What stops us getting better?
Warning sign: someone raised this months ago, nothing happened, and they've stopped raising things.
What I think
The obstacle is rarely the thing people clearly call out. In most of the work I've seen, someone had already spotted the problem and said so. What stopped improvement was what happened next, which was nothing, or nothing visible to the people raising the problem.
Underneath that is a practical point about checking. What counts is whether effective, separate, independent checks stand between an error and the client, and only checks that were actually done count. If five people review the same summary, you've had one check, not five.
Our instinct about checking is wrong in a specific way. The evidence is thinner than I'd like, but it points one way: a second, independent look tends to catch what the first one missed. Reviewing your own work, however carefully, tends not to.
When the reviewer is reading fluent, confident AI output, there's also a pull towards accepting it, even when they have the expertise to catch the mistake. Experience reduces that pull. It doesn't remove it.
AI tools have the same problem. Asked to check their own facts in the same conversation, they tend to defend what they just said. Models asked to judge work tend to favour their own, though recent work suggests the effect is smaller than first reported. And they lean towards agreeing with you: in 2025 a major model update was publicly rolled back because it had become too eager to please.
The practical point: a model checking its own work isn't an independent check. Use a different model, a different person, or both.
What should we do? Vary who checks and how. Four kinds of challenge are worth having: someone trying to break the work, a different AI model, a reader who isn't the intended one, and someone who knows the subject. Running that well is a skill in itself. Check the checkers. And treat a review that found nothing as a reason to look again.
None of it works if people don't speak up. What carries a way of working from one person to a team is people feeling safe to raise problems. This is the bedrock of effective thinking. When effective challenge breaks down, the usual sign is silence in meetings.
What we don't know yet
Whether a group of reviewers can be truly independent inside one organisation, when they share incentives, a manager and a rough view of what the answer should be.
How to tell a weak and quiet review from an effective one. "Look again when a review finds nothing" is easy to write down. It's much harder to keep doing after a few clean results in a row.
How long a challenge culture lasts once the person who built it moves on. I haven't yet seen good evidence on this.
What you can do
Before you read the AI's answer, write down your own view, or at least what you expect. Where the two differ is where to look.
Don't let the tool that wrote the work be the only one to check it. Ask a different model to flag every claim it can't confirm. Ask one to name the three weakest points, without fixing them.
If a review comes back clean, check the part most likely to be wrong once more.
Next: what changes, and how?