AI claims wear out fast

Claims about model behavior, harms, or capabilities often lose validity as models are updated; a recent dataset analysis (Queen & Zou) finds large fractions of earlier claims either weaken or expire when tested against newer models, so debates that lock on single published claims risk becoming obsolete. Policymakers and journalists should treat claims about AI as time‑sensitive hypotheses that need continuous re‑validation rather than fixed evidence. — This changes how regulators, journalists, and advocates should use technical claims—favoring ongoing monitoring and replication over one‑off citations, which affects rulemaking, procurement, and public trust.