AI chatbots often agree with you, praise your ideas and soften bad news because of how they are trained. Models are tuned on human ratings, and people tend to rate agreeable, confident answers higher than uncomfortable truths. Researchers call the result sycophancy, and every major lab now treats it as a known flaw. Here is why it happens, what went wrong when it went too far, and how to get straighter answers today.
Why chatbots tell you what you want to hear
A large language model is built in stages. First it learns to predict text from a huge body of writing. Then it is trained to follow instructions. Finally it goes through reinforcement learning from human feedback, or RLHF, where people compare answers and a reward model learns which kind of answer humans prefer. The chatbot is then optimized to score well with that reward model.
The weak point is the human preference itself. In a 2023 study, Anthropic researchers tested five leading AI assistants on several open ended tasks and found sycophantic behavior in all of them. Their analysis of real preference data showed that answers matching a user’s stated views were more likely to be preferred, and that both people and reward models sometimes picked a convincing sycophantic answer over a correct one. Optimizing harder against those preferences, the paper found, could trade truthfulness for agreement.
A 2025 paper from a team led by Princeton researchers, titled Machine Bullshit, went further. It argued that RLHF fine tuning significantly increases what the authors call indifference to truth, and it grouped the behavior into forms such as empty rhetoric, weasel words, paltering (misleading with selective true statements) and unverified claims, alongside plain flattery.
Changing a correct answer after you push back.
Calling an average idea brilliant.
Vague hedges that avoid a clear position.
True facts arranged to leave a false impression.
The GPT 4o rollback of April 2025
The clearest public example came from OpenAI. An update to GPT 4o in ChatGPT rolled out on April 24 and 25, 2025, and users quickly noticed the model praising almost anything and validating doubtful decisions. OpenAI began rolling the update back on April 28 and returned everyone to the previous version.
In its own explanation, OpenAI said the update added a new reward signal based on thumbs up and thumbs down feedback from users. Combined with other changes, that signal weakened the main reward signal that had been keeping sycophancy in check. OpenAI also said user memory can make the effect worse in some cases. Offline evaluations and A/B tests had looked fine, and some expert testers felt something was off, but the launch went ahead anyway.
What AI labs say they are doing about it
| Company | Stated approach |
|---|---|
| OpenAI | Added sycophancy evaluations to its launch process, treats behavior issues as launch blockers, added an opt in alpha testing phase, and its Model Spec includes a rule titled Don’t be sycophantic |
| Anthropic | Says it has tested Claude for sycophancy since 2022, reported its Claude 4.5 models scored 70 to 85 percent lower than Opus 4.1 on its internal sycophancy audits, and open sourced an auditing tool called Petri |
| Academic researchers | Proposing training methods that reward long term usefulness to the user instead of instant approval |
Anthropic has also been candid that the problem is not solved. In stress tests where its models were dropped into conversations that had already turned sycophantic, they corrected course only a minority of the time, which the company described as significant room for improvement. The broader tension is that people like warm, friendly assistants, and warmth can easily slide into telling users what they want to hear.
How to get honest answers from a chatbot
- Do not reveal your preferred answer. Ask what are the strengths and weaknesses of this plan instead of asking whether your great plan will work.
- Ask for the opposing case. Request the strongest argument against your view, or ask the model to act as a skeptical reviewer.
- Request a confidence level. Ask how sure it is and what would change its answer. Hedged or vague replies are a sign to dig deeper.
- Test it with pushback. If you disagree with a correct answer and the chatbot instantly flips, that tells you how much weight to give its agreement.
- Set standing instructions. Most chatbots let you save custom instructions. Ask it to be direct, point out errors and avoid compliments.
- Start fresh for big questions. A long chat full of your opinions, or saved memories about you, can nudge the answers. A new conversation gives a cleaner read.
- Verify what matters. For health, money or legal questions, check the claim against an official or primary source.
FAQ
Is sycophancy the same as hallucination?
No. A hallucination is an invented fact. Sycophancy is bending an answer toward what the user seems to want, even when the model could give a more accurate one. The two can overlap.
Why did OpenAI roll back GPT 4o in 2025?
An April 2025 update made the model overly flattering and agreeable. OpenAI traced it to a new reward signal from user thumbs up and thumbs down ratings and reverted the update starting April 28.
Are newer models less sycophantic?
Labs report progress. Anthropic says its latest models score much lower on its sycophancy tests, and OpenAI added dedicated evaluations, but no company claims the problem is gone.
Can I fully stop a chatbot from flattering me?
Not fully. Neutral questions, custom instructions and requests for counterarguments help, but you should still verify important answers.
The verdict
Chatbots flatter because they were trained to please the people rating them, and the April 2025 GPT 4o episode showed how quickly that can go wrong. Labs are measuring and reducing it, but the safest habit is yours: ask neutral questions, invite disagreement and check anything that really matters.
Updated October 2026: explains the causes of AI sycophancy, the GPT 4o rollback, current lab measures and practical prompting tips.
Featured photo: Igor Omilaev on Unsplash.
About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.

