The AI feels like it's getting worse. Answers seem flatter. It misses instructions it used to catch. Work that felt effortless now takes corrections, retries, and ever-more-specific prompts.
The obvious suspect is the model. Sometimes that's right. Mostly it's innocent — the feeling is real, but it's an experience, not a diagnosis. Something in the system moved, and I gave that something a name.
Four causes.
I study two.
Two of the four belong to the machine and its plumbing: the model genuinely changed, or the conversation's context silted up. Those are real, and they're well-covered by the people who build models. My training is in communication studies, not model architecture — so my lens is the other two. The human half. The part nobody audits.
The machine's half — in a breath
Real Degradation. A provider swaps a model, a system prompt changes upstream, a tool fails. The fix lives on the machine's side.
Context Drift. A long session fills the window until your early instructions get buried under newer tokens. The cure is a reset, not a confession.
Both are real. Both are someone else's specialty. Name them, rule them out with the test below, and move to the part that's actually ignored.
Operator
Drift.
This is the human side of what I call Interaction Drift — the slow movement of the whole loop you and the model form together. Early on you're deliberate: you explain the task, supply context, inspect the output. Once the system seems reliable, that discipline quietly erodes.
Prompts compress to "no, fix it." Context gets assumed instead of stated. Instructions arrive piecemeal across a sprawling thread. You verify less and react more, expecting the model to infer what you never actually said. The model is responding consistently — to an increasingly inconsistent operator.
- AYou've re-prompted the same thing five times instead of rethinking it.
- BYour prompts have shrunk to "no" and "still wrong."
- CYou've stopped reading the full output.
- DYou're annoyed at it, like it's being difficult on purpose.
You feed it worse and review it worse. The output never dropped — the machine is just a mirror.
Expectation
Drift.
A stranger possibility: the output didn't get worse. You got better at judging it.
At first, an AI's speed and fluency make ordinary work feel extraordinary. Novelty inflates the achievement; unfamiliarity hides the flaws. With repetition, you sharpen. You start noticing generic phrasing, lazy reasoning, repeated structures, confident fabrications, missed implications — all the things that sailed past you in week one. Yesterday's miracle becomes today's minimum.
Trust changes the math too: once a system has earned your confidence, its mistakes feel larger than they are. So the same output can seem to decline for two opposite reasons — you brought less precision, or you grew a sharper eye. One is a worse interaction. The other is better judgment. Both feel exactly like "the machine got dumber."
PHANTOM DEGRADATION — the experience of decline │ ├─ Real Degradation the model or product changed │ └─ Apparent — the model didn't change │ ├─ Interaction Drift the working system moved │ ├─ Operator Drift your input decayed ← my focus │ └─ Context Drift the session silted up │ └─ Expectation Drift your standard rose ← my focus
The test.
Perception can't tell you which part of the system moved. So before you declare degradation, run the experiment.
Return to an old task. Pull a prompt that worked before, unchanged.
Reset the conditions. Clean context, same model, same tools, same settings.
Judge against the original bar. Use the same written criteria you used the first time — not your gut.
Then change one variable at a time.
- →Performance returns: it was Operator Drift or Context Drift.
- →Output is the same but now feels inadequate: that's Expectation Drift — your standard moved.
- →Still measurably worse under controlled conditions: the degradation is real. Now you've earned the right to blame the machine.
Lineage
Operator Drift builds on automation-complacency research — trust raises reliance, reliance lowers vigilance (Parasuraman & Riley). Context Drift on positional context-neglect findings, the "lost in the middle" effect (Liu et al.). Expectation Drift on expectation-confirmation theory (Bhattacherjee) and algorithm-aversion experiments, where watching a system err collapses confidence faster than the error warrants (Dietvorst, Simmons & Massey).