Discussion about this post

User's avatar
Cameron Hedrick's avatar

Very helpful... thank you! This concept of 'agency laundering' (thank you - new term for me) is already in play. A couple of my clients have ER cases where AI produced consequential analysis. It was wrong and unchecked, and things went downhill fast. The associated humans are blaming AI for the ensuing issues. "It was the AI, not me!!" This is very common in human-only environments (sadly), but, as your article notes, the zone of control and associated accountability were easier to see when only humans were involved.

Weird and wild world, isn't it!

Thanks, Shonna...

Bilel Othmen's avatar

"The cue that matters most, though, is a standard." This is the line I keep coming back to. In code review I have started asking one plain question of AI-generated changes: would you be happy to maintain this? It helps because the reviewer holds the output against something, instead of a feeling that it looks fine. Your METR example shows how unreliable that feeling can be. In my view each department needs its own version of that question, written down and owned by someone close to the work, otherwise approval slowly becomes a habit rather than a judgement. You say the fourth question may be the hardest. When a team does manage to write its standard down, who usually ends up owning it: the manager, the reviewers, or the people doing the work?

3 more comments...

Ready for more?