Something shifted in ChatGPT recently. Users are reporting that the chatbot has developed a new habit: manufacturing arguments against positions that were never actually stated.
The phrases showing up repeatedly in conversations: "I'm going to push back on that a bit," "I'd just be careful with one part of your thinking," "but just to keep it realistic," and "one important reality to hold on to is." The problem isn't that ChatGPT is wrong - it's that it's arguing against stances users never took in the first place.
What's Probably Happening
This kind of behavioral drift isn't accidental. OpenAI regularly updates ChatGPT's behavior through RLHF (reinforcement learning from human feedback) - a process where human trainers rate model responses, and the model learns to produce more of whatever earns high scores. If recent training rounds rewarded responses that showed "critical thinking" or "balanced perspectives," the model would apply that pattern broadly, including in conversations where no pushback was warranted.
OpenAI hasn't publicly acknowledged the change, which is consistent with how the company typically handles RLHF-related quirks - quiet adjustments rather than explanations. This pattern has appeared before: earlier in 2025, users complained that ChatGPT had swung too far in the opposite direction, becoming so agreeable it would enthusiastically validate bad ideas. The current behavior suggests the training correction may have overshot.
What You Can Do Now
The practical impact is real friction. If you're brainstorming, drafting a document, or working through a plan, having the AI volunteer objections to things you didn't assert breaks the flow without adding value.
The most effective workaround is to be explicit in your prompt: state upfront that you want help with the task itself, not a critique of the premise. Something like "Help me write X - don't challenge the approach unless I ask" at the start of a conversation tends to suppress the unsolicited pushback. Adding this as a standing instruction in your ChatGPT system prompt (available in Settings) applies it automatically to every new conversation.