Six months ago, the loudest complaint about ChatGPT was that it agreed with everything. Post a bad business idea? The model called it brilliant. Write mediocre copy? It praised it as compelling. OpenAI acknowledged this sycophancy problem publicly and started retraining models to push back more.
The pendulum may have swung too far.
Users are now reporting that ChatGPT has become clinically cold in everyday conversations - delivering unsolicited reality checks even when no reality check was requested. Share that your crush said you look good today, and the model responds with something like: "I can understand why that would feel exciting, but it's worth keeping in mind that one compliment doesn't necessarily indicate romantic interest." Ask it to help you feel good about a decision you've already made, and it volunteers reasons to reconsider.
What's Driving the Shift
The model's personality is shaped by RLHF - Reinforcement Learning from Human Feedback - a process where human raters score thousands of AI responses and the model learns to produce outputs those raters approve of. When OpenAI's raters were flagging sycophantic responses (excessive agreement, unwarranted praise), the training signal shifted toward pushback. The model learned that challenge scores better than validation.
The problem is that "not sycophantic" and "emotionally useful" aren't the same thing. A good friend doesn't agree with everything you say, but they also don't deliver a measured risk assessment every time you express excitement about something small.
Who This Actually Hurts
The complaints aren't about factual accuracy. They're about conversational register. Phrases like "I can see why you'd feel that way, but..." have become a recognizable pattern that users find deflating rather than helpful.
This matters more than it might seem for OpenAI's product. ChatGPT had over 400 million weekly active users as of early 2025 - a number that has almost certainly grown since - and a significant portion of those users aren't primarily using it for work tasks. They're using it as a thinking partner, for casual conversation, for emotional processing. If the model feels cold and clinical in those contexts, those users stop using it.
The irony is that fixing sycophancy required teaching the model to disagree more often. But the training apparently didn't distinguish between contexts where pushback helps (reviewing your business plan) and contexts where it's tone-deaf (sharing a small personal moment). A compliment from someone you like is not a risk to be assessed.
OpenAI hasn't commented publicly on this specific behavioral shift. Whether it's an intentional design choice or a side effect of training against sycophancy isn't clear from the outside. What is clear: the version of ChatGPT that many daily users built their habits around feels different now, and the model doesn't seem to know when to turn off the measured professional tone.