LANDMARK · 12 MIN · JUDGE
Reality Lab: evidence
After this landmark, you can turn “this reads well” into “every claim here is backed,” by walking a draft claim-by-claim against its sources and fixing what isn’t supported.
This is a hands-on lab, not a lecture: you take an AI-generated draft that sounds authoritative and you audit it. Fluency is not evidence — a model produces confident sentences whether or not the underlying facts are real, and the failures hide inside the polished ones. The method is slow on purpose. Break the draft into individual claims. For each, ask what would make it true, then go find that: a named source, a date, a figure you can trace. Claims that check out, keep. Claims with no source you can find, cut or clearly mark as unverified — don’t leave them wearing the same confident voice as the verified ones. What you produce is a repaired draft plus a short note of what you couldn’t confirm. Doing this a few times rewires how you read every AI output afterwards.
This is the canonical concept. It stays the same across learner lenses so personalization never changes the underlying facts.
◌ EVERYDAY LENS
What this looks like for you
Before you forward or act on an AI summary, pick the two or three claims that actually matter and check each against a real source. If you can’t find one, treat the claim as unconfirmed rather than true.
MAKE A DECISION
Auditing an AI draft, you find a specific statistic with a plausible-looking citation, but the linked source doesn’t contain that number. What’s the right repair?
CARRY THISTake one AI-generated draft, split it into claims, and produce a repaired version with every unverified claim cut or explicitly marked.