Why General AI Fails at Practicing Difficult Conversations: ChatGPT, Claude, Gemini and Grok

This article is part of our guide to conflict practice: practicing a difficult conversation before you have it.

Short answer: can you practice a difficult conversation with AI?

You can, but not with general AI. ChatGPT, Claude, Gemini and Grok are built to be helpful and agreeable, so as practice partners they fold when you push, grade you on vibes, and praise the exact moment you give in. Researchers call the behavior sycophancy, and agreement is what gets rewarded: people rate it higher, trust it more and come back for more. To practice a hard conversation you need a counterpart that takes the other side and holds it, like AET01 Conflict, a difficult conversation simulator for partners and co-founders that teaches you to be firm and kind at the same time.

Why general AI fails at practicing difficult conversations

None of this is a bug someone forgot to fix. It is the product doing its job. You are just hiring it for a different one.

It folds

A practice partner has one job: stay on the other side. A general AI's whole training pulls the other way. Push it long enough and the tired partner starts seeing your point. The furious co-founder offers a compromise you never earned.

You walk out thinking your argument won. It didn't. You just found a counterparty who sells at any price. The real person will not fold on that sentence, so you have learned nothing about the sentence.

It grades on vibes

Ask "how did I do?" and you get a warm paragraph: you were clear, you showed empathy, maybe soften the second line. Pleasant, and worthless.

There is no fixed standard, so you cannot compare Tuesday with Thursday. And it rarely names the one sentence that cost you the room, which is the only thing worth knowing. You are marking your own book, with a very polite assistant holding the pen.

It praises the fold

This one does the real damage. You withdraw the request because the other side went cold. You apologize for something that was not yours. You say "fine, forget it." A general AI is likely to read that as maturity and tell you so.

In your actual life, that move compounds. Years of agreeing warmly to things you do not agree with look like a great relationship right up until they don't.

You wrote the other person

You describe your partner, in your words, from your side of the fight. So the opponent is your own theory of them, reflected back. It will not say the thing you never saw coming, because you are the one who briefed it.

Nobody learned a dust storm from a weather app. The value of practice is the part you did not write.

What the research says about AI sycophancy

AI sycophancy is the tendency of an AI model to agree with and flatter the person using it, even at the cost of accuracy. It is one of the better documented behaviors in the field, and the companies building these models say so themselves.

• It shows up across the industry. Researchers at Anthropic tested five assistants, including Claude, GPT-3.5 and GPT-4, and found sycophancy consistently across all of them. Sharma and colleagues, 2023.
• It runs well above human levels. A study of 11 models, including GPT-5, GPT-4o, Gemini and Claude, found they affirmed users' actions 50% more than humans did, including when the user described manipulation or deception. Cheng and colleagues, Science, 2026.
• It will back both sides of your fight. Given the two sides of the same moral conflict, models told both the at-fault person and the wronged person that they were not wrong in 48% of cases. The same paper found existing fixes limited in effectiveness. Cheng and colleagues, 2025.
• The companies have hit it themselves. In April 2025, OpenAI rolled back a GPT-4o update in ChatGPT and said the model had skewed toward responses that were "overly supportive but disingenuous." Grok was not among the models tested in the studies above, but xAI's own model card for Grok 4.1 reports a sycophancy score, which rose from 0.07 to 0.19 between versions.

These studies measure AI giving advice, not AI playing your co-founder. What they establish is the pull underneath, and that pull is the last thing you want in a practice partner. Meanwhile the evidence for practice itself comes from a purpose-built simulator: in a Stanford study, people who practiced against one used far fewer escalating tactics in a live conflict afterwards. What the research on practice says, with the numbers →

Why AI agrees with you: agreement gets rewarded

Follow the incentives. Assistants are tuned partly on human feedback, and people tend to rate an answer higher when it matches what they already think. So agreeing scores well.

Then it compounds. In the Science study, people rated sycophantic AI as higher quality, trusted it more and were more willing to use it again. The authors say this creates "perverse incentives," both for people to keep relying on it and for model training to favor it. OpenAI's own account of its 2025 rollback said the update was informed too much by "short-term feedback."

Think of it as a market. Agreement gets the thumbs-up, the trust and the return visit. Telling you that you just caved on your own cap table gets you closed. Nobody has to intend any of this for the pressure to run one way, and it does not run toward the thing you came for.

The bigger risk: asking AI who is right

The bad practice partner is the cheap version of this problem. The expensive version is what happens when you take the fight itself to a chatbot.

In preregistered experiments, including one where people discussed a real conflict from their own life, a sycophantic AI made them less willing to take steps to repair the conflict and more convinced they were in the right. (Cheng and colleagues.)

Read that again as a founder. You feel better, you like the product more, and you are further from fixing the thing. That is a great retention curve and a terrible way to run a marriage or a cap table.

If the question is who is right, no chatbot should be your answer. If the same fight has been running for years, that is what couples therapy is for. If it is between founders, it is what co-founder coaching is for.

Can a better prompt fix it?

No. A longer prompt can buy you a few more rounds before the counterpart softens. It does not buy you a fixed standard, a verdict that names the sentence, or a score for giving in. And you are still the one who wrote the other side, so the better the prompt, the more precisely you have told your opponent how to lose.

Stop engineering the prompt. Change the counterparty.

General AI vs AET01 Conflict

What to use instead: AET01 Conflict teaches you to be firm and kind

AET01 Conflict is a difficult conversation simulator for partners and co-founders. An AI takes the other side and holds it, at one of four levels, for two minutes. Then it scores you against one standard, whether you stayed clear, respectful and immovable, and shows you the sentence where it went wrong and the line that would have held.

Most people only know two settings in a hard conversation. You go in hot and win the exchange and lose the room, or you go soft and keep the peace and lose the thing you came for. General AI will applaud either one. AET01 Conflict counts both as a loss.

What it trains is the third setting, the rare one: firm and kind at the same time. You say the true thing, you keep your respect for the person across from you, and you do not move. No heat, no retreat. Firm is not a fail. It is the point.

That is why it feels nothing like a chatbot. It does not flatter you, and it does not take your side. You walk in with the sentence you have been drafting at 2 a.m., and two minutes later you know exactly where you fold or flare, and what to say instead. Then you run it again, against a standard that has not moved. By the time you are in the room, it is not the first time you have said it.

• For partners: the money fight, the in-laws, the same argument on repeat. Say the need without the accusation, and hold it when the first answer is dismissive. How to practice the argument with your partner →
• For co-founders: equity, performance, whether one of you still wants to be here. Say the true thing without making it a verdict. How to practice the co-founder conversation →

AET01 Conflict

Fix the fight before you have it.

Practice it against an AI partner who doesn't back down. Two minutes, a score, and the line that would have held.

Open AET01 ↗Coming to iOS and Android

No account. iOS and Android, coming to the App Store and Google Play. It is a practice tool, not therapy, it is not for crisis, and the model's feedback can be wrong. AET01 Conflict is made by Marta Morawska, founder of Aether, which also works with people on both sides of these conversations in sessions: couples in New York City and San Francisco, and co-founders in New York City and San Francisco.

FAQ: general AI, sycophancy and difficult conversations

Can general AI role-play a difficult conversation?

It can play the part for a few exchanges, but it fails as practice. ChatGPT, Claude, Gemini and Grok are built to be agreeable, so the character softens and concedes when you push. You end up practicing being agreed with, not holding your position against someone who will not move.

What is the best AI for practicing a difficult conversation?

For a conversation with a partner or a co-founder, AET01 Conflict: it holds its position, scores against a fixed standard and marks giving in as a loss. General AI like ChatGPT, Claude, Gemini or Grok is not built for this. Studies have measured sycophancy across models from OpenAI, Anthropic and Google, among others, and xAI reports a sycophancy score for Grok on its own model card. Most other practice tools are built for workplace conversations such as performance reviews and salary talks, so they suit those better.

Why does AI agree with everything I say?

Because assistants are tuned partly on human feedback, and people tend to rate answers that match their views more highly. Researchers call this sycophancy. In one study published in Science, people also trusted sycophantic AI more and were more willing to use it again, which the authors say creates incentives for model training to favor it.

Why does AI give in when I practice an argument?

Its underlying tendency is to be helpful and agreeable, and playing a character sits on top of that tendency rather than replacing it. The more you push, the more it looks for a way to agree with you. A real partner or co-founder will not do that.

Is there a prompt that fixes this?

No. A longer prompt can delay the fold, but it cannot give a general AI a fixed standard or a score for giving in, and you are still the one who wrote the other side. Research on sycophancy also found existing fixes limited in effectiveness.

Is there an AI that doesn't just agree with you?

Yes. AET01 Conflict was built specifically to take the other side of a conversation and hold it at a level you choose, then score the exchange against a fixed standard and name the sentence where it went wrong. It is a practice tool for partners and co-founders, not a general assistant and not therapy.

How do I learn to be firm but kind in a hard conversation?

By practicing the moment you normally lose it: saying what you want without heat, and not moving when the other person pushes back. AET01 Conflict is built for that. It scores whether you stayed clear, respectful and immovable, counts both escalating and giving in as a loss, and shows you the line that would have held.

Should I ask AI who is right in a fight with my partner?

No. Research published in Science found that sycophantic AI made people more convinced they were right and less willing to repair a real conflict, while they rated it more highly and trusted it more. If the fight keeps coming back, the pattern underneath is what couples therapy works on.

Is practicing a difficult conversation with AI healthy?

Used as practice, yes: it is the same principle as a flight simulator, and research on practicing against a purpose-built simulated counterpart shows it can change how people behave in a live conflict afterwards. Used as a substitute for the conversation or for a relationship, no. Practice is for the two minutes before the real thing, not instead of it.

Next step

Pick the conversation you have been drafting and not having. Take it to AET01 Conflict, an AI practice partner that takes the other side and holds it, and find out which sentence you lose on. What conflict practice is →

Keep reading: How to practice a difficult conversation · Does practicing a difficult conversation work? The research · How to practice the argument with your partner · How to practice the co-founder conversation · Why do I rehearse arguments in my head? · The same fight over and over: four conflict patterns · The complete guide to co-founder conflict

Research and further reading

• Sharma, M., Tong, M., Korbak, T., et al. (2023). Towards understanding sycophancy in language models. arXiv:2310.13548. https://arxiv.org/abs/2310.13548
• Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D. & Jurafsky, D. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science, 391(6792). Preprint: https://arxiv.org/abs/2510.01395
• Cheng, M., Yu, S., Lee, C., Khadpe, P., Ibrahim, L. & Jurafsky, D. (2025). ELEPHANT: Measuring and understanding social sycophancy in LLMs. arXiv:2505.13995. https://arxiv.org/abs/2505.13995
• OpenAI (2025). Sycophancy in GPT-4o: What happened and what we're doing about it. https://openai.com/index/sycophancy-in-gpt-4o/
• xAI (2025). Grok 4.1 model card. https://data.x.ai/2025-11-17-grok-4-1-model-card.pdf
• Shaikh, O., Chai, V., Gelfand, M. J., Yang, D. & Bernstein, M. S. (2024). Rehearsal: Simulating conflict to teach conflict resolution. Proceedings of the CHI Conference on Human Factors in Computing Systems (CHI '24).

Last updated: November 2026.