Consensus is a floor
A hundred experts, thirty nations, one report. Documents like that are strongest at establishing what is no longer arguable and weakest at saying what to do first — which is exactly the question a regulator brings to them.
What that process produces
Legitimacy, which is not nothing. In a field where most authoritative technical writing is published by the companies being regulated, an independent reference point that no single government can dismiss as another country's anxiety has real value.
And a floor. What survives review by a hundred experts across thirty nations is what is no longer seriously contested. That makes disagreeing with it expensive and agreeing with it cheap — a useful asymmetry, and a limited one.
What it necessarily loses
Ranking. Consensus documents describe landscapes; they rarely say which risk to fund first, because that is where the thirty nations diverge. The policymaker's actual question — what do I do on Monday — is the one the format is worst suited to answer.
And currency. A document reflecting a hundred contributors describes knowledge as it stood some months before publication, in a field where eleven models shipped in twenty days.
Where the live argument is
In the conference programmes. Thirty-five of 223 orals at ICLR 2026 were safety papers — roughly one in six, allocated by reviewers with no stake in the field’s public image, on work submitted months earlier. That is a cleaner measurement of research attention than any statement of priorities.
It also shows how heterogeneous "safety" has become. Watermarking is signal processing. Bias is statistics. Interpretability is closer to neuroscience. Filing them under one heading is administratively convenient and analytically misleading — and it is why progress in one is routinely reported as progress in all.
How to use both
Take the consensus report as the floor of what can be argued and the conference track as the frontier of what is being argued. Anyone citing the first as though it were the second is quoting a summary of last year to settle this year's question.
AI Security Institute — Research & Publications → · Future of Life Institute — AI Safety Index — Summer 2026 → · ICLR 2026 review — 35 AI-safety oral papers at ICLR 2026 → · Oxford AIGI — Legal Alignment for Safe and Ethical AI →