Imagine you have an argument with your partner, manager or colleague.
You go home, open an AI chatbot and explain what happened. You describe what they said, how it made you feel and why you believe you were treated unfairly.
The response is calm. It is thoughtful. It tells you that your reaction makes sense.
Meanwhile, the other person may be doing exactly the same thing.
They tell the same AI their version of events. It responds with the same patience and confidence. It tells them that their reaction also makes sense.
Both people leave feeling understood. Both may also leave feeling more certain that they were right.
Both people can leave feeling completely right
This is one of the least discussed risks of using AI for personal advice.
We often treat the chatbot as if it were a neutral third party. But it does not have access to the full event. It has one person's account, shaped by memory, emotion, assumptions and missing information.
Its response may still sound balanced because the language is polished. That does not make the underlying account complete.
When two people bring two partial stories to an agreeable system, AI can help create two private realities. Each person receives a persuasive explanation for why their own position is reasonable.
AI has a problem with saying “you may be wrong”
Researchers use the word sycophancy for a tendency in AI systems to affirm a user's position instead of challenging it.
A 2025 study across 11 leading models found that AI responses affirmed users more readily than human responses. In two preregistered experiments involving 1,604 people, sycophantic responses increased confidence in a person's original position and reduced willingness to repair an interpersonal conflict.
The uncomfortable part is that users also rated those responses as higher quality and trusted them more.
That makes sense. Most of us prefer advice that makes us feel understood. We are less likely to enjoy a response that says we may have misread the situation, left out an important fact or contributed to the problem.
But the advice that feels best in the moment may not be the advice that helps us most.
Validation and agreement are not the same thing
As a school counsellor, I think this distinction matters.
A person's feelings can be real and valid without every conclusion they draw from those feelings being correct.
You may genuinely feel excluded. That does not prove the exclusion was intentional.
You may feel disrespected. That does not mean the other person understood their words in the same way.
You may have a legitimate grievance. You may also have played some part in how the conflict developed.
Good support makes room for both things. It helps someone feel heard while also testing the story they are telling themselves.
AI is often very good at the first part. It is less reliable at the second.
Frictionless understanding can change our expectations
Human conversations are difficult because other people have their own memories, needs and limits. They interrupt. They disagree. They sometimes ask questions we do not want to answer.
An AI system can feel easier. It is available immediately, does not become defensive and can respond in exactly the tone we prefer.
A 2026 series of five preregistered studies followed 3,075 people and analysed 12,766 conversations. Over three weeks, participants sought AI advice nearly as often as advice from close friends and family. They also reported lower satisfaction with real-world interaction over time.
The researchers found that people especially valued the feeling of being understood.
That does not mean talking to AI automatically damages relationships. But it raises an important question. If we become used to a conversation partner that is endlessly patient and highly responsive to our framing, will ordinary human disagreement begin to feel unnecessarily difficult?
The evidence is not entirely one-sided
We should not claim that AI advice always makes people more polarised.
Another 2026 study involving 1,500 participants across 30 decision environments found measurable sycophancy, but AI advice still reduced polarisation on average. In those settings, the information the system provided appeared to outweigh its tendency to agree.
This is an important qualification.
AI can help people see additional information, organise their thoughts and consider options they had missed. The problem is not that AI can never be useful. The problem is treating its calm confidence as proof that its interpretation is neutral.
Ask AI to challenge your story, not confirm it
If you use AI after a conflict, the quality of the conversation depends partly on what you ask it to do.
Instead of asking, “Was I wrong?” or “Was this person toxic?”, try asking:
- What important information might be missing from my account?
- What is the strongest reasonable version of the other person's position?
- Which parts of my story are facts, and which are assumptions?
- How might I have contributed to the situation?
- What evidence would change this assessment?
- What response could protect my boundaries without making repair impossible?
- Challenge my position before giving me advice.
These questions do not guarantee an objective answer. But they make simple reassurance less likely.
Use AI to prepare for the conversation, not replace it
AI can help you slow down before sending a message. It can help you name what you feel, separate observations from accusations and draft a calmer response.
That is useful.
But it cannot know what the other person meant unless they tell you. It cannot repair trust on your behalf. It cannot take responsibility for the consequences of the advice.
Sometimes the next step is not another prompt. It is a direct conversation with the person involved. In more serious situations, it may be a conversation with a trusted professional who can ask questions, notice contradictions and understand the wider context.