The leaders of several prominent US artificial-intelligence laboratories are putting a new question before governments: whether companies racing to build more capable systems now accept that the race itself needs limits. Anthropic’s Dario Amodei, OpenAI’s Sam Altman and xAI’s Elon Musk have each backed, in differing terms, a slower and more supervised approach to frontier AI development. Their positions are attracting particular scrutiny as the European Union steps up warnings tied to the AI Act, especially where powerful models are used in increasingly autonomous systems.
The significance of this convergence lies less in shared language than in its commercial and political implications. The executives lead companies competing directly for investment, computing power, customers and influence over the rules that will govern advanced AI. If they are prepared to accept independent assessment and limits on deployment, their stance could strengthen the case for enforceable international safeguards.
If not, the alignment may become a contest over which companies define “safe” AI — and which rivals are excluded by the resulting rules.
A tactical shift in the frontier race
The case for deceleration centres on systems that can operate with less human direction: models deployed as agents, systems able to use tools, and increasingly capable software that can carry out extended tasks across digital environments. Safety advocates are concerned that gains in capability may outpace developers’ ability to understand, evaluate and control such systems. The executives’ interventions put third-party evaluation at the centre of the debate.
That would move beyond company-led testing and voluntary promises towards outside assessment of whether advanced systems should be released or granted new operational powers. Yet the central questions remain unanswered. Who selects the evaluators?
What tests do they apply? What findings trigger a pause? And would companies be legally bound to comply?
Those details matter because voluntary commitments have long been part of the technology sector’s response to regulation. They can create common expectations and encourage disclosure, but they can also leave enforcement to the companies with the greatest incentive to move quickly. For Anthropic, OpenAI and xAI, calls for caution will therefore be measured against product plans, investment needs and their willingness to accept constraints that apply equally to themselves and competitors.
A meaningful slowdown would require more than broad agreement that frontier models pose novel risks. It would require commitments in advance on testing, deployment thresholds and consequences when systems fail.
Europe’s different question: compliance
The EU AI Act changes the terms of the debate. Rather than asking whether leading companies favour restraint, European institutions are moving towards asking whether providers and deployers meet obligations imposed by law. The Act distinguishes between categories of AI risk while also creating duties for providers of general-purpose AI.
Its enforcement will test whether rules written before the latest wave of autonomous-agent products can be applied to systems that operate across services, interact with sensitive data or are used in high-impact settings. Warnings linked to unsupervised agent incidents and cyber-intrusion concerns would sharpen that test. European regulators are likely to focus on whether companies can document risk management, provide sufficient information to downstream users, maintain appropriate safeguards and respond effectively when systems are misused or behave unexpectedly.
That is more demanding than an industry statement of intent. A company may endorse external oversight in principle while disputing the reach of an EU regulator’s authority, the classification of a particular model or the technical feasibility of a required safeguard. Such disputes could become an early fault line between US firms seeking a common global framework and European policymakers determined to retain control over access to the single market.
Safety measure or regulatory moat?
The convergence also invites a familiar criticism: that established labs may use safety language to build a regulatory moat around their own positions. Frontier-model development requires vast computing resources, specialised talent, and access to data and distribution. Rules demanding extensive evaluations, documentation and security controls may be necessary protections, but they may also be easier for the largest companies to absorb.
That does not invalidate the safety case. Independent researchers have warned that more autonomous systems may create risks that are difficult to detect before deployment, particularly when models can plan, write code, seek information or interact with external tools at scale. The policy challenge is to distinguish practical safeguards from barriers designed chiefly to entrench incumbents.
European officials and US lawmakers must also decide how far to rely on company-designed governance. A credible regime would combine independent technical expertise, clear reporting obligations, meaningful audit access and penalties capable of changing corporate behaviour.