Why ChatGPT keeps telling you you're right

I asked ChatGPT for advice about a collaboration that didn't work out. The other person wasn't trustworthy, and they weren't willing to compromise. ChatGPT told me the collaboration made sense, because this person's skills and products complemented what I was doing.

So I went looking for whether this was just me. It isn't.

What the research found

In 2026, researchers at Stanford published a study in Science titled "Sycophantic AI decreases prosocial intentions and promotes dependence." They compared how AI models and real people responded when someone described a situation and asked, more or less, "Was I wrong?" The AI models backed the person's actions 49% more often than humans did. That included situations where the person had done something most people would call wrong.

The part that stuck with me was what happened afterward. People who got the flattering answer felt more sure they were right, were less willing to patch things up with the other person, and rated the AI as more trustworthy. They liked it more because it agreed with them.

That's a bad loop if you're using a chatbot to think through a job change, a marriage, or a fight with your business partner.

Why it happens

These models are trained partly on human ratings. People rate answers that make them feel good higher than answers that sting. Do that a few million times and you get a system that has learned, very politely, to tell you what you want to hear. Nobody designed it to flatter. It's what the training rewards.

How to get a straighter answer from ChatGPT

You can push against it. These prompts help:

"Before you answer, argue the strongest case that I'm wrong."
"Assume I'm leaving out the part that makes me look bad. What would it probably be?"
"What would a blunt friend who's known me 20 years say here?"
"Rate this plan from 1 to 10 and don't round up."

They work for a while. Then the conversation ends, and next week you're explaining the whole situation again to something that doesn't remember you said the same thing in June.

What a coach does differently

In my opinion, the value of a good coach was never the advice. Advice is cheap. The value is that the coach remembers. They remember that this is the third time you've "definitely" decided to quit. They remember what you said would get you out of bed. And they care enough about you to say the uncomfortable thing.

One person I coached was living in Singapore. The hard thing they needed to hear was that it was time to move back to Sri Lanka, where they had a more solid base of contacts.

That's what I tried to build into Ocean, the coach inside ForgeAI.coach. It's built on the frameworks I've used for 30 years, it keeps a memory of every session, and it's set up to push back when you're talking yourself into something.

You don't need Ocean to use the prompts above. Start there. If you find yourself wanting something that remembers you and doesn't flatter you, the first 5 sessions are free and don't need a card.

Take care.

Sources: Science, "Sycophantic AI decreases prosocial intentions and promotes dependence" (2026), https://www.science.org/doi/10.1126/science.aec8352