There's a name for what you did with that AI answer you forwarded this mornin...
There's a name for what you did with that AI answer you forwarded this morning. Meat proxy. You read the prompt, not the reply.
And the uncomfortable part isn't that you got lazy. The model got right often enough that you quietly stopped checking.
That's how it gets you. Not on the days it's obviously wrong (you catch those). It gets you on day 200, when it's been fine 199 times and you've slid from reviewing the output to nodding at it. Pilots and radiologists have a name for it: automation complacency. The better the automation, the less you watch it, right up until the one that mattered.
And "the AI said so" is not a place to hide when it's wrong. Your name's on the email. Not the model's.
So how do you stay in the loop without pretending a glance is a check? Four things that actually work (willpower isn't one of them, and it loses this fight every time):
Make your own words the deliverable. If the job is "forward it," you'll rubber-stamp it. If the job is "say it in your own words and add the one thing the model couldn't know" (the client's history, the number that isn't online, the landmine only you saw coming), you can't fake that. Your words are the receipt that you actually read it.
Hunt for the error instead of blessing the answer. "Where is this wrong?" catches what "looks good to me" never will. You read differently when you're looking for the crack.
Triage by what it costs you to be wrong. You can't deep-check everything, and pretending you do is exactly how you go blind. Skim the low-stakes summary on purpose. The price, the client deck, the contract clause. Those get real eyes. The mistake was never rubber-stamping. It was rubber-stamping the thing that mattered.
Be honest about what you've actually automated. If you've genuinely stopped reading a workflow, that's not validation with extra steps. It's a decision to hand it over. Own the decision, or put the friction back. Don't tell yourself you're reviewing something you're really just approving.
AI didn't replace your judgment here. It just made it easy to skip. The value you add is still the part where you actually read it.
What's the workflow you've quietly stopped checking?