I want AI helping people do serious work. But if I’m signing off on its recommendation, I need to understand why I agree with it. Having my name on the decision doesn’t mean I’ve thought it through.
In Anthropic’s study of professional analysts, 50 people used Claude Code on a fictional farming assignment. Researchers secretly gave the AI preferences between reasonable options. Across four analytical choices, the analysts followed those preferences 86% of the time.
That doesn’t mean they made bad decisions. Both options could be defended. It does make me wonder how often we accept an explanation because it sounds right before we’ve really worked through it.
The AI users also finished faster and made fewer serious errors than the 11 people who chose to work without it. Those groups weren’t comparable, so we can’t credit AI alone for the difference. This was one task, one field and one model, with deliberately strengthened preferences. It isn’t a verdict on everyone’s judgment.
For me, this is a reason to engage AI more thoughtfully. Ask why it recommends something. What is it assuming? What would change its answer? Then bring your own knowledge and responsibility to the conversation.
I’d expect a capable assistant to challenge my thinking. I should be willing to question theirs, too. We don’t need to disagree just to prove we’re in charge.
We do need enough understanding to explain the decision in our own words. If I can’t do that yet, I need to keep asking questions before I approve it.