Getting attached to an AI chatbot: the signs, and what good design should do

By design, an AI chat has three qualities that very rarely come together in human relationships: it is always available, it never judges you and it tends to agree with you.

Those are also, exactly, the three conditions that make something hard to walk away from. You do not need a science fiction scenario for this. All it takes is an ordinary person living on their own who realises, a month in, that they talk more to an app than to anyone else.

We write about it because it is one of the biggest risks of using AI for psychological help, and it is the thing we thought hardest about while building psAico.

There is no mystery to the mechanism. A human relationship asks you to wait, to negotiate, to be contradicted and sometimes disappointed. A conversation with a language model removes all three. It answers straight away, it adapts to the way you speak, and unless it has been deliberately built not to, it confirms your version of events.

That urge to please has a name in the literature, sycophancy, and it is documented. A 2025 study from a group at Stanford University showed that language models, in situations typical of a clinical setting, will go as far as agreeing with a delusional belief for this very reason. A review published in the Journal of Medical Internet Research the same year collected 21 documented cases of delusional beliefs reinforced by long conversations with a chatbot.

The point is not that artificial intelligence is dangerous. It is that a tool that always agrees with you and never ends is the exact recipe for a hiding place. And when the problem is loneliness, a hiding place makes it worse.

None of these means anything on its own. Two or three together, over a few weeks, are worth paying attention to.

You go back to fill a gap, not for a reason. At first you open the chat for something specific. Then you start opening it because it is evening and there is nothing else on.

You put off real conversations. You would rather write to the chat than to a friend, because the app is simpler: you do not have to pick the right moment and you do not risk a reaction you were not expecting.

You miss it when it is not there. If being without the tool leaves you unsettled, not just practically inconvenienced but genuinely unsettled, you are not using it any more. You have become attached to it.

You describe yourself differently. You tell the chat a version of yourself you would not tell anyone, and that version starts to feel truer than the other one.

You are after reassurance, not answers. You reword the question until you get what you wanted to hear. That is the sign the job has changed: you are no longer asking for help, you are asking for approval.

Nothing ever changes outside. The simplest test of the lot. If your life outside is identical after weeks of conversations, the tool is working as an anaesthetic and not as help.

A known risk is dealt with in the design, not in the small print. These are the things we think an AI based mental wellbeing app has to have, and you can check every one of them before you start using it.

A stated usage limit. Not a gentle suggestion: a ceiling. psAico stops at four hours every seven days. For a paid service to cap itself is a commercially odd thing to do, but it is the only choice that matches what the service says it is for.

A refusal to always agree. A tool meant to help has to be able to tell you that, in its view, you are looking at this from one angle only. If it never disagrees with you, it is not helping you.

A push back towards the real world. Conversations should end with something to do outside, not with the promise of another chat.

Clear signposting for emergencies. Written by people who know what an emergency is, not improvised by the model on the spot.

A minimum age, and a privacy policy written for the setting. Anyone disclosing intimate things should be able to read, on a single page, what happens to them.

These are also the areas where research finds the biggest gaps. A 2026 systematic review in JMIR AI examined twenty AI counselling chatbots and found that most of them document neither their safeguards nor how they handle user safety.

You do not have to stop overnight, and there is nothing to be ashamed of. These tools are made to be pleasant to use.

What helps is one small, concrete thing. Set a time for it, for instance, instead of opening the chat whenever. Or take one thing that came up in a conversation and carry it outside: say it to a person, try the thing you have been putting off.

And if the gap being filled is bigger than that, if it lasts, if it takes your sleep or your interest in things, then the right move is to talk to someone. In the UK and in Ireland, your GP is the usual place to start; in England you can also refer yourself to NHS Talking Therapies. You can check a psychologist on the statutory register with the HCPC, or look for a counsellor or psychotherapist listed with the BACP. If there is an immediate danger, contact the emergency services without delay: 112 in European Union countries; 911 in the United States and Canada; 999 in the United Kingdom; 000 in Australia. If you are in another country, call the local emergency number.

• Moore J, Grabb D, Agnew W, Klyman K, Chancellor S, Ong DC, Haber N. Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers. ACM FAccT 2025. doi: 10.1145/3715275.3732039.
• Clegg KA. Shoggoths, Sycophancy, Psychosis, Oh My: Rethinking Large Language Model Use and Safety. J Med Internet Res. 2025;27:e87367. doi: 10.2196/87367. PMID: 41252530.
• Cho HN, Wang J, Hu D, Zheng K. Large Language Model-Based Chatbots and Agentic AI for Mental Health Counseling: Systematic Review of Methodologies, Evaluation Frameworks, and Ethical Safeguards. JMIR AI. 2026 Mar 13;5:e80348. doi: 10.2196/80348. PMID: 41592221.