If each country writes its own AI safety rules, can labs' safety evidence still be compared side by side?

This explores whether different national safety rules would make it harder to compare one lab's safety evidence with another's. The corpus never tests this directly. It does show that safety evidence is fragile even under a single standard, which suggests fragmentation would make an already hard problem worse. The clearest policy statement comes from OpenAI, which argues that international safety standards are 'as important to pacing the frontier as alignment research itself.' It frames fragmentation as a collective action failure, a situation where each actor follows its own rules and nobody can tell whether the whole field is moving safely Can global standards pace frontier AI as much as alignment research?. The Future of Life Institute reaches a similar conclusion from another direction. It argues that companies cannot police themselves and that binding oversight needs shared verification, down to the hardware level Can companies alone manage the risks of AI systems?.

The corpus also shows what a shared standard makes possible. The Frontier AI Risk Management Framework scored many recent models against the same seven capability areas and the same green/yellow/red thresholds. That common scale exposed a pattern you might not expect: most models crossed the warning line for persuasion and manipulation while staying in the green for cyber offense and self-replication Where do frontier AI models actually pose the greatest risk today?. With each lab reporting against its own national categories and thresholds, you couldn't see that pattern, because there would be no common scale to place the results on.

The less obvious point is that the method used to measure can change the result, even when the system being tested stays the same. In one study, the verdicts of a language model acting as a judge shifted 31% of the time on complex tasks. An agent that gathered evidence before judging cut that to 0.27% Can agents evaluate AI outputs more reliably than language models?. If one country required one evaluation method and another country required a different one, two labs could report different safety results for reasons that have nothing to do with their models. Work on whether errors stay visible and recoverable finds the same problem within a single field: partial measures exist for each piece, but none of them connects to the others How can we measure whether AI errors stay visible and recoverable?.

There is a warning in the other direction too, against assuming a harmonized checklist would settle the question. Systems can pass every local check and still fail as a whole, because the local checks test different properties than the ones that decide whether the full system behaves safely Can individual components pass safety checks if the system still fails?. Safety failures also tend to hide in evaluation habits. They look plausible, they spread across a system, and they come to seem normal inside everyday workflows Why do safety failures remain invisible to our evaluation methods?. Applied to regulation, each national regime could be internally consistent and still miss the same failures. Common standards make lab results comparable, but they don't guarantee that what's being compared is the evidence that matters.

The gap in the corpus: none of these notes compares actual national regimes, such as the EU, US, UK, or China, or measures how much divergent rules distort cross-lab comparison. The argument above is assembled from neighboring research on measurement, not from direct evidence.