Drift

Slow, silent decline. Nothing broke; things moved. Your prices changed, a supplier renamed their forms, the AI company quietly updated the model — and a setup tuned to yesterday's reality no longer quite matches today's.

A van that's still perfectly drivable but hasn't had its tracking checked in two years. Nothing failed. It just isn't going where you point it any more.

Why it matters: AI doesn't fail loudly the way crashed software does — no error, no red screen, just answers getting slightly worse each month while everyone carries on trusting them. So the fix is a smoke detector rather than a repair: scheduled eval runs — the same test questions, re-run on a timer — tell you it's slipping while it's still cheap to correct.

Related: Eval · Golden set