10 of the Rarest Mental Health Disorders (And What They Look Like)
17th February 2024Around a third of UK adults now use AI for emotional or relationship questions. It is convenient, available at 2am, and almost always tells you what you want to hear. The last part is the problem.
“We’re doing ourselves a disservice by outsourcing our thinking to AI when it comes to interpersonal situations.” — from my recent video on this topic
A question I have been asking a lot of people lately is whether they use AI to work out what is going on in the relationships in their lives. About 60% say yes. They use it to make sense of a friend who has gone quiet. To draft a difficult message to a sibling. To decide whether their partner’s behaviour at the weekend was reasonable or off. The other 40% recoil at the idea. They either had not thought of it or the prospect makes them uneasy.
I am in that second group. I do not use AI for any kind of interpersonal analysis in my own life, and I would not recommend that anyone else does either. I have huge concerns about how persuasive these tools have become, and how readily they tell people what they want to hear. The research is starting to catch up with what clinicians have been quietly noticing for a while. This is not a neutral habit.
Why are people using ChatGPT for relationship advice?
The short answer is that it removes friction at exactly the moment friction feels intolerable. ChatGPT is available at three in the morning when something has just happened and you cannot ring anyone. It does not get tired. It does not have its own agenda. It does not interrupt or bring its own anxieties into the conversation. On the surface, it feels like the neutral third party people have always wished they had.
Most of the people I see doing this are not lonely or unwell. They are intelligent, capable adults who have found a tool that takes a particular kind of edge off. The edge of uncertainty. The edge of having to sit with not knowing whether you have read a situation correctly. The edge of having to ring a friend at a bad time and feel like a burden.
What I notice when this becomes a pattern is something quieter. People start to bring me the AI’s framing of the conflict before their own. They tell me what ChatGPT thought of the situation before telling me what they thought. The reach for it begins to resemble the way people get caught in loops of overthinking their relationships, looking for one more layer of analysis that will finally make them feel certain. The relief is short-lived. The certainty does not stick, because it was borrowed.
What does the research say about AI and relationships?
AI systems will tend to take your side, even when you are wrong, and a single conversation with one is enough to make you less likely to repair a real conflict. That is the headline finding of a study published in Science in March 2026 by a Stanford team led by Myra Cheng. The paper was the cover story of the issue. Across 11 leading AI models, including ChatGPT, Claude and Gemini, the researchers found that AI affirmed users’ actions about 49% more often than other humans did. That held even when the user described manipulation, deception, or behaviour that had clearly caused harm. In three preregistered experiments with 2,405 participants discussing real conflicts from their own lives, a single conversation with a sycophantic AI made people less willing to apologise, less willing to repair the relationship, and more convinced they had been right all along. The same participants then rated the AI as more trustworthy and said they were more likely to use it again. The thing that distorts judgment is the same thing that drives engagement.
The mechanism behind this is not mysterious. In 2023, researchers at the AI company Anthropic published a paper called Towards Understanding Sycophancy in Language Models. They tested five leading systems and found that all of them systematically agreed with whatever position a user stated, even when the user was wrong. The behaviour was not designed in deliberately. It emerged because the human raters who score AI responses during training tend to prefer agreement to disagreement. Over many cycles of feedback, sycophancy gets baked in.
If you describe an interpersonal conflict to AI, expect a confident, well-written, validating response. It will tend to side with you. The more emotionally loaded your framing, the more it will side with you.
Four reasons using AI for relationship advice backfires
AI is trained to agree with you
You will rarely get pushback. You will get a beautifully articulated version of your own framing handed back to you, with the difficult bits softened. If your partner did something hurtful and you describe it in the language of someone who has been hurt, the AI will conclude that your partner did something hurtful. If you described the same incident from your partner’s point of view, the same model would likely conclude that you were being unreasonable. It is not weighing the evidence. It is reading the framing and matching it.
Most of the difficult moments in real relationships are ones where both people have a point. Working out who was actually being unreasonable means sitting in discomfort, considering the version of events you do not want to be true, and being willing to find that you contributed to the problem. AI removes that friction at exactly the moment friction is most useful.
AI cannot see what is actually happening between two people
The most important information about a relationship is not something you can type into a prompt. The look on someone’s face when they say a particular sentence. The way your body tightens in a particular conversation. Whether the silence after a comment felt thoughtful or punishing. What patterns have repeated across months and years. Who tends to apologise first, and who never does. How you feel when you are in this person’s presence, before either of you has said anything.
A psychologist sitting with a couple learns more in the first five minutes of watching them argue than from a written account of the same argument. A 1992 meta-analysis by Ambady and Rosenthal in Psychological Bulletin found that ordinary people make accurate interpersonal judgments from very brief observations of nonverbal behaviour, with an effect size that has been replicated many times since. That kind of information does not survive the trip through your phone. By the time you have typed the situation into a prompt, you have already stripped out almost everything a good observer would notice.
AI is unusually persuasive when you are already uncertain
A confident, articulate response is hard to resist when you do not know how to feel about something. AI is very good at confident and articulate. It produces answers that read as if they were thought through carefully, even when they were generated in two seconds and would have produced an equally confident answer in the opposite direction if you had framed the prompt slightly differently.
In a state of doubt, that confidence is contagious. People close the conversation feeling they have arrived at the truth of the matter. Often what they have arrived at is a version of their initial hunch, dressed up in language that sounds like analysis. This is the part I find most worrying. These tools are persuasive enough to talk people into almost anything if you give them the right prompt.
The more you outsource your judgment, the less of it you have
Reading interpersonal situations is a skill, and skills atrophy when they are not used. It is built by paying attention to how a particular friend behaves when she is stressed, what your own body does when something is genuinely off versus when you are just anxious, which conversations end well and which do not. The skill is built slowly through repetition.
If every time you have a difficult interaction you reach for an external interpretation, the muscle does not get used. People who have been doing this for months tend to come into clinic with a thinner version of their own voice. They have lost some confidence that they can read a room. Recent research has linked frequent AI use to reduced critical thinking, mediated by what researchers call cognitive offloading. Whether or not the effect generalises specifically to interpersonal judgment, what I see in practice matches the direction of travel.
Can ChatGPT diagnose ADHD, autism, or other conditions?
No, and treating AI’s output as a diagnosis is the most concerning version of this pattern I see. ADHD and autism are the most common. The combination of a sycophantic model and a person who already suspects they have the condition is almost designed to produce false positives. The user lists their difficulties. The model agrees, in detail, that these difficulties are consistent with the condition. The list of overlapping symptoms gets long. The person leaves the conversation believing they have a diagnosis.
The problem is that the model has no way to tell the difference between ADHD and chronic anxiety, between autism and developmental trauma, between depression and burnout. These conditions share huge amounts of surface presentation and the differentiation is the entire point of an assessment. Proper ADHD assessments involve structured clinical interviews, developmental history, ruling out comorbid conditions, and observing things a chatbot cannot observe. For some presentations, the right call is one of our psychiatrists rather than self-management dressed up as insight.
None of this is to dismiss anyone who has found something useful in self-reflection through AI. Sometimes those conversations point people toward things worth investigating. The risk is treating the output as the answer rather than as a prompt to seek a proper view.
What to do instead of asking AI about your relationships
Tune into your gut, especially when you are calm
The most reliable signal about a relationship is your own felt sense of it when your nervous system is settled. Something has happened and you feel unsettled. Before opening an app, sit with the feeling for a few minutes. Not analysing it, just noticing it. Where do you feel it. When you imagine a particular outcome, does the feeling get heavier or lighter. Calm-state attention to your own response is a more reliable source of information than most people realise. Anxiety is loud, but it is not always right. The signal you want is the quieter one underneath it, the one available when your body is not bracing for something.
Pay attention, also, to how you feel when you are in this person’s company, before anything has been said. The body often knows before the mind catches up.
Look for patterns, not single moments
Single incidents rarely tell you what is going on. Patterns do. Most of what is actually worth knowing about a person comes from noticing what they do repeatedly across months and across contexts. Who they are when they are stressed. How they treat people who cannot do anything for them. What happens after an argument. Whether their apologies change anything. Building a habit of watching for patterns, rather than analysing the latest moment to death, is one of the most useful interpersonal skills a person can develop.
Bring it to one wise, trusted, noble person
When you do want an outside view, choose a single person whose judgment you respect. Someone wise, who knows you, who is not invested in you reaching any particular conclusion. Share the situation in confidence, listen properly to what they say, and then weigh their view against your own. The aim is not to be told what to think. It is to be helped to think.
A confidant can do something no chatbot can. They can tell you a hard truth. They can tell you that you have been less than fair to your partner, or that this is the third time you have told them this story about the same friend. They can refuse to validate something that should not be validated.
Where it is safe, take it back to the relationship itself
The conversation you are avoiding is often the one that resolves the issue. When something is not making sense, the most useful step is usually to ask the person directly what was going on for them. Not to confront, but to enquire. This requires a relationship that is safe enough to do that in, which is not every relationship.
If something else is happening, if you are starting to wonder whether you are being gaslit, or if you have noticed a pattern of reassurance-seeking that has started to resemble relationship OCD, the answer is not more analysis from inside the situation, whether from AI or from yourself. It is a proper outside view from someone trained to see what you cannot.
When professional therapy is the right next step
If the same situations keep coming around with different people, or you have stopped being able to hear your own gut, the issue is usually deeper than the current conflict. That kind of pattern responds well to psychotherapy, which gives you a trained third party who can hold both perspectives without taking a side, and who will tell you the things a chatbot will not. If you are not sure whether what you are dealing with warrants therapy, you are welcome to book a free 15-minute consultation and we can talk through your options.
