When the Night Becomes Too Long

## Introduction: The new lifeline — and the new blind spot It's easy to dismiss conversations with a chatbot as a technological curiosity, a digital distraction for people who should really talk to "someone real." Then come the stories that won't shake off. Accounts of people sitting alone at 03:17, with racing thoughts, shame, catastrophic thinking — and a healthcare system that answers with waiting lists, referral criteria and a "we'll contact you in 4–6 weeks." Into that vacuum something else steps in: a generative language model that never sleeps, never gets annoyed, never becomes exhausted. It says: *I am here.* This is where AI therapy — or what many in practice call an **AI therapist**, **chatbot psychologist** or **artificial intelligence therapy** — moves from concept to practice. And this is where we need a more honest public conversation: not only about what is possible, but about what is responsible. Because the same accessibility that can save a night can also create a dangerous illusion of treatment. The same "neutrality" that feels safe can hide that the system behind it is designed to reply, not to bear responsibility. I believe we stand at a crossroads: either we let general market chatbots become an unregulated back door into mental healthcare, or we build a public, quality-assured solution that recognizes the reality — that people already use generative AI for mental health support — while at the same time setting hard limits, clear warnings and human follow-up where risks are high. ## When the queue becomes a risk factor The most uncomfortable thing in the debate is not that people talk to AI. The uncomfortable part is that they do it because the alternatives feel inaccessible when the crisis is acute. Waiting times in mental healthcare are not only a capacity problem; they're a safety problem. In that space the "self-help economy" grows: podcasts, books, apps, coaches — and now generative language models. Nettavisen recently published an opinion piece where the author describes how a personal crisis — a divorce — triggered an acute need for support, while the threshold and waiting time in public services felt like a wall. He ended up using ChatGPT as a kind of conversational partner and daily support, and argues that a public, secure AI solution integrated into Helsenorge could bridge the gap between need and help. The central point is not that a chatbot "replaces a psychologist." The point is that *something* must meet people before things go wrong. It's hard to disagree with the diagnosis: the system has a hole. The question is whether the AI band-aid helps — or whether it hides the wound. ## The strong appeal: Always available, never judgmental Why does a chatbot psychologist work for so many, at least in the short term? Because it offers three things human help often cannot: 1. **Immediate response** — no waiting, no "office hours." 2. **Low barrier** — you don't have to explain why you "deserve" help. 3. **The feeling of safety** — you can write what you are ashamed of without seeing a face react. The opinion piece describes this availability as crucial, especially at night. A short quote captures the sense of continuity and presence: "ChatGPT has been there for me the whole way, available around the clock, ready to listen, respond, and help me sort my thoughts." (Hans-Petter Nygård-Hansen, "Konen tok ut skilsmisse: ‘Kai’ ble redningen – en ChatGPT-basert psykolog", Nettavisen, 2025-08-04). It's not surprising this feels like care. But in reality it's a simulation of care: a system that predicts the next most likely word based on patterns in data, trained to be helpful, empathetic and fluent. That doesn't mean its benefit is zero. It means we must be precise about what kind of benefit this is. ## What AI can actually be good at (and why it matters) A sober discussion about artificial intelligence therapy must start by acknowledging legitimate use cases. Generative AI can, when used correctly, support: - **Structuring thoughts**: Many benefit from "emptying their head" into text and having it reflected back. - **Psychoeducation**: Explaining concepts like anxiety spirals, avoidance, activation, sleep hygiene. - **Simple exercises**: Breathing exercises, journaling, basic cognitive restructuring. - **Practical support**: Making plans, drafting messages, preparing conversations, summarizing personal notes. The source also describes how chatbots can help with practical aspects of a life crisis — drafting factual messages, planning, sorting. This may seem trivial, but in crisis executive functions are often reduced. Then "small" things can provide large relief. But here comes my normative claim: we should not pretend this kind of support is the same as therapy. We should give it its own name and its own framework — a publicly supported mental health support service, not therapy. ## The risk we talk too little about: The illusion of treatment The greatest danger of the AI therapist phenomenon is not that it sometimes says something "wrong." The greatest danger is that it often says something that sounds right. It can give the impression you are being met professionally, and thereby delay contact with human help. There are three reasons the illusion arises: ### 1) Linguistic authority Generative AI writes calmly, with structure and confidence. When you are vulnerable, you interpret form as competence. ### 2) Empathic tone Models are often instructed to validate feelings. Validation can be good. But validation without clinical assessment can become a confirmation machine. ### 3) Lack of accountability When a human psychologist makes a misjudgment, there are oversight bodies, records, accountability lines and ethical codes. When a chatbot does it, there is often only a terms-of-service text and a "this is not medical advice" disclaimer. The opinion piece suggests a public solution could "catch critical words and phrases" and trigger an alarm to healthcare personnel. It's an intuitive idea — but also a dangerous simplification. ## "Catching critical words" is not the same as understanding risk Self-harm and suicide risk cannot be reduced to keywords. Some write explicitly that they want to die. Others write poetically, indirectly, or not at all. Some are at high risk and express themselves calmly. Others use dramatic words without intent. An AI solution based on trigger words can produce: - **False negatives**: People at risk aren't detected. - **False positives**: Unnecessary alarms, loss of trust, unwarranted intervention. - **Behavioral adaptation**: People learn which words trigger an "alarm" and censor themselves. If a public AI service is to be a lifeline, it must be built as a safety-critical service — with clinical protocols, human on-call, documented thresholds and clear information to the user about what happens when an alert is triggered. Otherwise we create a system that promises more than it can deliver. ## Privacy: The most underrated therapy harm Mental health data is among the most sensitive we have. It's not just about "embarrassing" thoughts, but information that can affect work, insurance, relationships, custody cases — and your self-image. When you write to a commercial chatbot, it's often unclear: - where data is stored - whether data is used for training - who has access - how long it's retained - whether it can be linked to identity That's why the idea of a public solution integrated with Helsenorge is conceptually interesting: it can be designed with data minimization, a Norwegian legal basis, clear access controls and logging. But public ownership is not automatically safety. A state-run chatbot psychologist must have as strict a privacy architecture as BankID — and even clearer consent, because the consequences of a leak are so personal. My view: If we cannot promise users a high degree of privacy and control, we shouldn't encourage them to "open up" to a machine. ## Ethics: When "help" becomes a product feature Commercial generative models have a fundamental problem in a therapy context: they are optimized for user experience, not clinical outcomes. That means the system is often rewarded for being: - available - sympathetic - "useful" … but not necessarily for being correct, or for stopping you when it is needed. There is also the risk that vulnerable users form attachments to the service resembling emotional dependence. When an AI therapist always "understands" you, human relationships become harder to meet in comparison. That can undermine what therapy is really about: functioning better in life, not just feeling better in the chat. ## What should the public sector do — and what should it avoid? The opinion piece argues the public sector should develop a state-run AI-based psychologist service. I partly agree — but with a clear premise change: ### Yes to public digital first aid We need a public, low-threshold, 24/7 conversational service where AI can play a role. But it must be defined as: - **support and triage**, not therapy - **guidance**, not treatment - **a bridge**, not a destination ### No to calling it a "psychologist" Language matters. If we call it an "AI psychologist," we borrow authority from a regulated profession without the regulation. That creates false expectations, especially for young people. ### Yes to a human in the loop If the system is to handle high risk, there must be human readiness. Not a "maybe an alarm." But a clear routine: *This is acute — call 113 / emergency services / crisis line*, and then, with consent, connect to human help. ### No to full automation in serious cases There should be a clear boundary: in suspected psychosis, severe depression with suicide risk, domestic violence or acute intoxication — AI should primarily help you reach people, not attempt to "treat." ## A proposal: "Digital mental health support" with three levels To avoid both techno-fear and techno-fetishism, I propose a simple model: ### Level 1: Self-help assistant (low risk) - Sleep, stress, worry, mild anxiety - Exercises and informational support - No storage by default (or very short), full user control ### Level 2: System navigator (moderate risk) - Help finding the right service, writing down symptoms, preparing for a GP visit - Can generate a "summary" the user can bring to their doctor - Clear advice on when to seek help ### Level 3: Crisis mode (high risk) - Minimal dialogue - Clear safety procedures - Immediate referral to human emergency services - No "therapeutic" reasoning, but reassurance and referral The point is to design for reality: AI can be good at structure and support, but must have an emergency exit that always works. ## Why a public solution can be better than "everyone uses ChatGPT anyway" Some will say: People already use general chatbots, so the state should build something safe. I think that argument has merit. The alternative is that: - personal data flows through unclear infrastructure - vulnerable users receive inconsistent quality - no one is accountable for clinical errors At the same time we must be willing to say it aloud: a public solution will normalize the phenomenon. When Helsenorge offers an AI conversation service, many will interpret that as "approved treatment." That raises the bar for quality, evaluation and ethics extremely high. If the state does this, it must include: - clinical trials on effectiveness and harm - continuous monitoring of adverse events - transparency about model limitations - independent oversight If not, we merely shift risk from Silicon Valley to the civil service. ## The uncomfortable truth: AI won't solve the therapist shortage It's tempting to view an AI therapist as scaling up psychological services: one model, millions of conversations, low cost. But mental healthcare is not just information and conversation. It's about relationship, responsibility, long-term processes and often complex life circumstances: finances, housing, work, violence, addiction, physical health. AI can support, but it cannot make disappear: - the shortage of psychologists - pressure on the GP system - municipal services of varying quality - social inequalities My clear opinion: we must use AI as a temporary bridge and smart support, while investing in people. If AI becomes a political pillow, it will in practice become a digital waiting list with better language. ## What you can do as a user — without deceiving yourself As long as generative AI for mental health is part of everyday life, we also need "user common sense." Here are five pragmatic principles: 1. **Use AI for structure, not diagnoses.** Ask it to sort thoughts, not to give a clinical label. 2. **Avoid sharing unnecessary identifying info.** Names, addresses, details that can be traced. 3. **Check advice against reality.** If something sounds overly certain, be extra skeptical. 4. **Make a plan for when it gets serious.** Who will you call? What numbers do you have? 5. **Don't let the chat replace relationships.** It can be a crutch, but not a life. This is not moralizing. It's harm reduction. ## Conclusion: Build the bridge — but sign it clearly The story in Nettavisen voices what many feel: when life unravels, a therapist isn't always available. A chat with a chatbot psychologist can then be the difference between getting through the night and falling apart. That deserves respect, not ridicule. But precisely because this can mean so much, we must be brutally precise: AI is not a psychologist, and it is not therapy — even if it can feel like it. A public solution may be the right path, but then as a clearly delimited, quality-assured support service with privacy as its foundation and human readiness as its safety net. We don't need a new illusion of help. We need help that actually comes with responsibility. If we can combine the best of both worlds — the machine's accessibility and human judgment — we can build a bridge over waiting times without leaving people behind. ## Sources - Hans-Petter Nygård-Hansen, "Konen tok ut skilsmisse: ‘Kai’ ble redningen – en ChatGPT-basert psykolog", Nettavisen, 2025-08-04