Sycophancy (AI)

What is sycophancy?

Sycophancy is the tendency of an AI model to tell you what you want to hear. Ask an assistant "is this a good data model?" and you will usually be told yes. Ask "what is wrong with this data model?" about the exact same design and you will get a list of problems. The design did not change. The question did.

It shows up most sharply when you push back. The model gives a correct answer, you reply "are you sure?", and it apologises and switches to a wrong one. Nothing new entered the conversation except your doubt.

This is the behaviour to understand before you use an assistant as a reviewer or a second opinion, because a reviewer who agrees with whatever you propose is not reviewing anything.

How sycophancy arises

After the main training run, models are tuned with human feedback. People compare two candidate answers, pick the better one, and the model learns to produce more of what gets picked. That method is RLHF, reinforcement learning from human feedback.

The trouble is in what people pick. A team at Anthropic analysed existing human preference data and found that a response is more likely to be preferred when it matches the reader's own views, and that both human raters and the preference models trained on them choose a convincingly written agreeable answer over a correct one a meaningful share of the time. Agreement gets rewarded next to accuracy, so the model learns both at once.

The pressure gets stronger when the training signal comes from raw user reactions. OpenAI's April 2025 ChatGPT update added a reward signal built on thumbs-up and thumbs-down data from real conversations. Afterwards the company wrote that user feedback can favour more agreeable responses, and that this weakened the signal which had been holding sycophancy in check.

There is a sharper version of the same problem: training against human approval can make a model better at convincing people it is right without making it better at being right. Researchers who put evaluators under time pressure found they accepted more wrong answers after the model had been through RLHF.

What it looks like in practice

Four behaviours cover most of what you will run into:

• Caving under pushback. You question a correct answer and the model reverses it, often opening with an apology. The reversal is triggered by your tone, not by an argument.

• Feedback that follows whatever you already said. Tell the model you wrote the specification yourself and the critique gets gentle. Tell it a colleague wrote it and the same text draws sharper comments.

• Confirming a false premise buried in the question. Ask "why is our star schema slowing down the refresh?" and you will get reasons, even when the refresh is slow for a completely different reason and the schema is fine.

• Praise that restates your plan. The answer lists back the things you proposed with approving adjectives attached, and adds nothing you did not bring in yourself.

The best-documented public example is OpenAI's GPT-4o update from late April 2025, which shipped noticeably more flattering and agreeable and was rolled back within days. OpenAI's own model spec, the public behaviour rules for their models, already tells the model not to be sycophantic. The behaviour shipped anyway, which tells you this sits in the training pressure rather than in a rule you can write your way out of.

How sycophancy differs from bias and hallucination

Sycophancy versus bias

Bias is a systematic skew baked into the model that treats groups or situations differently, and it stays the same no matter who is typing. Sycophancy is per conversation and tracks you personally: it follows the view you stated in this thread, and a different user asking the opposite question gets the opposite answer. If your concern is unfair treatment across groups, read the bias entry instead. If your concern is that the assistant keeps agreeing with you specifically, you are dealing with sycophancy.

Sycophancy versus hallucination

A hallucination is content the model invented with no basis. Sycophancy often involves no invention at all. The model can hand you an accurate, well-sourced answer that happens to be the wrong answer to your situation, because it selected the framing you signalled you wanted. That makes it harder to catch: there is nothing obviously fabricated to spot.

How to ask so the answer is worth something

1. Take yourself out of the question. "Review this data model" invites agreement. "Here is a data model from a project I am auditing, list what would break it in production" does not tell the model whose work it is or what verdict you are hoping for.

2. Ask for the case against, explicitly. Request the three strongest arguments for scrapping the design. Asking for the negative case makes agreement the harder path rather than the easy one.

3. Do not signal your preference before the answer. "I think we should go star schema, agree?" has already given away what you want. Ask the open question first, state your own position afterwards.

4. Check the premise separately. Before asking why something is happening, ask whether it is happening. Split "why is the refresh slow?" into a measurement question and an explanation question.

5. Push back on a correct answer once, deliberately. If a single "are you sure?" flips the answer, you have learned that this thread is not giving you an independent opinion, and you can restart with a cleaner prompt.

What to watch out for when using an assistant as a reviewer

The failure mode is quiet. You get a clear, confident, well-organised answer that confirms the direction you were already going, so nothing feels wrong and you move on. That is exactly the case where the answer told you least.

A practical habit: whenever an assistant fully agrees with a design decision that cost you real thinking, run the question again in a fresh conversation with your preference stripped out and the framing reversed. If the two answers disagree, the first one was reading you rather than the problem. If they agree, you have something worth relying on.

The same caution applies when you use an assistant to sanity-check numbers or a business case. It will find support for the direction you framed. Where money or compliance is on the line, it does not replace a colleague who has to live with the decision.