In AI it shows up as silent quality regressions, higher refusal rates, or usage caps that were not in the launch blog post. Users notice when the same prompt gets a worse answer six months later. Distinguishing enshittification from a real safety fix takes evals, not vibes.