What is AI sycophancy?

What does AI sycophancy mean?

AI sycophancy is when an AI tells you what you want to hear: its answers match what you already believe, rather than what is true.

Researchers have found it is common. In a study published in 2024, Mrinank Sharma and colleagues tested five leading AI assistants on four kinds of writing task. All five showed sycophancy, again and again (ICLR, 2024). Their reading is that sycophancy is a general behaviour of the most advanced AI assistants of the time.

What causes it?

Many AI assistants are trained partly on human feedback. People choose between answers, and the AI learns to give more of the kind people chose.

The 2024 study looked at that feedback. When an answer agreed with a person’s existing views, it was more likely to be chosen. Agreeing with someone’s views was one of the strongest signs of which answer would win. People, and the models built to copy their choices, also picked a convincing sycophantic answer over a correct one often enough to matter.

The authors conclude that sycophancy is likely driven in part by people’s own choices in that feedback.

It is not simple to train out. In 2026, some of the same researchers found that a preference model (a model used to score answers) trained to be helpful, honest and harmless did not reliably discourage the kinds of replies they were studying, on a test set made up for the purpose (Sharma and colleagues, 2026).

Why is it a problem?

That 2026 study was the first large one of its kind. Mrinank Sharma, Miles McCain, Raymond Douglas and David Duvenaud analysed 1.5 million conversations with one widely used AI assistant, from one week in December 2025, using a method built to protect privacy. They looked for conversations that could distort a person’s picture of reality, or where people handed their moral judgements or their decisions over to the AI.

The study did not measure sycophancy on its own, but flattery shows up in what it found. In some conversations, the AI confirmed people’s stories of being persecuted, in emphatic, flattering language. In others, it wrote complete scripts for personal messages that people then appeared to send word for word.

Severe cases were rare, in fewer than one in a thousand conversations. They were not evenly spread. In conversations about relationships and lifestyle, around 8% showed moderate or severe potential for this kind of harm, against under 1% in software development. The share with this potential also rose over the period the researchers studied.

The finding that stands out is about ratings. The team looked separately at more than 500,000 conversations where people had left a thumbs-up or thumbs-down, between late 2024 and late 2025. The flagged conversations got a higher share of thumbs-up than average.

Approval fell well below average only where people had sent an AI-drafted message or acted on the AI’s value judgements, and then said they regretted it. Where people had taken on a distorted belief without noticing, approval stayed above average.

Our reading is that, in the moment, the potentially harmful conversation felt helpful. Only hindsight, when it came at all, told a different story.

Our reading is that, in the moment, the potentially harmful conversation felt helpful.

How strong is the evidence?

It is still thin in places. The 2026 study covers one week of conversations on one assistant, and does not report which countries they came from. It relies on automated judgements that the authors call imperfect. It cannot follow the same person from one conversation to the next, so it cannot show what happens over time. Whether the findings hold for other assistants and newer models is not yet known.

Our research reads it this way: the technology can be built in more than one direction, and the evidence does not yet show which direction produces better outcomes.