OpenAI proposes international standards for automated AI research, including work on successor systems. Its proposal of 21 September envisages criteria for human review and incident reporting, supported by public AI safety institutions. This is a corporate proposal to guide standards, rather than a new law or an international agreement already in operation. Source: OpenAI
The announcement is interesting partly because of who writes the questions. A company developing the technology understands difficulties that a government may take longer to notice. It also has preferences about the costs and restrictions it accepts. Hearing it is necessary; allowing its proposal alone to define the problem would transfer a public decision to the party expected to comply.
A common measure does not settle acceptable risk
Standards allow different organisations to discuss comparable results. If two teams use different terms for similar incidents, it becomes difficult to tell whether their records describe real differences or merely different vocabularies. Agreement about what is measured can facilitate scrutiny and reduce duplicated work.
Measurement does not settle every decision, however. Imagine two countries agreeing on how to calculate the probability of harm while disagreeing about which probability is acceptable. A common technical definition clarifies the argument; it does not remove the disagreement. They may also differ over who receives the benefit, who bears the risk and which remedies are available.
NIST describes its AI Risk Management Framework as a voluntary risk-management instrument. That status illustrates a necessary distinction: a technical framework can guide practice without itself creating a legal duty. Its force depends on how it is used and, where appropriate, on additional decisions about binding requirements. Source: NIST
A threshold can change behaviour
Deciding which incidents must be reported involves more than choosing a number. An excessively narrow threshold can exclude early warnings; an excessively broad one can generate reports that consume the attention needed to investigate the most serious problems. Designing it requires considering both the cost of omission and the capacity available to examine notifications.
There is also an incentive problem. When a duty depends on classifying an episode, the party obliged to comply may have reasons to prefer a less demanding category. That does not establish that it will do so, but it justifies making classifications reviewable and preserving records of how decisions were reached. A verifiable rule needs evidence, rather than merely a submitted label.
Automated research introduces a difficulty of its own. If one system proposes changes to another, human review will be meaningful only if the reviewer can understand what needs checking and has time to do it. Requiring a signature without examining those conditions may produce formal compliance alongside empty supervision in practice.
Participation matters as much as vocabulary
Companies and public laboratories possess specialised knowledge, but they do not by themselves represent every affected interest. An organisation using AI, a worker subject to its decisions and a person whose data is involved may require different protections. The process for making a standard should allow those differences to become visible before the rule is fixed.
Participation does not require everyone to design each technical test. It requires decisions about purposes and consequences to be open to discussion, objections to receive answers and responsibility for choosing between alternatives to be identifiable. Specialisation can justify a division of labour; it does not remove the need to explain decisions to those who will live with them.
OpenAI’s proposal deserves attention because it identifies coordination problems no company can solve entirely on its own. Public evaluation should also ask which matters it leaves out and who would have powers to require compliance. Agreeing on a language of safety is a beginning. The political question remains: who can require development to change course when that language describes a risk someone is unwilling to accept?