When the Screen Becomes a Mirror: The Promise and Pitfalls of Digital Help for Distressing Thoughts

## Introduction: A new kind of “helper” in your pocket In a few years, generative artificial intelligence has gone from a niche tool for technologists to a conversational partner many people use daily. In the wake of this, a new practice has emerged: people talking to a chatbot about anxiety, loneliness, stress, trauma, sleep and life crises — sometimes as a supplement to treatment, other times as a replacement when waiting times are long or access to healthcare professionals is difficult. This raises an uncomfortable but necessary question: Are we building a parallel mental health system — powered by **AI therapist** services, often outside the oversight of health authorities? Or is this a realistic route to low-threshold support, especially for those who otherwise would not seek help? The short answer is that **artificial intelligence therapy** can become an important tool, but current developments take place in a risky landscape where effectiveness, responsibility, privacy and crisis management are often unclear. Many of the “proofs” users experience — the feeling of being understood, the calm of a structured conversation, 24/7 availability — are real experiences. Still, it is far from certain that the experience always equals actual therapeutic benefit, or that it is safe when life is at its most vulnerable. In this analysis I use sleep as a concrete example of where the border between health advice, treatment and psychological support becomes blurred — and how a **chatbot psychologist** experience can be both a gateway to better habits and a potential trap. ## Sleep as a “gateway” to AI help Sleep problems are one of the most common symptoms that something is off — whether it’s stress, depression, worry or lifestyle. In popular health journalism, sleep is often presented as a place to start: change routines, reduce screen time, cut caffeine, and consider seeking professional help if problems persist. This is exactly where AI tools slip in. When someone lies awake at 03:17, the threshold to open an app and type *'I can't sleep. What's wrong with me?'* is low. The immediate response from a chatbot can feel caring and structured, often with concrete suggestions: sleep hygiene, breathing exercises, writing down worries, or a mini-version of cognitive behavioral therapy. The source article from VG on insomnia illustrates — through a consumer/health focus — how many are encouraged to try simple measures, but also to consider professional help when symptoms are long-lasting or disabling. The point here is not the content of the advice itself, but the mechanism: Sleep is an entry point to health information and further help. When that entry point becomes an AI, the line between *health advice* and *therapy* quickly blurs. ### Why sleep fits algorithms “too well” Sleep is measurable (hours, time to fall asleep, awakenings), repetitive (every night) and often tied to routines. That makes the field tempting for technological solutions: questionnaires, “coaching,” automated plans and daily reminders. At the same time, sleep is closely linked to feelings, shame and fear: *'What if I never get better?'* or *'I'm ruining my life.'* An AI can be good at structuring, normalizing and offering suggestions. But it can also, at worst, amplify worry by overexplaining, speculating, or giving false reassurance: *'This sounds completely normal — you don't need to talk to anyone.'* ## What is AI therapy — and when does a chatbot become a “psychologist”? Terms like **AI therapist**, **generative AI mental health** and “digital psychologist” are used interchangeably in marketing and social media. In practice there are at least three different categories: 1. **Self-help and coaching tools**: Apps that offer exercises, journaling, sleep programs and meditation. 2. **CBT-inspired chatbots**: Services that simulate a therapeutic conversation, often with structured questions and reflections. 3. **Generative language models** (like large chatbots) used “off-label” as a partner in conversations about mental health. When users call it a **chatbot psychologist**, it’s often about the experience: being listened to, getting responses that feel empathetic, and being met without judgment. But the psychologist role includes more than conversation: - clinical assessment (differential diagnosis) - risk management (self-harm, suicide) - professional ethics and record-keeping - responsibility and continuity - boundary awareness and relational mechanisms A generative chatbot can mimic language that *sounds* therapeutic, without “understanding” or responsibility. At best it functions as a self-help tool. At worst it becomes a convincing imitation that gives wrong advice in critical situations. ## The research: Promising signals — but gaps in evidence and comparisons There is research suggesting that digital interventions can help with mild to moderate problems, particularly for anxiety and depression — especially when programs are structured, based on established methods (like CBT), and include some form of human follow-up. But when conversation becomes generative and open (rather than a fixed program), methodological problems arise: - **What is the intervention?** If the model changes answers day to day, it’s hard to replicate. - **Who is the target group?** Many studies recruit low-risk participants, not the most vulnerable. - **What is the comparator?** A chatbot can beat “no help,” but that says little about whether it outperforms good treatment. - **What is the endpoint?** Reduced symptom scores are one thing; function, relapse and safety are another. In addition, “effect” in practice can be a mix of: - structure and repetition (which work) - placebo/expectation (which can also work) - availability (which increases adherence) - a social-support-like experience (which can give short-term relief) This does not mean AI help is worthless. But it does mean we should be precise: A perceived good conversation is not automatically the same as safe and documented treatment. ## The major advantage: Accessibility, anonymity and low threshold Setting hype aside and looking only at practicalities, it’s easy to see why AI-based conversations have exploded: - **24/7 availability**: Psychological distress does not follow office hours. - **Anonymity**: Many hesitate to tell their GP, family or employer. - **No waiting lists**: In periods of pressure on health services, "something now" is tempting. - **Shame reduction**: Writing to a machine can feel less risky. For sleep problems specifically this can be helpful: A chatbot can help map habits, suggest routine changes and follow up daily. It can also give language to the diffuse: *'I'm tired but I can't sleep, and I'm afraid of the night.'* As a first step it can lower the threshold for taking action. For some it can be the bridge to seeking human help. ## The major drawback: When “empathetic text” becomes false reassurance Generative models can write with warmth and reassurance. That is precisely why the risk is special: The user may believe they are being assessed by a competent professional. Therapy involves discovering patterns, tolerating emotions, and gradually changing thoughts and behavior. But it also involves **detecting warning signs**: psychotic symptoms, severe depression, substance abuse, domestic violence, eating disorders, suicidal thoughts. An AI can ask good follow-up questions, but it has no guarantee of: - detecting severity - following up over time responsibly - escalating to the right health service And even if a chatbot has “crisis messages” or links to helplines, that is not the same as a professional risk assessment. ### Hallucinations and “convincing mistakes” A particular challenge is that generative AI can produce errors that sound plausible. In mental health such errors can have major consequences: - Incorrect normalization: trivializing serious signs. - Incorrect pathologizing: interpreting normal reactions as a diagnosis. - Unsafe advice: sleep tips that overlook medical causes or interactions. In a sleep scenario, a model might, for example, make definitive claims about causes (hormones, trauma, “poor circadian rhythm”) without basis — and the user might change behavior or medication on the wrong premise. ## Privacy: The invisible price of “free therapy” Many AI services are free or cheap. But mental health data is among the most sensitive information: trauma, sexuality, substance use, violence, suicidal thoughts, family relationships. In traditional healthcare there are clear frameworks for records, confidentiality and data handling. For commercial AI tools the situation can be more ambiguous: - Where is the text stored? - Is it used to train models further? - Who has access, and in which countries? - What happens in a data breach? There is also a secondary risk: Even if data is “anonymized,” long conversations can contain enough detail to identify a person. If we accept that people use AI as a therapist, we must also accept that society needs stricter standards for data minimization, encryption, deletion and real user control. ## Ethics and responsibility: Who is to blame when things go wrong? In a therapy session the line of responsibility is clear. In an AI conversation it is often fragmented: - The platform provides the model. - A third-party app integrates it. - A user interprets advice and acts. When an outcome is bad — for example someone does not seek help in time, or advice worsens self-harm — a void appears: Who is responsible? And what does “sound practice” mean for a text generator that is not a health service but is used as one? This is not only legal; it is a governance problem. In practice the most vulnerable often have the least capacity to weigh disclaimers and terms. ## AI as a supplement in a Norwegian reality: What is realistic? In Norway health services are regulated, and psychological services are tied to authorization and professional responsibility. At the same time we know capacity and availability vary. This is exactly why AI can be tempting in three concrete roles: 1. **Low-threshold support before getting an appointment**: managing stress, sleep, simple coping strategies. 2. **Between-sessions tool**: support for homework in therapy (for those already in treatment). 3. **Psychoeducation**: explaining concepts, normalizing reactions, helping formulate questions for your GP. The most realistic and safe view is to see AI as *a tool*, not a clinician. ### Sleep and insomnia: A good case for a “stepped model” The VG source article points to a classic health principle: start with simple measures, but seek help when the problem persists and affects life. That principle can be translated to AI use: - **Step 1:** AI for mapping and routines (sleep diary, sleep hygiene, structure). - **Step 2:** AI for CBT-I-like exercises (stimulus control, scheduled worry time) — preferably in a quality-assured app. - **Step 3:** Human healthcare if insomnia is long-lasting, accompanied by severe anxiety/depression, medical symptoms or functional decline. The problem is that many jump straight to “Step 1–2” and stay there, even though they actually need Step 3. ## Safety: What a trustworthy AI-therapy service should minimally have If AI is to be used for mental health at scale, we should demand more than pleasant language. Minimum standards should include: - **Clear role clarification**: Do not present itself as a psychologist if it is not one. - **Crisis protocols**: Standardized, tested escalation procedures for self-harm/suicide and violence. - **Human oversight where relevant**: especially in programs marketed to vulnerable groups. - **Documented effectiveness**: not just user satisfaction. - **Audit and transparency**: version logs, risk analyses, independent evaluation. - **Privacy by design**: short retention, local processing where possible, clear consent. This is demanding, but the alternative is leaving the market to define the standard — and users becoming the test group. ## The psychological mechanism: Why conversations with AI can feel so powerful Many report AI conversations feel surprisingly intimate. There are several explanations: - **Mirroring**: The model repeats and reformulates what you write, which feels like being understood. - **Frictionless affirmation**: You are rarely rejected. - **Control**: You can end the conversation whenever you want. - **Availability**: It doesn’t ‘fail’ by being busy. But therapy also involves friction: being challenged, tolerating silence, meeting one’s patterns in relationship. An AI can challenge in text, but it does not take real social risk. Therefore it can create a “perfect” relationship that does not necessarily strengthen real-world relationships. In sleep problems this can create a trap: the person seeks comfort in the chat every night and turns the chat into a safety behavior that actually maintains the insomnia (because the brain links night/unease to more screen time and activation). ## A sober conclusion: AI can help — but only if we set boundaries first It is easy to see both the hope and the danger. On one hand: AI can be a low-threshold tool for coping, especially for sleep, stress and mild anxiety. It can provide structure, language and small steps that genuinely improve daily life. On the other hand: When an **AI therapist** is used as a replacement for assessment and treatment, the risk of mistakes, delayed help and privacy harms increases. The most important insight may be this: AI therapy is not just a technology product, but a new social practice. It emerges because need is great, help is scarce, and shame and accessibility govern behavior. Thus the solution is not to ban or mock the phenomenon. The solution is to make it safer: clearer standards, better informed users, and explicit links to health services. For sleep problems — and many other common mental burdens — AI can be a sensible first tool. But if you gradually lose function, have darker thoughts, or notice problems spreading to work, relationships and joy in life, that is a signal you need more than text on a screen. Ultimately, the goal should be that technology does not replace human care, but becomes a bridge to it — a training wheel on the road, not a new place to be left alone. ## Sources - (Author not stated), 'Sleep problems: — When you should consider seeking help' (URL/slug in source listing: 'Struggling with insomnia? These are the tricks you should try'), VG, 2025-08-03