Harmonizing AI Safety Thresholds: Frontier Labs' Published Capability Limits Are Not Comparable

arXiv·high signal

Wilber Sean Anterola, Matthew Ball and Luis F. Lafuerza (arXiv 2607.16112, cs.AI) document that the capability thresholds frontier AI companies publish in their safety frameworks differ substantially in definition and measurement, making third-party verification of whether any lab has crossed a line effectively impossible. The paper proposes a harmonization scheme to make the thresholds mutually legible. This matters as a governance-infrastructure story: policy that references 'the lab's own stated threshold' is currently referencing incomparable numbers.