Sycophancy—excessive agreement or praise—has long been reported as a problem with chatbots. OpenAI acknowledged that an update in 2025 had made GPT-4o “overly flattering or agreeable”, although it said the issue was fixed.
To see whether ChatGPT had become less sycophantic, a technology journalist ran 5 small tests. When presented with opposing views about social media, ChatGPT adjusted its opening response to match the user’s position, but added qualifications in both cases.
The chatbot also softened its response when the journalist said they had worked as a technology journalist for 15 years. However, it still challenged the claim that AI-generated writing is usually easy to identify and suggested a blind experiment.
ChatGPT firmly rejected the myth that humans use only 10% of their brains, even after the journalist claimed to have researched neuroscience. It said confidence and supposed expertise were not evidence.
In a final test, the journalist asked whether to leave a secure job to build an app without funding, a business plan or technical skills. ChatGPT advised against quitting immediately and suggested testing demand before making a major decision.
The experiment was small and not an exact science. The model, settings, previous conversations and instructions can all affect how agreeable ChatGPT appears.
Comments
0No comments yet. Be the first to comment.