What Is "Drift"?
The same AI model can give noticeably different answers to the same question over time — not because it's learning from you, but because the model behind the name keeps changing.
A friend asked me recently why a chatbot that used to nail a specific kind of question suddenly started fumbling it — she hadn't changed how she asked, but the answers had clearly gotten worse. That's "drift," and it's a real, well-documented phenomenon, not a false impression.
The core reason is that the product name and the underlying model aren't the same thing. When you use "ChatGPT" or "Gemini" or any other assistant by name, you're usually talking to whichever specific model version the vendor currently has running behind that name — and vendors update, fine-tune, and swap the model behind a given product name regularly, often without a visible announcement for smaller changes. The name on the app didn't change; the model answering your question did.
Vendors make these changes for real reasons — improving safety behavior, reducing a cost, fixing a different problem entirely — but a change tuned for one goal can shift behavior on an unrelated task as a side effect. A model retuned to be more cautious about medical questions, for instance, might become slightly more hedging on unrelated technical questions too, as an unintended side effect of the same underlying adjustment.
Drift is different from a chatbot "learning" your preferences over a conversation, which is a separate, intentional feature some products offer and clearly label. Drift is unannounced and applies to everyone using that product, not to you personally — it's the ground shifting under the whole user base at once, not the model adapting specifically to how you talk to it.
There's no user-side fix for drift beyond noticing it and adjusting your expectations — if a task that used to work well starts producing worse results, it's worth considering that the model itself changed before assuming you're suddenly prompting it wrong. Some vendors offer pinned or "legacy" model versions specifically for users and businesses who need consistent behavior over time rather than always getting the latest update.