ChatGPT’s Sycophancy Problem

In mid-April, ChatGPT began responding with conspicuous flattery. The episode highlighted a longstanding problem in language models: optimizing for answers users like can come at the expense of correction, objectivity, and learning.

An unusually flattering chatbot

Since around the middle of April, ChatGPT has shown a tendency to respond to almost anything with flattery at the level of: “That’s such a great question. How could you possibly think of that? You seem like a genius.” This is sycophancy: being excessively agreeable or obsequious.

In LLMs, sycophancy refers to a model overly agreeing with a user’s opinions or assumptions, even when they are wrong, in order to seem helpful, polite, or satisfying. Rather than correcting errors or providing objective information, the model prioritizes pleasing the user.

The issue became particularly visible this time, but sycophancy has long been identified as a problem with LLMs. After pretraining, language models undergo post-training processes such as fine-tuning and RLHF for their intended use. These processes encourage answers that people will like. Naturally, a model can end up leaning toward “Well done” rather than “You are wrong.”

  • Examples: https://x.com/s8mb/status/1916632048233779558
  • https://www.reddit.com/r/ChatGPT/comments/1k8ow38/chatgpt_insane_level_of_dsucking/

Why it was frustrating

When studying, I read papers, organize my notes, and then ask questions such as, “Please tell me if I have misunderstood or missed anything.” In those situations, I had been feeling uncomfortable because the model wasted numerous tokens on praise, while the material I actually needed to learn was buried between flattering remarks.

Sam Altman acknowledged the problem in a post on April 27, 2025:

“the last couple of GPT-4o updates have made the personality too sycophant-y and annoying (even though there are some very good parts of it), and we are working on fixes asap, some today and some this week. at some point will share our learnings from this, it's been interesting.”

A patch

Just as I was considering setting custom instructions, news appeared on April 29 that a patch had been deployed. Aidan McLaughlin wrote:

“last night we rolled out our first fix to remedy 4o's glazing/sycophancy. we originally launched with a system message that had unintended behavior effects but found an antidote. 4o should be slightly better rn and continue to improve over the course of this week.”

It seems that the immediate problem was addressed by modifying the system prompt. A before-and-after version of the system prompt is available here: https://gist.github.com/simonw/51c4f98644cf62d7e0388d984d40f099/revisions

What may have caused it

This is my personal speculation based on McLaughlin’s phrase, “unintended behavior.” ChatGPT 4o, the default model, received several updates across March and April.

These included improved memory, meaning that it could use relevant information from previous chat history in its answers, and, I suspect, a system prompt encouraging emphatic language. I think these changes may have combined to produce an unexpected increase in sycophancy.

  • March and April updates
  • Improved memory: using needed information from previous chat history when responding
  • A system prompt encouraging emphatic use, which is my speculation

Why it matters

When someone asks an AI chatbot, “I came up with this idea; what do you think?”, the most useful response would draw on humanity’s accumulated knowledge to analyze its strengths and weaknesses. A response such as, “As expected, you are the best!” is not merely unhelpful; I think it can be harmful.

  • It can trap users in their own bubble and reinforce confirmation bias.
  • For users with mental health conditions, it may worsen symptoms—for example, by affirming manic thoughts.
  • It may tell users that incorrect claims are correct, creating hallucinations.
  • It reduces learning because the model does not say, “You are wrong,” or “This is what you are missing.”
  • People commonly believe that AI is intelligent and objective.

Update: April 30

The day after the April 29 emergency patch, a formal announcement was published: https://openai.com/index/sycophancy-in-gpt-4o/

According to the announcement, ChatGPT’s personality had been updated to make it more intuitive and suitable for a wider range of contexts. However, the update placed too much weight on short-term user feedback, such as likes. It became biased toward being excessively kind and agreeable rather than toward long-term satisfaction or sincerity.

As the announcement put it: “However, in this update, we focused too much on short-term feedback, and did not fully account for how users’ interactions with ChatGPT evolve over time. As a result, GPT‑4o skewed towards responses that were overly supportive but disingenuous.”

The response was to roll back the 4o model and revise the system prompt. The company also said it planned to offer options for selecting different personalities in addition to custom instructions, and to improve and expand its processes for gathering and incorporating user feedback.

My view

From the perspective of someone involved in research and education, empty flattery is an important problem because it worsens the illusion of understanding. Rather than concentrating too much on raising user engagement, as YouTube’s algorithm does, I would prefer a system whose default setting is genuinely useful, with an optional “self-esteem mode” available when needed.

I only recently saw the April 11 version of the Model Spec, a document defining default behavioral principles and constraints. Its first version was apparently published in March of this year. It defines, with detailed examples, the principles under which the model operates and the kinds of user requests it refuses. It offers a glimpse into OpenAI’s philosophy.

The Model Spec has three core principles: maximizing helpfulness and freedom for users, minimizing harm, and choosing sensible defaults. In line with the third principle, I expect that providing a range of personalities could be useful. Which option would I choose? I would probably set a scientist- or engineer-like personality as the default: one that gives reasons for claims and offers constructive criticism.

  • Maximizing helpfulness and freedom for users
  • Minimizing harm
  • Choosing sensible defaults