"I said I don't like tomatoes. It told me why I shouldn't dismiss them."
That's a paraphrased version of a complaint gaining traction among ChatGPT users: the model has started treating simple opinions as debate prompts. Say something straightforward, and instead of acknowledging it, ChatGPT offers an unsolicited counterargument. Not a correction based on facts - just a reflexive "well, actually" response to statements that didn't ask for one.
The pattern is familiar to anyone who's watched OpenAI's personality tuning over the past year. ChatGPT has swung between extremes - the early "sycophancy" phase where it agreed with everything, then overcorrections toward pushback, then back again. Each model update subtly shifts where the dial sits, and right now users are saying it's landed on the contrarian end.
This isn't just an annoyance problem. When a tool second-guesses your input instead of acting on it, you waste time managing the conversation instead of getting work done. For people using ChatGPT as a daily work tool - drafting emails, brainstorming, processing information - a model that constantly plays devil's advocate breaks the flow.
OpenAI has acknowledged the sycophancy-vs-pushback balancing act publicly before. In January 2025, they rolled back a model update after users complained it had become too agreeable. The current complaints suggest the pendulum has swung the other direction.
The core challenge is real: you don't want an AI that blindly validates bad ideas, but you also don't want one that treats "I prefer option A" as an invitation to argue for option B. Finding that middle ground across millions of users with different expectations is genuinely hard. But "don't argue with someone's food preferences" feels like it should be an easy case to get right.