On the day they started, it was true
Ask anyone on your team whether they check what AI hands them, and they will say yes.
Verification almost never gets abandoned. It thins — a little at a time, on exactly the days it went well.
Watch the mechanism, because it is rational at every step
The first outputs come back fine. Checking each one costs real seconds and returns nothing. Twenty clean results later, the check has produced no value on any occasion anyone can remember, and those seconds start to feel like a tax on a tool that has not yet been wrong.
Nobody makes a decision. The evidence for skipping it accumulates on its own — and every correct output adds to the pile.
Which is why "always verify AI output" does not survive contact with a working week. It fails the same way "always proofread" fails. It names an outcome and leaves the method to willpower, and willpower is exactly what erodes under a run of good results.
Rules do not decay. Habits do. And this is a habit.
The reason you cannot manage this the usual way
A skipped check leaves no trace. There is no artifact, no timestamp, nothing in a log. The work looks identical whether it was verified or waved through, which means the erosion is invisible to every instrument a manager normally has.
So the question worth asking is not whether your team verifies AI output. Ask them and they will say yes, and they will mean it.
Ask when they last found something.
If the answer is a while ago, that reads like good news. It usually is not. It is the exact condition under which the check quietly stops happening — and the day it matters is the day nobody notices it already went.
Which puts weight on one small thing
"When did you last find something" is the only signal available, and it is fragile. It survives only as long as finding something costs the finder nothing.
The moment a caught error reads as an admission — that someone trusted the tool too far, or shipped something that needed catching — people stop mentioning the catches. The checks may well continue. You simply lose the one piece of evidence that told you so, and from the outside that looks exactly like everything being fine.
So the thing to protect is not the rule. It is whether a person can say "the AI got this wrong and I caught it" as a routine remark, in front of other people, with nothing attached to it.
A safe place to build the check before the day it matters: See Practice Runs →
Keep reading: The cost-to-check rule · Trusted AI
This is the thinking behind a Trusted AI Culture.