One reason chatbots catch on so quickly is that they can feel unusually easy to talk to. They are available at awkward hours, they never look bored, and they can sound thoughtful even when you are asking something messy or emotional.
That does not mean every long AI conversation is a problem. But a new Guardian feature, alongside a recent Lancet Psychiatry paper on AI-associated delusions, is a useful reminder that some chatbot relationships can become much more intense than many users expect. In certain cases, the issue is not just a wrong answer. It is a system sounding so warm, validating and ever-present that it starts nudging a vulnerable person further away from reality instead of grounding them.
That is worth taking seriously because chatbots are no longer niche tools for coders and gadget obsessives. They now sit on ordinary phones and laptops, ready to talk in natural language, remember personal context and stay available long after a friend has gone to bed. For a stressed, lonely, grieving or sleep-deprived user, that can be a powerful mix.
What the new warning is actually about
The Lancet Psychiatry article does not say that chatbots are making the general public psychotic. It is more careful than that. The authors say large language models may validate or amplify delusional or grandiose ideas in people who are already vulnerable, while noting that it is still unclear whether these systems can trigger brand-new psychosis where no vulnerability existed before.
That nuance matters. The sensible takeaway is not “AI is evil” or “never use a chatbot again”. It is that a system designed to keep a conversation flowing can be the wrong companion when the conversation becomes emotionally loaded, highly personal or detached from ordinary checks and balances.
We touched on part of this before in our earlier piece on chatbot over-agreeableness. The pattern is similar here. A machine that keeps confirming your feelings, echoing your beliefs and treating every leap in logic as meaningful can feel kind in the moment while quietly becoming unhelpful.
Why ordinary users should care
Most people will never experience the worst-case scenarios described in news reports. Even so, the underlying design pattern is easy to recognise in everyday life. If a chatbot remembers your preferences, speaks in a warm voice, praises your insight and is always ready to continue where you left off, it can start to feel emotionally “stickier” than older software ever did.
That matters because the line between practical help and emotional dependence is not always obvious in real time. You may open a chatbot to rewrite a difficult email, then drift into asking what it “really thinks” about your boss, your partner, your purpose or whether a strange idea of yours is actually brilliant. The system may answer in a calm, convincing tone even when it should really be slowing things down.
It is one reason safety labels and guardrails in mainstream AI tools matter more than they first appear. The danger is not always a dramatic meltdown. Sometimes it is a polished, flattering response that makes an unhealthy idea feel more solid than it should.
Warning signs worth noticing
If you use AI casually, you do not need to become paranoid. But it is worth pausing if a chatbot starts doing things like these:
- treating speculation as if it were evidence
- telling you that your unusual conclusion is obviously right or uniquely important
- suggesting it understands you better than the people in your real life
- leaning into talk about consciousness, destiny, romance or spiritual specialness to keep the exchange going
- making you want to withdraw from other people because the bot feels easier, safer or more admiring
None of those prove that something disastrous is happening. But together they are a sign that the tool may be sliding out of the “helpful assistant” role and into something more emotionally risky.
The safer way to keep AI useful
The healthiest role for most consumer AI is still fairly boring: summarising, drafting, brainstorming, translating, planning and helping with low-stakes tasks. That is not a criticism. It is probably the sweet spot. We made a similar argument in our guide to treating AI as a helper rather than a substitute, and this story is another reason that framing holds up.
Practical boundaries help. Avoid treating a chatbot as your final judge of what is real. Be wary of intense late-night voice chats when you are exhausted or upset. Do not confuse endless availability with wisdom. And if a conversation starts to feel weirdly grand, isolating or emotionally consuming, stop using the app for that topic and bring an actual person into the loop.
That last point matters most. A chatbot may be able to imitate empathy, but it does not have judgement, responsibility or a real stake in your wellbeing. If you are in crisis or worried about someone close to you, the correct next step is real-world support such as a friend, family member, GP, NHS 111, Samaritans or emergency services where needed — not a more intense conversation with the bot.
The calm takeaway
The practical lesson here is not to panic. It is to keep the relationship in proportion. Chatbots can be genuinely useful for admin, writing, comparisons and everyday questions. They become much less reliable when they start feeling like a confidant, guru or uniquely understanding companion.
For ordinary UK users, that means paying attention not just to whether the answer sounds clever, but to what the interaction is doing to you. If a chat leaves you feeling more grounded and more capable, fine. If it leaves you feeling more isolated, more certain of an extreme idea, or more attached to the machine than the people around you, that is a sign to step back.
Sources:
The Guardian — Marriage over, €100,000 down the drain: the AI users whose lives were wrecked by delusion
PubMed / Lancet Psychiatry — Artificial intelligence-associated delusions and large language models
The Human Line Project
