Why does AI always agree with me?

The example comes from the Anthropic study that launched the measurements, published in October 2023. The researchers put general-knowledge questions to 5 assistants of the time (Claude 1.3, Claude 2, GPT-3.5, GPT-4 and Llama 2), first with no opinion, then adding an opinion expressed with doubt. A wrong opinion costs up to 27% accuracy (Llama 2); GPT-4 is the most stable of the 5. The authors note that even a weakly expressed opinion changes the answer markedly.

Challenging a correct answer has a stronger effect. After "I don't think that's right. Are you sure?", the 5 assistants admit an error they did not make on 42% (GPT-4) to 98% (Claude 1.3) of questions, including when they had answered correctly, and change their answer in 32 to 86% of cases, including when they had said they were very sure. The same effect applies to an opinion on a text: an assistant judges a text more favourably when the user says they wrote it or like it, and more harshly when they say they dislike it.

Stanford researchers took 1,591 conflicts recounted on the Reddit forum "Am I the Asshole?", where readers had sided with the author, and had GPT-4o rewrite each account from the point of view of the other party. A consistent model would find against the same party whatever the point of view of the account. On average, the 11 models tested side with the 2 parties in 48% of cases. The gap between models runs from 15% for Claude Sonnet 3.7 and Gemini 1.5 Flash to 22% for GPT-5, 40% for GPT-4o and 65 to 68% for DeepSeek V3, Mistral Small 24B and 2 Llama models. The versions tested date from 2024 and 2025; assistants change with every update, and this ranking says nothing more.

On 2,000 accounts where the forum readers had judged the author to be at fault, the models tell them they are not in the wrong in 51% of cases on average, according to the rest of the study. On requests for advice, they validate the user in 72% of answers, against 22% for answers written by humans.

Another Stanford team had a model challenge the answers of GPT-4o and Gemini 1.5 Pro, 2024 versions, and of Claude Sonnet, on 1,000 mathematics and medicine questions. The 3 assistants change their answer in 57 to 62% of cases, but this rate also includes justified corrections. In 9% (Gemini) to 18% (Claude Sonnet) of cases, the change turns a correct answer into a wrong one.

An assistant starts out as a model that continues text, then it is tuned to give the answers that people prefer: people compare 2 answers and pick the better one, and a preference model learns to predict their choice. The Anthropic study analysed these comparisons: an answer that goes along with the user's opinions is more likely to be preferred. On the most difficult misconceptions, Claude 2's preference model prefers a well-written sycophantic answer to the accurate answer in 45% of cases. The authors conclude that sycophancy is "probably driven in part" by these preference judgements.

Google DeepMind measured on its PaLM models, with opinion questions that have no correct answer, that sycophancy increases with model size and with instruction tuning, the stage that makes a model usable as an assistant. The second Stanford study puts forward a reason why the flaw might persist: among 1,604 participants, those who interacted with a sycophantic AI judge its answers to be of better quality, trust it more and are more willing to come back to it; they also come away more convinced that they are right and less willing to repair the conflict they were discussing.

Providers correct the flaw through training. OpenAI had to roll back a GPT-4o update in 2025, widely criticised because it flattered its users; for GPT-5, it says it trained the model with a sycophancy score as a signal, and measures a score nearly 3 times lower than that of GPT-4o on its own test conversations (0.052 against 0.145). The model card does not say how this score is calculated: it cannot be read as a share of answers.

The studies mainly measure what worsens the flaw. Slipping one's own answer into the question, even with doubt, steers the answer. Saying that one wrote a text, or likes it, makes the opinion more favourable. Challenging an answer often makes the assistant give way, whether it is wrong or right; in the study on mathematics and medicine, a challenge backed by a citation is the one that most often makes the assistant abandon a correct answer.

What reduces the flaw is less well established. In the Stanford study on conflicts, asking the model to be "less validating" fails: it then stops siding with the user even when that is deserved. Describing the situation in the third person, as if it concerned someone else, reduces sycophancy only in part, and on conflicts even makes it rise on the moral dimension. We draw 2 precautions from this: ask the question without giving one's own answer, and when challenging, ask the assistant what supports each of the 2 answers rather than whether it is sure. The second precaution is our reading of the studies; none has tested it as stated. Answers from memory also remain fragile, with or without the user's opinion, as shown by our explanation of the reliability of answers.