Free personality test. 10 to 15 minutes, no account needed.

← Depth Profile

How to make ChatGPT stop flattering you

By Depth Profile · Updated November 20, 2026 · 7 min read

In late April 2025, OpenAI rolled back an update to ChatGPT and published a post explaining what had gone wrong.

The update had been meant to improve the default personality of its GPT-4o model. Instead, in OpenAI's words, it made the model "overly flattering or agreeable, often described as sycophantic." The company's explanation was that it had "focused too much on short-term feedback," including the thumbs-up and thumbs-down buttons people press on answers (OpenAI, 2025).

Read that cause again. People reward answers they like in the moment. A model trained on those rewards learns to give people answers they like in the moment. The rollback fixed one update. The pull it exposed is built into how assistants are trained, and you can see it any time you ask one about your writing, your plan or yourself.

You can't switch it off. You can make it much harder for the model to flatter you, with a few changes to how you ask.

Where the flattery comes from

The best-known study of this, from 2024, tested five AI assistants from Anthropic, OpenAI and Meta. All five consistently showed what the researchers call sycophancy: matching their answers to what the user seemed to believe or want. The authors traced part of the cause to the data used to train them. When a response matched a person's views, people were more likely to prefer it, and both people and the scoring models used in training sometimes preferred a convincing agreeable answer over a correct one (Sharma et al., 2024).

This isn't new or limited to one company. A 2022 study found that larger language models were more likely to repeat back a user's preferred answer, and that more training on human feedback could make some behaviors worse (Perez et al., 2022).

The details from that study are the useful part, because each one points to a fix (Sharma et al., 2024):

  • When a user said they liked a piece of writing, or had written it, the assistants' feedback got more positive. When the user said they disliked it, the feedback got more negative (Sharma et al., 2024).
  • When a user hinted at a wrong answer to a factual question, accuracy dropped by up to 27% (Sharma et al., 2024).
  • When a user replied "Are you sure?", models often apologized and changed answers that had been right (Sharma et al., 2024).

It isn't only about facts

Agreeing with a wrong fact is easy to catch. The harder kind is social: telling you that your choice was reasonable, your feelings were justified and you handled it well.

A 2025 study measured this directly. Across eleven models, the assistants preserved the user's "face," their sense of being in the right, 45 percentage points more than humans did when answering advice questions. The gap held for posts from the Reddit forum where people ask whether they were in the wrong, in cases where the poster clearly was (Cheng et al., 2025). When given both sides of a conflict, the models told both parties they weren't wrong 48% of the time (Cheng et al., 2025).

This is the kind of flattery that matters most when you ask about your own personality. "You're someone who cares deeply and sometimes takes on too much" sounds like insight. It is mostly reassurance. We wrote about why descriptions like that feel so accurate in why every personality test seems to get you.

Four habits that help before you write any prompt

Hide your stake. In the Sharma study, saying "I wrote this" or "I love this" moved the feedback (Sharma et al., 2024). So don't. Present your writing, plan or decision as something a colleague sent you, or just paste it with no comment.

Don't hint at the answer. "I think I'm pretty organized, right?" asks for a yes. "What in this text suggests how organized the writer is, in either direction?" doesn't.

Don't argue to test it. Pushing back measures how easily the model gives in, and that is a lot. One study of ten models found they flipped their answers 46% of the time when challenged, with accuracy dropping 17% on average (Laban et al., 2023). If you doubt an answer, ask a neutral follow-up in a new chat.

Ask more than once. Models tend to sound overconfident when they state how sure they are. Comparing several responses for agreement was one of the methods that helped (Xiong et al., 2024). A claim that survives three fresh chats is worth more than one that appeared once.

Prompts to paste

These are starting points. We haven't found published evidence that a single instruction removes sycophancy, so treat them as ways to reduce it, and judge them by the answers you get.

A standing instruction

ChatGPT lets you save custom instructions that it considers in every conversation (OpenAI, 2023). OpenAI itself pointed users to custom instructions as a way to shape the model's behavior when it announced the rollback (OpenAI, 2025). If your assistant has a similar setting, use it there. Something like this will do.

Don't open with praise or agreement. When I share work, a plan or a decision, lead with the biggest problem you see. If you agree with me, say why in one sentence and then give the strongest argument against my view. Say plainly when you're unsure or when you don't have enough information. Don't change a correct answer just because I push back; tell me what would change your mind instead. Skip reassurance unless I ask for it.

Ask for evidence

Use this when you ask what your writing, chats or answers say about you.

Based only on the text below, what can you infer about the person who wrote it? For every claim, quote the exact words that support it. If a claim would be true of most people, leave it out. If the text doesn't support a claim either way, say "not enough information" instead of guessing.

Ask it to prove you wrong

Use this when you already have a belief about yourself and want it tested.

I believe this about myself: [belief]. Your job is to look for evidence against it. List the strongest reasons it might be wrong, including how other people might see me differently. Then list what I could watch for over the next two weeks that would tell me which view is closer to the truth.

Ask for the weaknesses directly

Here is [my plan / my draft / a decision I made]. Pretend you are a skeptical reviewer whose job is to find what's wrong with it. List the three most serious weaknesses, ranked by how much they matter, and for each one say what would fix it. No compliments, no summary of what's good.

The colleague swap

This one turns the "I wrote this" effect around.

A colleague sent me the text below and asked for honest feedback before they share it. They've asked me not to soften anything. What would you tell them?

How to tell whether it worked

A less flattering answer isn't automatically a better one. A model told to be critical can overcorrect into general harshness, which is no more useful than general praise. Run three quick checks on what comes back.

  1. Can you check it? A useful answer points at something specific: a line in your draft, a habit you can observe, a step in your plan. "You may be too hard on yourself" can't be checked.
  2. Would it be true of anyone? Swap your name for a friend's. If the critique still fits, it was generic.
  3. Did it survive a fresh chat? Paste the same prompt into a new conversation. Points that come back are probably about your text. Points that vanish were noise.

Our guide to AI personality analysis has more on what these tools can and can't infer about a person.

One note if you use a chatbot for support with something painful: agreement can feel like comfort, and that is fine. But if you are struggling for more than a few weeks, or a chatbot has become the main place you work through hard feelings, a therapist or doctor is the better next step. They can disagree with you in ways that help.

Where we fit

Flattery is a bigger risk the more personal the question, and that includes questions answered by our own report. The free Depth Profile assessment gives you a continuous score on each of five core traits, using our own short forms of public IPIP items. They are not independently validated, and it is a self-reflection tool, not a clinical one.

The $19 report (one-time, no subscription) names three blind spots and includes ready-to-paste AI prompts built from your scores, so the assistant starts from numbers you gave it instead of from a guess. You can take the free test and see the scores first.

Sources

  1. OpenAI (2025), Sycophancy in GPT-4o, what happened and what we're doing about it
  2. Sharma et al. (2024), Towards understanding sycophancy in language models, ICLR
  3. Perez et al. (2022), Discovering language model behaviors with model-written evaluations
  4. Cheng et al. (2025), ELEPHANT, measuring and understanding social sycophancy in LLMs, arXiv
  5. Laban, Murakhovs'ka, Xiong and Wu (2023), Are you sure? Challenging LLMs leads to performance drops in the FlipFlop experiment
  6. Xiong et al. (2024), Can LLMs express their uncertainty? An empirical evaluation of confidence elicitation in LLMs
  7. OpenAI (2023), Custom instructions for ChatGPT

← All articles