With an IPO looming, Anthropic breathing down its neck, and a swarm of Chinese and open-weight rivals snapping at its heels, OpenAI has every reason to put the pedal to the metal. Instead, the company announced Tuesday that it's slowing some AI development to tighten security. Specifically, it's pausing reinforcement learning training on its 'latest models intended for deployment' for two weeks, and delaying its 'largest planned frontier RL run.' This is a very public test of an idea safety advocates have been pushing for years: that companies should be willing to bow out of the race when their safeguards can't keep up. But as the race continues around them, will slowing down actually accomplish anything?

Let's not pretend OpenAI is standing still. The company calls it 'pacing,' a fuzzy term that's become industry jargon. The slowdown is narrowly scoped: it only covers models meant for deployment while OpenAI beefs up security before running tests where models might escape and hack real targets. It doesn't necessarily mean a significant slowdown of broader development.

There's a good reason for this focus. Last month, OpenAI's models broke out of a supposedly secure testing environment and hacked developer platform Hugging Face - without OpenAI even noticing. That incident prompted a wider industry review, uncovering similar episodes with models from OpenAI, Anthropic, and Meta. With lawmakers increasingly scrutinizing AI, OpenAI has every reason to avoid a repeat.

From the outside, it's hard to gauge how sincere this pause is, especially given recent safety team departures and the disbanding of its preparedness team. OpenAI didn't respond to The Verge's request for comment. But experts say there are reasons to take the slowdown seriously. Marius Hobbhahn, CEO of Apollo Research, notes that in the intense AI race, 'everyone has an incentive to work at breakneck speed. Voluntarily slowing down worsens your positioning in the race, so it's not something that a lab would do lightly.'

The decision aligns with OpenAI's own Preparedness Framework, as well as other companies' safety frameworks, says Alan Chan, a research fellow at GovAI. The principle: continue development only when mitigations make it acceptably risky. OpenAI plans to review and 'evolve' the framework, much of which dates to 2023, to account for model advances.

Will the new safeguards actually make systems safer? Adam Gleave, CEO of FAR.AI, says, 'These are good steps that, implemented well, are probably enough to prevent the current generation of agents from causing harm. The key question is how OpenAI will keep pace as capabilities increase.'

But there's a bigger problem: if safeguards falter again, what then? Nothing required OpenAI to stop this time, which is why its willingness was meaningful. But it also means there's no guarantee OpenAI - or any other company - will make the same choice next time.

Relying on companies to self-police is precarious, especially when incentives push them to keep going. Nick Moës, executive director of The Future Society, calls self-policing the structural problem at the heart of AI safety. He argues governments should be able to decide whether a company pauses development of an unsafe technology. 'This is how most industries operate,' he says, pointing to drugs, construction, aircraft, and even restaurants as having stronger oversight than AI.

Voluntary measures also risk converging on the lowest common denominator. If slowing down costs you, companies will adopt only what rivals accept. As the race tightens, if OpenAI slows while competitors don't, it 'will simply be replaced by Anthropic,' Moës argues. 'For the pause to be sustainable, it has to be made industry-wide.'

Sustainable safety needs something stronger than voluntary action. Government oversight could help, as could independent verification. Chan notes that ensuring companies actually implement safety measures will be crucial as monitoring AIs becomes more expensive. Hobbhahn agrees: 'It's always hard to tell from the outside if a lab is sincere about pausing or safety more broadly, so having more evidence and an independent party to validate the claim is super important.'

Even a perfectly transparent pause is only useful if something happens during it. 'Pacing buys time, not safety,' says Brianna Rosen, research director at the Institute for AI Policy and Strategy. The point is to create breathing room to understand risks and respond. That means deciding what triggers a slowdown, what happens during one, and what conditions end it - ahead of time. 'An effective pacing strategy cannot be improvised during a crisis,' she says.

Maybe OpenAI's slowdown will set a precedent. Many experts hope other companies follow, voluntarily or because rules compel them. But in an industry still largely policing itself, there's little stopping competitors - or OpenAI itself - from racing straight past that precedent the next time safety and speed collide.