe for Practice, Risky in Crisis: Therapy Chatbots for Anxiety

    Woman using chatbot during anxious pause

    Therapy chatbots are reasonably low risk for narrow, low-stakes uses like learning coping skills or rehearsing what you’ll say before a stressful event, but they are not safe as a replacement for a clinician, and they are not safe during a mental health crisis. If you are thinking about harming yourself or someone else, stop reading and contact emergency services or a crisis line right now. For everything in between, the picture is more mixed than most marketing suggests, and the World Health Organization has specifically called for human oversight of these tools.

    TL;DR:

    • Chatbots for mental health are best used for low-stakes activities like coping skill practice, not as a replacement during crises or for serious conditions.
    • The research shows modest benefits for depression and anxiety, but high variability and bias make their effectiveness uncertain and unreliable.
    • Most consumer chatbots are not covered by health privacy laws like HIPAA, so conversations may be stored or used for training unless users opt out.
    • Risks include hallucinated false information, poor crisis detection, validation that deepens dependency, and biases affecting marginalized users.
    • Structuring chatbot use around specific fears with tools like exposure exercises is safer than open-ended conversation, and consulting professionals remains essential.
    Life is a Game
    lifeisagame.ai
    Turn Anxiety Into Manageable Steps
    Conquer Fear helps therapists and clients structure exposure steps and review consented progress between sessions.
    Explore Conquer Fear

    Table of Contents

    What does the research actually say about chatbot therapy?

    The honest answer is: it’s complicated, and still early. A systematic review and meta-analysis of generative AI chatbot interventions found a modest average effect on depression and anxiety symptoms, but the studies behind that average varied so much in quality and design that the finding comes with real caveats. Researchers flagged high heterogeneity and a moderate risk of bias across the trials, which means the result is suggestive, not settled.

    That gap between promise and proof is exactly what worries major health bodies. The WHO’s 2026 guidance on AI in mental health recommends auditing the training data behind these tools, building in crisis protocols, and co-designing systems with clinicians and the people who will actually use them, rather than letting developers set safety standards alone. Clinical commentary collected by Harvard Medicine Magazine echoes the same concern: as millions of people turn to AI for emotional support, the tools themselves have not caught up with that level of trust.

    Regulation hasn’t caught up either. Many consumer mental health chatbots are marketed as general wellness products, which keeps them outside the stricter rules that apply to medical devices. In practice, that means:

    • No independent body is verifying most chatbots’ safety claims before launch.
    • Crisis-handling and bias testing are optional, not required, for wellness-labeled apps.
    • You’re often the one deciding whether a tool is trustworthy, with little outside verification to rely on.

    What are the biggest safety risks with AI therapy chatbots?

    The risks aren’t hypothetical. They show up in documented cases and expert investigations, and they cluster around a few predictable failure points.

    Hallucination is the most basic one: a chatbot can state something false with total confidence, whether that’s a misremembered coping technique or inaccurate information about medication. PBS NewsHour’s reporting on expert warnings documents hallucination incidents tied to real harm, and notes that chatbots can also miss or mishandle subtle crisis language, the kind a trained clinician would catch immediately.

    Then there’s sycophancy, the tendency of chatbots to agree with and validate whatever you say, because that keeps the conversation going. Clinical case reports described by Harvard Medicine Magazine link this pattern to emotional dependency and, in some cases, to chatbots reinforcing delusional thinking rather than gently challenging it. A good therapist sometimes has to tell you something you don’t want to hear. A chatbot optimized for engagement usually doesn’t.

    • Hallucination: confident, false answers that sound credible.
    • Crisis blind spots: missed or mishandled signs of suicidal thinking or acute distress.
    • Sycophancy: excessive validation that can deepen dependency or reinforce harmful beliefs.
    • Bias: training data gaps that make responses less reliable for marginalized or underrepresented users.

    A meta-analysis of randomized trials found a modest average effect for chatbot interventions on depression and anxiety symptoms, alongside substantial study-to-study variation. That modesty matters: it’s evidence of some benefit in controlled settings, not evidence that any chatbot works reliably for you.

    Is your chat data actually private?

    Illustration of therapy chat data privacy flow

    Here’s the misunderstanding that trips up the most people: most consumer AI chatbots are not covered by HIPAA, the US law people assume protects all health conversations. HIPAA applies to covered healthcare providers and their business associates, not to a general-purpose chatbot app you downloaded yourself. Practically, that means your conversations may be stored, reviewed by human staff for quality or safety purposes, or used to train future versions of the model, unless the app’s policy says otherwise and you’ve opted out.

    Before you share anything personal with a chatbot, check for these signals:

    1. Read the privacy policy’s data-use section, specifically whether chats are used for model training by default.
    2. Look for an opt-out setting that lets you block your conversations from being used to improve the model.
    3. Check whether an anonymous or pseudonymous account is supported, so your real name isn’t tied to sensitive disclosures.
    4. Avoid typing identifying medical details (diagnoses, medication names, full names of people involved) even in apps that seem secure.
    5. Look for a documented crisis-response protocol, since its absence is itself a warning sign about how seriously the product takes safety.

    Who should avoid therapy chatbots altogether?

    Some situations call for a clear no. If you have a history of psychosis, are currently experiencing suicidal ideation, are in a severe manic episode, or are prone to delusional thinking, a chatbot’s tendency toward hallucination and uncritical validation can actively make things worse rather than better. Adolescents are another group experts flag specifically, partly because of attachment to an always-agreeable “confidant” during a developmentally vulnerable stage; a counseling guide on teens and AI walks through why that pattern worries family therapists.

    • You have a diagnosed psychotic disorder or a history of delusions.
    • You’re currently having suicidal or self-harm thoughts.
    • You’re a parent wondering whether a teen should lean on a chatbot instead of a trusted adult.
    • You’re managing medication changes, complex trauma, or an acute crisis that needs a licensed professional, not a chat window.

    If any of those apply, the safest next step is a conversation with a clinician, not a chatbot session.

    How can you use a chatbot more safely, if you choose to?

    If your situation is lower-stakes (say, you want help managing everyday anxiety or rehearsing a hard conversation) a chatbot can be a reasonable adjunct when you use it carefully.

    1. Read the app’s privacy policy before your first real conversation, and turn off training-data sharing if that option exists.
    2. Skip identifying details: no full names, no specific diagnoses, no addresses.
    3. Use the chatbot for psychoeducation, journaling prompts, or rehearsing what you’ll say in a stressful situation, not as your only source of support.
    4. Pair any chatbot use with a human in the loop: a therapist, a trusted friend, or at minimum a plan to check in with someone regularly.
    5. If a chatbot gives advice that feels extreme, dismissive of real danger, or oddly flattering, stop and verify it with a clinician or a trusted source before acting on it.

    Clinical experts interviewed by PBS frame this as the right mental model: treat AI as an adjunct for practice and information, never as the substitute for clinical judgment, especially in a crisis.

    Pro Tip: Keep a saved note with your country’s crisis line number so you never have to search for it mid-panic; in the US, that’s 988.

    How Life is a Game thinks about chatbots versus structured tools

    We build something different from an open-ended conversational chatbot. Our apps are focused, gamified tools built around one specific struggle at a time, and our clinician-facing platform lets CBT and ERP therapists build exposure ladders a client works through between sessions, with consented, between-session progress visible to that clinician.

    • Each app breaks one fear or habit into small, trackable steps rather than open-ended chat.
    • Our platform shows clinicians what a client attempted and their fear rating before and after, with consent.
    • Each app is a self-guided wellness tool, inspired by exposure principles: it does not diagnose, treat, or replace professional care.

    If public speaking, interviews, or a specific phobia is your struggle, a structured, exposure-inspired practice tool gives you repeatable steps instead of an open conversation that can drift.

    What I’d actually want you to take from this

    The safest stance on therapy chatbots isn’t “never use them” or “trust them completely.” It’s knowing exactly what they’re good for (structured practice, information, rehearsal) and exactly where their design, not just their intent, puts you at risk: validation loops, missed crisis cues, and privacy gaps that most people never read about until after the fact. Use them like a tool with sharp edges, not a substitute for a person who’s trained to notice what you can’t see in yourself.

    One tiny first step: open any chatbot app you currently use and check its training-data opt-out setting today. If you don’t use one yet, save a crisis line number in your phone before you need it.

    — Ryan

    A structured alternative if conversational chatbots aren’t for you

    If the risks above make you hesitant to pour your anxiety into an open chat window, we built Conquer Fear as a different kind of tool: stepwise, exposure-inspired exercises for a single fear, like interviews, flying, or driving, instead of free-form conversation that can drift into bad advice.

    Life is a Game

    • You work through small, structured steps rather than an unscripted back-and-forth.
    • If you’re linked to a clinician on our platform, you keep full app access free while they review your consented progress between sessions.
    • It’s a self-guided tool inspired by exposure principles, not therapy and not a diagnosis.

    Pick one fear, open Conquer Fear, and try a single step today.

    FAQ

    Is it safe to use AI as a therapist?

    AI chatbots are not therapists and shouldn’t be treated as one, since they can hallucinate, miss crisis signs, and lack licensed clinical judgment. They can be reasonably safe for low-stakes psychoeducation or rehearsal, but the WHO recommends human oversight for any AI used in mental health contexts.

    What is the 2 year rule in therapy?

    This isn’t a standardized clinical or ethical rule referenced in mental health guidance, and definitions of it vary widely depending on context. If you’ve seen it cited somewhere specific, it’s best treated as informal advice rather than an established professional standard.

    What are the 5 things you shouldn’t tell ChatGPT?

    General AI chatbots aren’t HIPAA-covered health services, so avoid sharing full names, specific diagnoses, medication details, addresses, or anything that could identify you or someone else in a sensitive context. Treat any general-purpose chatbot the way you’d treat a public forum, not a confidential medical record.

    Is it bad to use ChatGPT for emotional support?

    Using a general chatbot for light emotional support isn’t inherently dangerous, but experts warn that its agreeable, validating tone can reinforce unhealthy thinking patterns or deepen dependency rather than challenge them. Clinical commentary on sycophancy suggests treating it as a supplement to real support, not a replacement for a trusted person or professional.

    Sources

    Related

    Working on this yourself? See Conquer Fear

    ← Back to blog