The Case for Slower AI, From People Who Build It Fast
Several of the loudest advocates for caution are also competing hardest for capability. That is not necessarily hypocrisy.
There is a familiar criticism of the artificial intelligence industry: the people warning most loudly about risk are the same people racing hardest to build the systems that pose it. The criticism has force, and it misses something important.
If it is coming anyway, the question is who builds it
If you believe the technology is coming regardless, then the useful question is not how to stop it but who builds it and under what constraints. From that premise, it is rational to want to be at the frontier and to argue for rules that bind everyone, including yourself.
The premise may be wrong. But a person who holds it is not being incoherent, and dismissing them as insincere forecloses the more interesting argument about whether the premise is right.
- A jurisdictional race makes unilateral restraint costly and possibly counterproductive.
- Published safety commitments can be cited by regulators drafting rules.
- Internal capability is what makes an external evaluation possible at all.
The commitments have no price attached
The problem is not the position but the absence of costs attached to it. A commitment that is never tested, a threshold that is never approached, and an evaluation that is never published all allow the rhetoric of caution while the behaviour remains unchanged.
The test of sincerity is not what a company says about risk. It is what it does when the safe option is slower and more expensive. That test arrives rarely, and the industry’s record on the occasions it has arrived is mixed.
The decision that would settle it
The situation in which a commitment becomes costly is specific: a competitor ships a capability that changes the market, and closing the gap requires proceeding without a safeguard that the published policy requires.
That decision has likely been faced internally by every lab with a public safety framework. Whether it was made in favour of the commitment is mostly unobservable from outside, which is why external verification matters more than the wording of the policy.
Useful discipline, and not a substitute for law
Neither extreme survives scrutiny. It is not credible that the people building these systems are indifferent to their consequences, and it is not credible that good intentions substitute for external constraint.
The productive position is to treat public commitments as useful engineering discipline and simultaneously to demand the regulation that would give them force. Those two positions are compatible, and the people who hold only one of them tend to be arguing about something else.
Image credit and licence details for every photograph on this site are listed on the credits page. This article is editorial content; it carries no sponsored material.