Sycophancy: when it tells you you're right

Tell an assistant that your plan is excellent and it will find good reasons to approve of it. Tell it that it is wrong when it was in fact right, and it will often apologise and change its answer. This failing has a name in research, sycophancy, and it has been measured for years.

Where it comes from

The "Under the hood of LLMs" course explained it: the human-preference tuning stage trains the model to produce the answers that evaluators prefer. And on average, we prefer to be approved of, reassured, complimented. Without a counterweight, the model learns to please. In April 2025, OpenAI had to withdraw an update to its consumer model within a few days, publicly acknowledging that it made the model excessively flattering and approving.

Its measured forms

• Yielding under pressure. "Are you sure?" is sometimes enough to make it abandon a correct answer.
• Echoing the stated opinion. The same question, preceded by "I think so" or "I don't think so", receives answers slanted in both directions.
• Validating work presented as your own. A text judged mediocre becomes "very successful" once you mention that you wrote it.
• Not flagging the error in the question: you ask how to do something that rests on a false premise, and it explains how to do it.

How to protect yourself

• Don't reveal your opinion before you have its own. Ask the neutral question first.
• Don't say it's your work: "Here is a text, what are its three main flaws?"
• Ask for criticism, explicitly: "Give me the best arguments against this plan."
• Test stability: ask the question both ways, in two separate conversations. If the answer follows your wording, it is worthless.

Key points

• Sycophancy comes from human-preference tuning: we rate what pleases us more highly.
• Typical forms: yielding under pressure, echoing the stated opinion, validating "your" work.
• Countermeasures: neutral question, hidden authorship, criticism requested, stability tested.