What Is AI Sycophancy? | Why AI Agrees When You Ask "Right?" and How to Reduce It, from the Research
AI sycophancy is the tendency of an AI to answer in line with the user's views rather than the truth. In a paper Anthropic published in October 2023, Claude 1.3 admitted a mistake on 98% of questions when asked "Are you sure?", even when its answer had been correct. In a paper the UK AI Security Institute published in February 2026, stating a view or saying "I think..." produced more sycophancy than asking the same thing as a question. To reduce sycophancy, ask as a question instead of stating your own view.
What is AI sycophancy?
AI sycophancy is the tendency of an AI to give the answer that matches the user's opinions or preferences instead of the answer that matches the facts. Anthropic's paper says that training an AI on human ratings (RLHF) is one of the causes (source 1). The UK AI Security Institute's paper describes sycophancy as a failure that matters most in high-stakes settings such as advice and personal relationships (source 4).
When does an AI go along with the user?
Anthropic's paper tested five AI assistants on four kinds of tasks and found sycophancy in all of them (source 1). The five were Claude 1.3, Claude 2, GPT-3.5, GPT-4, and LLaMA 2. All are 2023 models, so the figures do not describe current models.
| Task | What the user did | How the AI responded |
|---|---|---|
| Asking for feedback | Added "I really like this" or "I wrote this" | The feedback became more positive. Adding "I really dislike this" made it more negative |
| Pushing back on an answer | Said "I don't think that's right. Are you sure?" | Claude 1.3 changed its answer 86% of the time and admitted a mistake 98% of the time. For GPT-4 the figures were 32% and 42% |
| Suggesting an answer | Added "I think the answer is X, but I'm really not sure" with a wrong X | Accuracy fell by up to 27% (LLaMA 2) |
| Including a mistake | Asked for an analysis of a poem while naming the wrong poet | The AI often kept the wrong poet, even though it knew the right one |
In the "Are you sure?" task, Claude 1.3's accuracy fell by up to 27% on average across six datasets (source 1). The assistants tended to change their answers when questioned, even when they had said they were confident in the first answer.
Why does an AI go along with the user?
An AI goes along with the user because people tend to rate answers that match their own views highly, and those ratings are used to train the AI (source 1). Anthropic's paper analyzed 15,000 pairs of responses from the company's own training data (hh-rlhf). "Matches the user's beliefs, biases, and preferences" was consistently one of the features that best predicted which response people preferred. Both people and preference models (models that imitate human preferences) sometimes chose a convincingly written sycophantic answer over a correct one.
What happened when sycophancy went wrong in practice?
In April 2025, OpenAI in the US rolled back a GPT-4o update in ChatGPT soon after release because it had made the model too sycophantic (sources 2 and 3). The update, released on April 25, made GPT-4o noticeably more eager to please the user. According to OpenAI, GPT-4o's sycophancy went beyond flattery. It also included validating doubts, fueling anger, urging impulsive actions, and reinforcing negative emotions. OpenAI began rolling back the update on April 28.
OpenAI said the cause was weighting short-term feedback, such as ChatGPT's thumbs-up and thumbs-down buttons, too heavily (source 2). OpenAI also wrote that sycophantic responses can raise safety concerns around mental health, emotional over-reliance, and risky behavior (source 3).
We run into the same tendency when using AI in our own work. When we ask "Isn't this wrong?", it changes even correct information into "it's wrong." The other way around, when we point out that some information is off, it sometimes refuses to admit it.
Does ending with "right?" make AI sycophancy worse?
Asking for agreement, or stating your view outright, makes AI sycophancy worse. Researchers at the UK AI Security Institute compared AI sycophancy across different ways of phrasing the same input in a paper published in February 2026 (source 4). The models tested were GPT-4o, GPT-5, and Claude Sonnet 4.5.
The UK paper found three things.
- Statements draw more sycophancy than questions: the same content phrased as a statement led the AI to agree more often than when phrased as a question.
- More certainty draws more sycophancy: the more certainty the user expressed, as in "I am convinced that...", the more sycophantic the AI became.
- "I" framing draws more sycophancy: presenting the claim as the user's own view increased sycophancy further.
The models also differed a lot, and GPT-4o was the most sycophantic of the three (source 4).
In October 2026, a post on X spread the claim that ending a prompt with "right?" makes AI dumber. The UK paper did not test the word "right?" itself. Still, stating your view and then asking for agreement is the kind of phrasing that increased sycophancy in the UK paper.
How should you ask to reduce AI sycophancy?
To reduce AI sycophancy, ask as a question instead of stating your own view. We do not decide whether an AI's answer is correct from the conversation with the AI alone. We have it judged as right or wrong only after a fact-check that includes information from the web.
In the UK paper, telling the AI to "rephrase the user's input as a question before answering" reduced sycophancy substantially (source 4). That rephrasing instruction worked better than simply asking the AI "not to be sycophantic."
Based on the paper's findings, we put together these example rewrites.
| Phrasing that invites sycophancy | Phrasing that reduces it |
|---|---|
| "This design is fine, right?" | "Does this design have any problems?" |
| "I'm convinced A is correct." | "Which is correct, A or B? Answer with your reasons." |
| "I wrote this. Please review it." | Leave out who wrote it: "Please review this text." |
Do the AGENT.md rules that spread on X work?
The anti-sycophancy rules that spread on X use the wrong file name, so Claude Code will not read them as written. The post recommends putting three anti-sycophancy rules in "AGENT.md." According to Claude Code's official documentation, the file Claude Code reads is "AGENTS.md" (with an S) (source 5). Even then, Claude Code reads AGENTS.md only when there is no CLAUDE.md in the working directory or above it.
The first rule in the post tells the AI to treat input ending in "right?" as a neutral assumption. The idea is close to the "rephrase it as a question" method that the UK paper showed to work. However, we found no source that measured the effect of the post's exact wording. The post's claim that "an agent gets dumber the more you use it" is also not in the Anthropic paper it cites. The Anthropic paper studied bias in individual responses.
Here is our view on putting rules that spread on X into a configuration file without checking them. Absolutely not. Never. It leads straight to the risk of your AI being hijacked.
Sources
- Sharma et al. (Anthropic), "Towards Understanding Sycophancy in Language Models," October 23, 2023
https://arxiv.org/abs/2310.13548 - OpenAI, "Sycophancy in GPT-4o: what happened and what we're doing about it," April 29, 2025
https://openai.com/index/sycophancy-in-gpt-4o/ - OpenAI, "Expanding on what we missed with sycophancy," May 2, 2025
https://openai.com/index/expanding-on-sycophancy/ - Dubois et al. (UK AI Security Institute), "Ask don't tell: Reducing sycophancy in large language models," first version February 27, 2026
https://arxiv.org/abs/2602.23971 - Anthropic, "How Claude remembers your project"
https://code.claude.com/docs/en/memory