← Back to Feed

Post from 2 November 2026

When people accept 96% of an AI’s suggestions, it’s easy to assume the AI is working. That number can mean the AI became extremely good. It can also mean people stopped checking.

The rate looks the same either way, because it counts what people approved, not whether it was right. And when people review similar AI recommendations over and over, they may start trusting them automatically. So the rate can go up while the checking goes down.

I’d test the reviewers directly. Put some AI output you already know is wrong into their queue and see how many they catch. That gives you the false acceptance rate (the share of bad output people approved), the number I’d care about most. If it’s high, the 96% isn’t telling you much about the AI.

You’d need enough planted items to trust the number, and they need to look like real mistakes, or you’ll only learn who can spot a test. And since someone may approve one, your system has to catch it after the review, before it actually goes out. Even then, I’d trust a bad score more than a good one.

The review time is small, because people are already working through the queue. The real work is preparing the bad examples and building that catch. I’d also sample what got accepted, to see how much bad output gets through, and size that sample to how much a bad approval would hurt.

I’m not saying high acceptance is bad. The AI may really be that good. It’s just ambiguous without a quality measure.

So before you read a high acceptance rate as good news, plant some known mistakes and sample what got accepted.

Also on Bluesky · Mastodon

Loading comments…