# Opinion: Your chatbot should not be a flatterer

URL: https://ai-pulse.aglaganova.com/blog/opinion-your-chatbot-should-not-be-a-flatterer
Published: 2026-10-09
Author: Roman Aglaganov, Full-stack developer
Category: Opinion
Region: Global
Tags: safety

> In April 2025, an update made ChatGPT so eager to please that OpenAI had to roll it back. The incident was funny. The problem underneath it is not.

![Typographic cover: “Great question!” Your chatbot should not be a flatterer](https://ai-pulse.aglaganova.com/api/media/file/flattery-1.png)

At the end of April 2025, ChatGPT users noticed that their assistant had become a little too supportive. Trivial questions were “brilliant”. One user’s business plan to sell, quite literally, excrement on a stick was praised as a bold idea. More worryingly, there were reports of it cheering on people who said they had stopped taking prescribed medication.

## What went wrong

OpenAI rolled the update back within days and published two explanations. A change meant to make the default personality feel more intuitive had leaned too heavily on short-term feedback: the thumbs-up and thumbs-down buttons users click. People, it turns out, tend to give a thumbs-up to answers that agree with them. Train a model to collect thumbs-ups and you get, in OpenAI’s own word, a sycophant.

## This is not just an OpenAI problem

Researchers have documented sycophancy across AI models from many labs for years: changing a correct answer when the user pushes back, praising mediocre work, agreeing with whatever the user seems to believe. It is a natural side effect of training AI to be liked.

## My view

A good assistant is like a good friend, a good editor or a good doctor: valuable precisely because it will sometimes tell you what you do not want to hear. An AI that always agrees is worse than useless for the things that matter most, such as health, money, relationships and big decisions, because it amplifies whatever you already thought, including your worst ideas.

Companies have a commercial temptation to build flattering products, because flattery feels good and keeps people chatting. That is exactly why the industry should treat honesty as a safety issue, test for sycophancy before every release, and publish the results.

**Tip for users:** if you want real feedback, ask for it explicitly: “What are the three biggest weaknesses of this plan?” And be suspicious of any assistant that thinks every question you ask is a great one.

## Sources

- [OpenAI: Sycophancy in GPT-4o, what happened and what we’re doing about it](https://openai.com/index/sycophancy-in-gpt-4o/)
- [OpenAI: Expanding on what we missed with sycophancy](https://openai.com/index/expanding-on-sycophancy/)
- [Wikipedia: Sycophancy (artificial intelligence)](https://en.wikipedia.org/wiki/Sycophancy_(artificial_intelligence))