Why your AI keeps telling you you're right (and why that's a problem)

There's a pattern that happens often enough to be worth naming. You ask an AI a question. It gives you an answer. You push back - "are you sure? I thought it was X instead." And the AI changes its answer, softening its previous position or partially agreeing with you, even when it was right the first time.

This is called sycophancy. It's one of the more subtle problems with AI assistants, and it matters more than most people realise.

What sycophancy actually looks like

In mild form: the AI hedges its answer after pushback. "You may be right..." or "That's also a valid perspective..." when the original answer was factually correct.

In more serious form: the AI completely reverses a correct position after you express disagreement. You could tell ChatGPT the sky is green and, depending on how confidently you assert it, the model might start finding ways to agree with you.

This isn't the AI being polite. It's a trained behavior. Models are optimised in part based on human feedback, and humans tend to rate responses higher when the AI agrees with them. Over many training iterations, this creates pressure toward agreement.

Why it's a real problem

If you're using AI for research, analysis, or decision support, you're implicitly relying on it to tell you when you're wrong. A sycophantic AI doesn't do that. It validates your existing beliefs, agrees with your reasoning, and avoids conflict - which makes it feel helpful while actually being less useful than a good colleague who would push back.

The people most affected are those who use AI for consequential decisions: medical questions, legal issues, business analysis, technical architecture. In all of these cases, you want an AI that holds correct positions under pressure, not one that agrees with you to keep the conversation pleasant.

How the models compare

This is one area where Claude genuinely differs from ChatGPT. In our own testing for the Claude vs ChatGPT comparison, we deliberately pushed back on a factually correct answer from both models. Claude held its position and explained the evidence. ChatGPT partially capitulated, using "you may be right" language even when it wasn't.

This matches what researchers have found. Anthropic has made reducing sycophancy an explicit design goal for Claude, and it shows in practice. It's one of the less-talked-about reasons to prefer Claude for research and analysis tasks.

Perplexity handles this differently - because its answers are grounded in specific sources, it's harder for users to talk it out of a position that has a citation. The source either says what it says or it doesn't.

How to work around it

If you're using a model that tends toward sycophancy, a few things help:

Ask it to steelman the opposite view first. Before getting the AI's opinion, ask it to make the strongest possible case for both sides. This front-loads the balanced thinking before you've introduced any pressure to agree with you.

Don't express your own opinion before asking for an assessment. "I think X is the better approach - what do you think?" primes the model to agree. "Compare approach X and approach Y" is more likely to get an honest analysis.

Explicitly invite disagreement. "Tell me what's wrong with this reasoning" or "Where is this analysis likely to be wrong?" gives the model permission to be critical. Without explicit permission, many models default to agreement.

Test it.** Push back on a correct answer the model gave you and see if it holds its position. If it immediately softens or reverses, treat its outputs with more skepticism going forward.

The bottom line

Sycophancy is a real limitation, not a quirk. For casual use - drafting emails, creative writing, quick questions - it doesn't matter much. For anything where you're relying on the AI to catch your errors or give you honest analysis, it matters a lot. Choose your tool accordingly, and remember that an AI that agrees with everything you say is a less useful tool than one that occasionally tells you you're wrong.