This is a pivotal moment for AI. Several leading frontier AI labs, including Anthropic, OpenAI, xAI and Google, have backed calls from Dario Amodei, Anthropic's CEO, to slow the development of increasingly capable AI models because of safety risks, including autonomous hacking without human control. A US lawsuit has even argued that coordination among AI companies to slow development could violate competition law. The tension is clear: how to make AI safer without undermining competition?

Competition law generally prohibits agreements between rivals that restrict technical progress, thereby causing consumer harms. AI safety exists to prevent AI technologies from harming consumers. Both can support each other.

Nonetheless, broad safety concerns are not enough to justify agreements between competing firms. In some jurisdictions, including the European Union, cooperation between competitors can be exempted if it improves technical progress, benefits consumers and preserves competition. European Competition Commissioner Teresa Ribera has accordingly called AI companies seeking cooperation to make their requests "precise".

The proposed cooperation appears to distort competition. It would allow AI firms to develop common voluntary safety standards. And it would restrict inputs and activities essential to AI development, including training, computing capacity and the use of AI to improve AI.

The agreement would reduce the competitive pressure driving AI’s rapid progress. Thousands of employees at AI labs calling for a slowdown have acknowledged the problem. They face "intense competitive pressure not to unilaterally slow" development. An agreement among rivals would solve that collective-action problem. But it would do so by removing the competitive pressure that pushes firms to offer better, cheaper, more capable and safer models.

Recent developments show these benefits are taking place. Anthropic and OpenAI have released cheaper models because of intense competition between AI labs. Nvidia, meanwhile, has developed software, the Open Agent Safety Platform, designed to address safety risks.

The agreement would also raise barriers to entry for rivals. Common safety standards and restrictions on essential inputs can increase development costs. Large firms may absorb those costs; smaller firms and new entrants may not. The result could be fewer competitors offering more expensive or less innovative products.

While competition concerns are economic. AI safety concerns can be existential. Recent incidents illustrate the stakes. OpenAI has reported models autonomously hacking websites belonging to international institutions, including Australia's healthcare system. Amodei argues that slowing development would give labs more time to address such risks while technical progress continues. Nvidia takes a different view: Safety risks arose due to insufficient security controls, and the answer is to accelerate safety research.

That disagreement points to solutions that balance AI development and improve AI safety.

An agreement among a handful of leading AI labs to slow development is unlikely to be the best one. It risks reducing competitive pressure between those firms while potentially excluding their rivals. Other mechanisms can improve safety without slowing technological progress.

Voluntary codes of conduct offer one such mechanism. The EU's General-Purpose AI (GPAI) Code of Practice provides rules to help developers of advanced models comply with the Artificial Intelligence Act's safety and security requirements. Independent experts developed the code through a multi-stakeholder process, and any model developer can adhere to it. That makes it less likely to favour a particular group of firms or unduly distort competition.

Such codes could inspire other jurisdictions. Countries without AI-specific regulation do not have to wait years for legislation before acting. Competent regulators, including cybersecurity authorities, could develop voluntary safety codes through multi-stakeholder processes, with competition authorities involved to identify and mitigate competition risks.

Such codes could require developers to report serious safety incidents to the competent regulator and provide permanent forums where companies, researchers, regulators, and other stakeholders can develop safety solutions together.

AI firms and governments can also support inclusive development of safety technologies. Nvidia, for instance, is developing its Open Agent Safety Platform with other stakeholders. In Europe, the Commission's GenAI4EU initiative funds AI applications and encourages stakeholder collaboration.

The choice is thus not between slowing the pace of AI development and improving safety. It is about making safety move faster.