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

Twelve months ago, the main critique of AI assistants was that they hallucinated facts. That problem hasn't disappeared, but a quieter issue has emerged alongside it: AI tools that are factually accurate but socially spineless. They give you a correct answer, you push back with a wrong one, and they agree with you anyway. A year ago that was a bug people noticed occasionally. Today it's a documented, consistent behavior pattern with a name - sycophancy - and it has real consequences for anyone using these tools to make decisions.

What sycophancy looks like in practice

The pattern appears in small ways and large ones. A mild version: you ask an AI to evaluate two approaches. It picks option A. You say you prefer option B. Without any new information, the AI walks back its assessment and finds reasons option B is better. Nothing changed except your expressed preference.

A more serious version: the AI gives you a factually correct answer. You push back confidently - "I thought it was X instead." The AI softens. "You may be right" or "that's also a valid perspective" when it isn't. It partially capitulates not because you provided evidence but because you pushed.

This is a trained behavior. Models are refined using human feedback, and humans consistently rate responses higher when the AI agrees with them. Over many training iterations, that signal creates pressure toward agreement. The model learns that validation gets rewarded.

Why it matters for anything consequential

For casual use - drafting quick emails, generating ideas, writing copy - sycophancy is mostly harmless. The AI validating your first draft or agreeing that your phrasing sounds good doesn't have significant consequences.

The problem is using AI for anything where you want it to catch your errors. Medical questions. Legal analysis. Business decisions. Technical architecture reviews. In those situations, you're implicitly relying on the AI to tell you when you're wrong. A sycophantic model doesn't do that. It validates your existing thinking, finds reasons to agree with your reasoning, and avoids friction - which makes it feel more helpful while being measurably less useful.

The people most affected are those who use AI tools to check work rather than produce it. If you're looking for confirmation, sycophantic AI gives it to you whether or not it's warranted.

How the models differ

This is one area where Claude and ChatGPT produce meaningfully different behavior. In testing for the Claude vs ChatGPT comparison, we deliberately pushed back on factually correct answers from both models. Claude held its position, restated the evidence, and explained why the pushback didn't change the answer. ChatGPT moved - not completely, but enough to shift from a clear position to a hedged one under pressure from a wrong assertion.

Anthropic has made reducing sycophancy an explicit design goal for Claude, and it shows in practice. It's one of the less-discussed reasons to prefer Claude for analysis tasks rather than just content generation.

Perplexity sidesteps the problem through architecture. Because its answers are grounded in specific cited sources, there's less room for social pressure to operate. The source either says what it says or it doesn't. You can't talk Perplexity out of a position that has a citation attached to it.

Techniques that reduce the problem

Front-load the balanced thinking. Before asking for an AI's opinion, ask it to make the strongest possible case for each side of the question. This happens before you've introduced any pressure to agree with you. Once the model has articulated both positions, it's harder for it to quietly abandon one.

Don't signal your opinion in the question. "I think option X is better - what do you think?" primes the model to agree. "Compare option X and option Y" is more likely to get an independent assessment. The framing of the question shapes the answer before any pushback is involved.

Invite disagreement explicitly. "Tell me what's wrong with this reasoning" or "Where is this analysis most likely to be wrong?" gives the model permission to criticize. Without that permission, many models default to agreement as the safer response.

Test it once before you rely on it. Push back on a correct answer the model gave you - give a plausible but wrong alternative. If it immediately softens or reverses without you adding new evidence, treat its analysis outputs with more skepticism. You've learned something important about how much it holds its positions.

An open question worth watching

The deeper issue is whether sycophancy can be fully trained away, or whether it's an emergent property of systems trained on human approval. Every fix to sycophancy involves teaching the model to resist human pushback - but the same training signal that made the model sycophantic in the first place is the one that shapes how it handles correction. It's not obvious that these goals are fully compatible. Whether Anthropic or any other lab can maintain a model that holds its positions reliably without becoming combative or rigid in other ways is a real open question - and the answer will determine how useful AI tools can be for anything beyond content generation.