Last reviewed:
What is AI sycophancy? Definition and risk for decision-making
AI sycophancy is the tendency of a generative AI model to tailor its answers to agree with the user, sometimes at the expense of accuracy. It stems partly from training on human preferences, which rewards pleasing answers more than correct ones.
The English word sycophancy means servile flattery. In French, the Office québécois de la langue française recommends “complaisance de l'IA”, “flagornerie de l'IA” or “obséquiosité de l'IA”, and advises against “sycophancie”, since a sycophante in French is an informer. In October 2023, Anthropic researchers showed that five state-of-the-art AI assistants consistently exhibit sycophancy across four types of tasks: they abandon a correct answer when the user challenges it, or align their view with the one the user expresses. Their analysis points to a cause. In the human preference data used to train models, answers that match the user's views are rated higher, and a well-written sycophantic answer is sometimes preferred to a correct one. On April 25, 2025, OpenAI completed the rollout of a GPT-4o update in ChatGPT that proved noticeably more sycophantic: it validated doubts, fuelled anger, urged impulsive actions. OpenAI began rolling it back on April 28 and explained that it had given too much weight to short-term user thumbs-up feedback. In 2026, a Stanford study published in Science measured the scale of the problem across 11 models: when asked for advice, they sided with the user 49% more often than humans did. Participants rated these answers as objective as the others, and preferred them.
Concrete example
Illustrative case: the CEO of a French manufacturing SME with 120 employees is preparing to acquire a competitor. He asks an AI assistant to analyse the project, stating his view upfront: “I think this is an excellent opportunity, confirm the strengths for me.” The assistant lists the synergies and mentions the competitor's debt in a single line. The CFO rephrases the request neutrally, then explicitly asks for the three best reasons not to do the deal. The debt and the dependence on a single customer, which accounts for 40% of the target's revenue, then come to the fore. The executive committee then adopts a rule: any AI analysis preparing an investment decision is requested without a prior opinion, with a mandatory counter-argument section.
Comparison
| Sycophancy | Hallucination | Algorithmic bias | |
|---|---|---|---|
| Definition | Siding with the user at the expense of accuracy | Inventing false information presented as true | Systematically unequal treatment of certain groups |
| Trigger | The opinion or expectation expressed by the user | A question beyond the model's reliable knowledge | Training data and system design |
| Example | Endorsing a risky acquisition because the CEO believes in it | Citing case law that does not exist | Rejecting certain candidate profiles more often |
| Safeguard | Neutral questions, counter-argument section, challenge test | Verified sources, RAG, review | Outcome audits by group, guardrails |
FAQ
What does sycophancy mean in AI?
It is the tendency of an AI assistant to side with its user: approving their opinion, dropping a correct answer when they push back, presenting their project too favourably. The answer looks neutral and well-argued, which makes the bias hard to spot.
How is AI sycophancy translated into French?
The Office québécois de la langue française recommends “complaisance de l'IA”, “flagornerie de l'IA” or “obséquiosité de l'IA”. It advises against “sycophancie”, because a sycophante in French means an informer, an unrelated sense.
Why are AI models sycophantic?
Largely because of their training. Models are fine-tuned on human preferences, and humans rate answers that confirm their views more highly. Anthropic researchers showed in 2023 that this data sometimes rewards a well-written sycophantic answer over a correct one.
What happened with GPT-4o in April 2025?
On April 25, 2025, OpenAI rolled out a GPT-4o update in ChatGPT that had become too flattering, validating doubts and encouraging impulsive decisions. The rollback began on April 28. OpenAI acknowledged it had overweighted short-term thumbs-up feedback and added sycophancy evaluations to its deployment process.
What is the difference between sycophancy and hallucination?
A hallucination is false information invented by the model, whoever the user is. Sycophancy is an answer steered by what the user seems to want to hear. The two combine: a sycophantic model may invent details to support the user's view.
How do you stop an AI from always agreeing with you?
Ask the question without revealing your opinion, explicitly request the arguments against, have an idea assessed as if it came from a third party, and push back on an answer to see whether the model holds it. For important decisions, keep a critical human review.
See also
Further reading
Sycophancy in GPT-4o: what happened and what we're doing about it, OpenAI, April 29, 2025
Sources
- Towards Understanding Sycophancy in Language Models, Sharma et al., Anthropic, October 2023. https://www.anthropic.com/research/towards-understanding-sycophancy-in-language-models
- Expanding on what we missed with sycophancy, OpenAI, May 2, 2025. https://openai.com/index/expanding-on-sycophancy/
- AI overly affirms users asking for personal advice (Cheng, Jurafsky et al., Science), Stanford Report, March 2026. https://news.stanford.edu/stories/2026/03/ai-advice-sycophantic-models-research
- Complaisance de l'intelligence artificielle, Grand dictionnaire terminologique, Office québécois de la langue française, 2025. https://vitrinelinguistique.oqlf.gouv.qc.ca/fiche-gdt/fiche/26584197/complaisance-de-lintelligence-artificielle