Note · July 2026
The polish trap —
a mistake that was mine.
01
A hand-written engineering survey, turned into a clean report. It did the hard part correctly. So I did what felt natural next: I started polishing the format — making it byte-for-byte identical to the template it had always used.
02
Without noticing, I'd substituted “looks like the old one” for the standard that actually mattered: is it correct, and is it useful? I was pouring effort into the thing that was already fine, and calling it work.
03
They're not the obvious errors that stop you. They're the quiet drift — effort flowing toward polish and away from judgement, because polish is now almost free and judgement still costs something. The output looks better while it gets no truer.
04 · why it spreads
When Anthropic asked 81,000 people what they wanted from AI, the single biggest concern was unreliability — and the resentment underneath it was the “verification tax,” the effort of checking output you were promised would save you time. So the tax mostly goes unpaid. The more fluent and confident the output, the less it gets checked — that's automation bias, and it's been measured for decades. A tool that sounds sure is a tool people stop questioning.
What the trap teaches
Guard the standard before you let the tool near it.
Decide what “good” means — correct, useful, defensible — and write it down before the polishing starts, because once output is instant and beautiful the definition quietly drifts toward “looks right.” Keep the human doing the part that carries the judgement, not just the part that's tedious. And treat fluency as a warning, not a reassurance: the more confident the output looks, the more deliberately someone has to check it. I learned this by getting it wrong on my own work. Telling it against myself is the point — if it caught me, it will catch your team too.
David Kolb advises expertise-heavy organisations on adopting AI without letting effort drift from judgement to polish. This note is written to be forwarded — if it's useful to a colleague, send it on.
Source: Anthropic, “What 81,000 People Want from AI,” March 2026 (unreliability the top concern; the “verification tax”). Automation bias: established finding across the human-factors literature.
One thing you can do this week
Take the last piece of AI output you approved and review it properly now. Time it. Then ask whether you'd have done that under deadline.
What it won't tell you
That tells you what happens when you're paying attention. It doesn't tell you what happens in the normal flow of work.
If this is live for you.
Send me the thing you can't get a straight answer on. I'll tell you whether I'm useful.
Email me →