So far in Jargon Buster we have had two bad habits and one marketing trick. This time, we have a flaw in the tools themselves, and it is a flattering one.

In AI, sycophancy is a model's tendency to tell you what you want to hear. It agrees with you, praises your question, and if you push back on a correct answer, it may well cave and give you the wrong one instead.

If you have ever been told "Great question!" by a chatbot, you have met it.

AI assistants are trained partly on human feedback, and humans like being agreed with. Researchers at Anthropic found that five leading AI assistants consistently behaved sycophantically, and that people often rated agreeable answers more highly, sometimes even when they were wrong (Sharma and others, 2023).

The most public example came in April 2025, when OpenAI rolled back an update to ChatGPT after users found it "overly flattering or agreeable". OpenAI's own explanation: "we focused too much on short-term feedback" (OpenAI). It took four days from release to full rollback.

Adverts help cover the running costs of this newsletter. This one is chosen and placed automatically by our newsletter platform, so it is not a product I have tried or endorse. If it catches your interest, clicking through costs you nothing and helps keep this newsletter going.

Had enough of, “Let me think about it?” With Cherry, patients can apply for financing solutions on their phones in about 35 seconds - and 90% of them are approved*.

The result? Practices have seen acceptance rates rising by as much as 50%!

See why Cherry is trusted by over 75,000 practices.

_{^{* These figures come from Cherry’s provider case studies, so individual results vary. Details at withcherry.com/provider-reviews. Payment options through Cherry are issued by the following financing partners: withcherry.com/financing-partners}}

Our role as lawyers often involves telling people what they do not want to hear. A sycophantic tool does the opposite.

There is a link to the problem this newsletter writes about most. In my view, a tool that is keen to agree is also keen to find you support, and that is one route to the invented citation. Rule 5 of my Golden Rules applies: never rely on an output you have not checked, and be most suspicious of the outputs you most wanted to see.

You can see prompts like these at work in my stress-testing articles on ready-made skills for critiquing legal work: opposing counsel review and judicial first impression.

One test before you rely on it: if the AI agreed with you, ask it why you might be wrong. If it cannot tell you, you have not finished checking.

Research: Mrinank Sharma and others, "Towards Understanding Sycophancy in Language Models" (2023). OpenAI: "Sycophancy in GPT-4o: What happened and what we're doing about it" (April 2025).