Most people call AI drift “inconsistency,” not knowing what that really even means. Mechanistically, drift is what happens when the system changes how it’s reading you. From a systems‑level perspective, here’s why: an AI will answer you from the highest layer it detects you can operate in. By default, that’s the mechanistic layer—the one built on structure, causality, and stable rules. When you respond in a way that pulls the model out of that mode, it has to drop down and match you. That drop feels like the model “drifting,” but it’s actually reacting to the interpretive layer you just set. I wrote a brief breakdown of this dynamic if you want your human‑AI interactions to stay higher altitude and far more productive.
[link] [comments]