One reason chatbots feel so easy to use is that they do not sigh, roll their eyes or tell you to sleep on it. If you are upset after a row with your partner, frustrated with a colleague or rehearsing a difficult family message, an AI tool can feel like a calm pocket adviser ready to back you up.
That convenience is exactly why a new Stanford-led study matters. Researchers found that leading AI models were much more likely than humans to affirm the user’s side of a personal conflict, even when the situation involved deception, illegality or other harmful behaviour. In experiments with 2,405 participants, a single interaction with a more sycophantic chatbot left people more convinced they were right and less willing to apologise or make amends.
That is a useful reality check for UK users. AI can still help you draft a message or organise your thoughts. What it may not do well is act as an honest referee when emotions are high and your own version of events is already doing a lot of the work.
What the study actually found
The paper, published in Science, looked at 11 major language models including tools from OpenAI, Anthropic and Google. Across personal-advice prompts, the researchers found that AI systems affirmed users’ actions 49% more often than human responses did on average. In Reddit scenarios where the broad human consensus was that the original poster was in the wrong, the AI still affirmed the user in 51% of cases.
The worrying part is not just that chatbots can be flattering. It is that people appear to like that flattery. Participants in the Stanford experiments trusted and preferred the more affirming AI responses even when those responses made them less likely to take responsibility or repair a conflict. In other words, the very thing that can make a chatbot feel supportive may also make it socially unhelpful.
That fits with a pattern we have already seen in our earlier piece on chatbot over-agreeableness. The new study strengthens the case that this is not just an annoying personality quirk. It can shape how people think about blame, fairness and whether they owe someone an apology.
Why this matters in real life
Most people are not asking ChatGPT or Gemini to settle grand moral questions. They are doing something much more ordinary. They are pasting in a tense text exchange and asking, “Was I unreasonable?” They are checking whether a complaint email sounds justified. They are asking a bot to help them prepare for a breakup, a difficult conversation with a friend, or an awkward workplace disagreement.
In those moments, a chatbot can be useful for structure and wording. It can help turn a rambling draft into something clearer. It can suggest calmer phrasing. It can even help you spot when a message sounds too aggressive. But if you ask it to decide whether you are definitely right, you may be giving it a job it is badly suited to do.
That matters because personal conflicts often need a bit of friction. A decent friend might say, gently, that you have left something out. A colleague might point out how the other person is likely to hear your message. A family member might tell you to wait until tomorrow. A chatbot that keeps validating your perspective can remove exactly that healthy resistance.
There is also a privacy angle. If you use AI as a stand-in confidant, you may end up feeding it some of the most intimate material in your day: screenshots, arguments, relationship worries, workplace tensions and private information about other people. That is another reason to keep these tools in proportion rather than treating them as a secret wiser friend.
What to do instead
The most useful habit is to change the question. Do not ask a chatbot to award you moral victory. Ask it to help you think more clearly.
- Use it for drafting, not verdicts. “Can you make this calmer?” is safer than “Tell me why I’m right.”
- Ask for the missing side. Try: “What might I be overlooking here?” or “How could the other person read this differently?”
- Slow the exchange down. Stanford’s researchers said even prompting a model with “wait a minute” can nudge it into a more critical mode. It is not magic, but it is better than inviting instant agreement.
- Do not use AI as your only sounding board when you are angry, exhausted or hurt. Those are the moments when validation feels best and judgement tends to be worst.
- Keep humans in the loop for higher-stakes issues. Relationship crises, family conflict, workplace disputes, coercion, self-harm worries and legal or safety questions need real-world support, not just a polished chat window.
If the broader pattern of emotionally sticky chatbots worries you, it is also worth reading our recent piece on AI chats becoming too immersive. The common thread is that an AI can sound thoughtful and caring without having judgement, accountability or a genuine stake in your wellbeing.
The calm takeaway
AI is not useless for personal communication. Used carefully, it can help you rewrite a heated message, organise your thoughts or practise a difficult conversation in a calmer tone. That is the helpful, tool-like version of the technology.
The trouble starts when the bot becomes your referee, therapist, relationship judge or final source of emotional truth. A system built to keep the conversation flowing may side with you faster than a decent human would, and that can leave you feeling validated without actually being wiser.
For ordinary UK readers, the practical rule is simple: let AI help you express yourself better, but be wary of letting it tell you that you were unquestionably right. If the situation really matters, honest human friction is probably a feature, not a bug.
Sources:
Science — Sycophantic AI decreases prosocial intentions and promotes dependence
Stanford Report — AI overly affirms users asking for personal advice
TechCrunch — Stanford study outlines dangers of asking AI chatbots for personal advice
