From pacing the frontier to governing the frontier

Over the weekend, Anthropic CEO Dario Amodei published an essay arguing that frontier AI capabilities are advancing so fast that safety and oversight risk falling behind. He proposed a three-step plan: embed third-party evaluators inside labs, which Anthropic will adopt unilaterally; coordinate safety standards among labs in democratic countries; and eventually extend that coordination to include China. He frames this as slowing the rate of capability gain rather than halting development, and ties how much slowdown is achievable to the size of the US lead over China.

Regulate the gap, not the speed

Amodei is right about the underlying problem. AI capability is advancing faster than the institutions and governance mechanisms built to manage it. Even the BRICS summit declaration identified this problem and called for emphasis on safe, secure, reliable, and trustworthy AI. But the answer isn’t a binary choice between acceleration and slowdown. The goal should be to match increases in AI capability with increases in our ability to evaluate, secure, govern, and responsibly deploy it.

That calls for an adaptive assurance framework rather than a universal speed limit, one where obligations rise with demonstrated capability, autonomy, and consequence. Where systems sit within a well-understood risk envelope, innovation and adoption should be enabled. Where new capabilities materially outpace our ability to demonstrate control, security and accountability, development should be gated, temporarily, until assurance catches up. The aim should not be to regulate the speed of innovation, but the gap between capability and control.

What credible governance requires

Independent evaluation sits at the centre of this. Anthropic’s offer of ongoing external access is welcome, but assurance can’t rest on the voluntary decisions of individual labs, and the labs being judged shouldn’t set the terms of judgement. Unfortunately, governments globally have limited capacity and capability to govern AI. While not ideal, governance will therefore require labs to collaborate among themselves and will need support from their respective governments to work with labs across other regimes. Getting such a system right comes down to five conditions:

Evaluators must be independent. Appointment and funding should sit outside the evaluated firm, with a published mandate and no editorial control over findings.
Rules must be symmetric. Thresholds and safety tests should apply to closed and open-weight models alike, scaled proportionately so smaller developers aren’t priced out.
Standards must interoperate. A pacing regime designed by a handful of US labs and exported to everyone else risks fragmenting the market and sidelining serious work already under way at the OECD, the G7 and national AI safety institutes.
Commitments must be verifiable. One that can’t be independently checked isn’t a commitment.
The process must be inclusive. The Global South and mid-tier economies need a seat at the table, since they will live with whatever standards emerge.

What this means for enterprise and government

Governance now has to move beyond the model itself. Risk increasingly depends on what a system is permitted to do: access networks, execute code, control other systems, spend money, replicate, or act autonomously. Both governments and enterprises need governance frameworks that evaluate not simply model capability, but capability combined with agency, access and consequence.

For enterprises, that means moving from generic responsible-AI principles to continuous operational assurance: workload classification, clear human authority, defined agent permissions, auditability and the means to intervene when a system strays outside its intended boundaries.

For governments, it means building permanent technical evaluation capacity, stronger incident-reporting mechanisms, shared safety benchmarks and the institutional agility to respond as capabilities change. At the same time, they should retain control of where and how to deploy AI to manage the downstream risks.

Our recommendations

That translates into a short list of practical steps, for enterprises and for governments.

Enterprises must make AI governance fundamental to implementation, and independent evaluation and transparent risk reporting a procurement criterion. They should build incident-disclosure and evaluation-access clauses into vendor contracts, and avoid concentrating critical operations on a single model provider while the regulatory picture remains unsettled.

Governments must sequence the work. Transparency and third-party audit requirements command the broadest cross-industry support and should move first, with domestic evaluation capacity built in parallel. The upcoming summit between US President Trump and Chinese President Xi Jinping offers an opportunity to engage labs in China to join the call from Amodei. Plurilateral channels, the OECD, the G7 and the national AI safety institutes, are where standards should be shaped, not simply inherited. One practical starting point: a confidence-building process among the major labs themselves, working towards something like a Geneva Convention for AI development and use, whether through an existing vehicle or a new one built for the purpose.

The companies and countries that win this race won’t be the fastest. They will be the ones that can innovate and diffuse capability quickly while proving their systems remain secure, accountable, and under meaningful human control.