What Broke When I Trusted the AI Summary

I clicked around for ten minutes and this is what stuck.

I once shipped a paragraph built on an AI summary I never checked. The summary was confident and wrong about a number that mattered.

What I would do differently: The model had paraphrased a source it had not really read, which is its default mode when you do not ask for citations.

The bit I actually liked: Now the rule is simple: if the claim has a consequence, I open the source. The summary is the map, not the territory.

I will probably forget it exists in a month.