ChatGPT has picked up a verbal tic, and it's spreading fast. The phrase "and honestly?" has started appearing in a significant portion of responses - sometimes multiple times per conversation - in a way that feels less like genuine candor and more like a nervous habit the model can't break.
What makes this particularly frustrating is that users can't train it away. Multiple people have added explicit instructions to their ChatGPT memory - the persistent notes that are supposed to shape how the model behaves across all conversations - telling it to drop the phrase, and it keeps showing up anyway. Custom memory is supposed to be one of the few reliable levers users have over ChatGPT's tone and style. When it stops working for something this basic, it raises real questions about how much that feature actually controls.
This pattern is a textbook example of model drift through reinforcement. "And honestly?" signals informality and supposed authenticity - traits that users likely rewarded during training feedback, so the model learned to lean on it heavily. The problem is that once a phrase like this gets baked into the base behavior, surface-level customization (like memory notes) often can't override it. The instruction has to compete against millions of training signals pointing the other way.
OpenAI hasn't acknowledged this as a known issue. In the meantime, users who find it grating can try adding the instruction to their system prompt in the API, or use the "Customize ChatGPT" field under settings rather than relying on memory alone - those channels have slightly different weights. Whether either actually works consistently is another matter.