A team of researchers has built a formal system that treats the right to publish an AI-assisted claim as a single-use permit, revocable the moment the underlying evidence doesn't match what readers actually see.
The paper introduces "Publication Authority" - a one-time, non-transferable permit that must be freshly earned before an AI-assisted claim can be published. Their implementation, called PAC-2026, checks six separate things before granting that permit: the evidence behind a claim, the runs and artifacts that produced it, how measurements are disclosed, who authorized publication, whether the reader-facing version matches what was verified, and whether the correction history stays intact over time. Fail any one check and the whole permit is denied - there's no partial credit. The team ran the system through more than 110,000 simulated states across ten models, plus historical test cases, and it correctly flagged violations in dozens of deliberately broken configurations.
This is a narrower problem than most AI trust tooling tackles. Provenance and content-authentication standards can show you where a claim came from, but they don't stop someone from publishing a claim whose visible version no longer matches the version that was actually checked. PAC-2026 is aimed squarely at that gap - the moment between verification and publication where an authentic-looking but stale or altered claim slips through.
The authors themselves are careful to say what this isn't: it doesn't establish that any claim is true, hasn't been tested against real-world editorial systems, and has no standards backing. It's a promising piece of plumbing, tested so far only in the lab that built it.