The AI slowdown debate is also a fight over control

Dario Amodei, Sam Altman and Elon Musk have put the pace of frontier AI development into public dispute. They are among the people with the power to change that pace. Their statements have also raised a question that will survive this week's exchange: how much control would their proposed safeguards give anyone outside the companies building these systems?

Amodei's proposal calls for slowing capability advances while safety work catches up, alongside outside scrutiny and wider coordination. He says Anthropic will embed independent evaluators, with substantial access and rights to publish findings subject to narrow redactions. Altman's Monday update says OpenAI requires safety cases before frontier reinforcement-learning runs. Musk has proposed peer review between competitors. Those statements differ in what they promise and how far they have progressed. They do not amount to a common agreement to stop building AI.

I have been reading the responses alongside research we conducted earlier this year. One condition respondents rated was the ability to ask for a human review of important workplace or school decisions. Respondents who already used AI as part of their regular work gave that condition particularly strong support. Regular use could coexist with wanting a route to human review.

The people using AI still want a way to challenge it

In our May–June survey of 2,584 respondents, we asked how important several conditions would be for trusting AI in a workplace or school. One was that workers or students could ask for a human review of important decisions. Across the sample, 79.5% rated that very or extremely important. Among the 604 respondents who said they used AI in a regular workflow involving files, data, code, images or other work, the figure was 93.2%.

The corresponding figure among people who had tried AI once or twice was 67.4%. That is a substantial descriptive difference, though it does not tell us that using AI caused someone to want review. The complete breakdown is below, including groups that do not fit a simple progression from less use to more support for safeguards. These are unweighted respondent results, collected May 24 through June 12, with the question wording and bases available here.

Our June research work had already identified acceptance of AI with conditions attached. The current results show why treating AI use and support for human review as opposing positions would misread these respondents. Someone can put AI to work regularly and still rate access to human review as important. That is a useful starting point for examining which safeguards a proposal provides.

The survey asked about decisions in workplaces and schools. It did not ask respondents to endorse a frontier slowdown, choose an evaluator or settle competition with China. Reviewing a model before further development and reviewing a decision made about a person also happen at different points. A person denied an opportunity needs a route to challenge that decision, even if an expert examined the underlying model months earlier.

What the objections reveal

Criticism from David Sacks and Aaron Levie concerns who would acquire power under the proposed arrangements. Sacks supports voluntary caution while objecting to rules that could entrench the leading laboratories. Levie questions delay and concentrated control. Those are objections a serious safeguard has to answer. A requirement can be presented as public protection while being easier for a large incumbent to meet than for a smaller competitor.

The practical difficulty is that scrutiny costs money and requires access to expertise. An evaluator needs enough knowledge and access to discover a problem, along with the freedom to report something the sponsor would prefer to keep private. Peer review between competitors could expose weaknesses that an internal team misses. It also leaves questions about how a smaller developer enters the arrangement and who represents people affected by the technology. A promise of review needs those details before we can judge its independence.

The China objection deserves its strongest version. If the United States restricts useful AI capability while competitors continue advancing, it could surrender economic advantage and military capability, as well as influence over the standards other countries adopt. A lasting, unilateral cap would require a convincing account of why its safety benefits exceed that strategic cost. Policy has to account for competitors continuing to advance.

Amodei acknowledges the constraint of maintaining an American lead. China's September 14 foreign-ministry response calls for inclusive international governance; it offers no reciprocal, verifiable limit on frontier development. We therefore have an unresolved coordination problem. Agreement that safety is desirable does not establish agreement about which capabilities to delay, who can verify the delay, or what happens when a participant breaches it.

That objection is strongest against an indefinite American capability cap. Its force against a particular security requirement or outside evaluation depends on what that measure costs and what it prevents. A requirement that protects models from theft, for example, could preserve an advantage. One that consumes scarce research time without detecting relevant failures could erode it. The useful argument has to get down to the actual measure. Saying that China will continue developing AI cannot, by itself, decide between those two cases.

The next thing to inspect is the agreement

I would like to see the proposed evaluator arrangements in enough detail to understand what happens after a bad finding. Amodei has described access and publication rights. The next evidence is whether those terms survive in the implemented arrangement, how evaluators are appointed and paid, and whether the company must do anything in response. A report can be public and still leave the decision entirely with the firm being examined.

Anthropic has separately reported an agreement with METR to investigate its cybersecurity incidents, initially for eight weeks with an option to extend. That gives us a concrete arrangement to examine alongside the broader proposal for ongoing embedded evaluation.

For the person using AI, or living with a decision made through it, there is a further question about remedy. Can they get a qualified person to reconsider the decision? Can that person change the outcome? The human-review condition in our research is concrete because it gives someone a way to act when the system has affected them. The frontier proposals should be examined with the same attention to what their safeguards allow another person to do.