Built a plan-picker widget last week. AI wrote 90% of the code. I caught 100% of the mistakes that would have shipped.
A few specific ones, because abstractions don’t help:
1The cart links
AI confidently wrote /api/v1.0/cart/packages/xxxxx for every plan. Looked right. Worked locally. It would have shipped a silent 1-year-term default in the cart while the page advertised 36-month pricing — a different price in the cart than on the recommendation card. The fix was a query param: ?term=xx. The AI didn’t know it existed because our internal docs said “term is set on the cart page, not the URL.” I knew it from production. Without the catch, every visitor sees a price change at exactly the wrong moment.
2The sales footer that wouldn’t render
I asked AI to make a “talk to sales” footer more visible without piling on more brand blue. Light grey: invisible. Warmer grey with deeper shadow: still invisible. After two failed iterations I told it: “any option other than blue?” That unlocked it. The right answer was tone inversion — flip the whole bar to dark navy, the same pattern Apple Pay and Stripe Connect use for their “talk to us” footers. AI knew the pattern. It didn’t reach for it because I’d framed the brief as a hue problem (“less blue”) instead of the actual contrast problem.
3The redundancy I had to find myself
When a secondary CTA wouldn’t fit above the fold, AI started shrinking padding. I stopped it and looked. The header said “Find your fit” with a subtitle “Made just for you.” The progress bar said “YOUR MATCH.” Inside the body there was a “YOUR MATCH” eyebrow above the plan name. Same idea, three times. AI never flagged that. About 50px reappeared the moment I deleted the duplicates.
The pattern is consistent: AI generates faster than any human can. It also makes specific, plausible-looking mistakes — wrong params, wrong framing, missed redundancies — that another AI won’t catch, because the output looks correct in isolation.
An experienced human plus AI, in tight feedback loops. AI alone is a sure-shot fail on any high-stakes surface — not because it’s dumb, but because it has no taste for what’s actually wrong, only for what looks right.
I’m building more AI-assisted surfaces this year. None will ship without a human on the validation pass.
What’s the last AI-generated thing you caught a real mistake in — that the AI was sure was correct?