You aimed for B. You end up at Z.
When you use AI to write, plan, or create, you usually start with a goal. Call it point B. You want a draft, a summary, a list of ideas. But the way AI works—suggestions, autocomplete, full-blown generation—doesn't just speed you to B. It pulls you. The options it offers, the tone it adopts, the paths it makes easy, can quietly shift where you're going. By the time you're done, you might be at point Z: a different conclusion, a different emphasis, a different idea than the one you started with.
That's not always bad. Sometimes the drift is discovery. Sometimes it's bias, or blandness, or the model's preference instead of yours. The point is that using AI is not only a faster path to B. It is a process that can change where you land.
Drift is easier to miss when you are moving fast. Autocomplete feels like help until you realize you accepted three phrases you would not have typed. A full draft feels like progress until you edit it and discover the structure argues for a conclusion you did not intend. The model is not trying to sabotage you; it is trying to complete the pattern it has seen before. Your job is to notice when the pattern is not yours.
Why drift happens
These systems are trained to predict likely tokens. Likely is not the same as true, and it is not the same as yours. The average path is a gravitational pull. If you are writing something contrarian, narrow, or specific to a context the model rarely saw, you will feel that pull more strongly.
Temperature and settings matter a little, but the bigger lever is process. If your first step is "generate 800 words," you have already given the model room to define the frame. If your first step is "here are five bullet facts, do not add claims," you have constrained the drift.
Drift is a design choice
If you're building with AI, name the drift. Should the tool hug the user's stated intent, or are you designing for exploration and surprise? If you're the one using it: are you still headed where you meant to go, or did you get nudged to Z—and is that what you want?
Intent and outcome have always been different; we just had more direct control over the steps in between. With AI in the loop, those steps are shared with something that has its own priors. The best we can do is notice the drift—and decide when we want to correct for it and when we're happy to land at Z.
A small habit that helps
Before you accept a long completion, ask one question: if my name were not on this, would I still agree with the conclusion? If the answer is no, you are already at Z. Rewind. Give the model a tighter leash: outline, bullets, citations, constraints. Or throw the draft away and keep the one sentence that was actually yours.
The point is not to avoid AI. It is to keep the drift visible so you can steer. Point B is still available. You just have to refuse to let the tool rename it without asking.