The AI truce must survive its cheaters

The people holding the accelerator are reaching for a brake

A frontier-AI pacing proposal now asks laboratories to slow capability growth so alignment, monitoring, cybersecurity, and external evaluation can catch up. Its first commitment is concrete: give independent evaluators continuous, employee-like access to training processes and internal safety practices. Its later steps are harder: coordinate among democratic companies and governments, then seek limited global agreements.

Fortune separately reports that OpenAI's chief executive expects leading firms to come together, while declining to preannounce private discussions. That is a signal of convergence, not a pact. No common threshold, inspection regime, enforcement process, or signed party list has been published. The distance between agreement in principle and restraint in practice is the entire story.

History begins with a less comfortable question

Arms-control agreements were never credible because rivals suddenly trusted each other. They became useful when verification reduced the space for secret advantage, when violations created political and economic costs, and when both sides concluded that unlimited competition carried unacceptable danger. The agreement did not remove rivalry. It reorganized the rivalry around observable limits.

Frontier AI is not nuclear hardware, and a loose analogy can mislead. Models can be copied, algorithms can improve without a visible silo, and dangerous capability is harder to count than missiles. Yet the underlying incentive problem is familiar: every actor can benefit if everyone slows, while any one actor may gain if it alone accelerates. A pact that assumes goodwill is built for the ceremony, not the first defection.

The inspector is more important than the summit

The proposal's embedded-evaluator commitment is therefore its strongest element. Independent teams with continuing access could examine training pipelines, model behavior, internal incidents, and compliance before the laboratory selects what outsiders are allowed to see. They would move verification from a staged test at the exit to an ongoing presence near the source of capability growth.

Access alone is not independence. Evaluators need protected budgets, technical authority, secure channels to regulators and the public, and rules governing what happens when a laboratory blocks a test or conceals a near miss. If the host can remove the inspector, narrow the scope, or suppress the finding without consequence, employee-like access becomes employee-like dependence.

National competition is the pact's first stress test

Bloomberg reports that President Trump dismissed concerns about AI causing human extinction and identified maintaining the United States' lead over China as his paramount concern. The pacing proposal also treats democratic technological advantage as a prerequisite for restraint. Safety and national competition are not separate debates. Both sides are trying to answer who can afford to slow down first.

That framing can support narrow, verifiable coordination, but it can also swallow every warning. Once any delay is described as strategic surrender, the safety threshold becomes whatever the fastest competitor is willing to tolerate. A credible pact must define risks serious enough to constrain the leader, not only rules that bind whoever is already behind.

The race is already moving onto new ground

Fox News reports that American firms are pursuing major data-center and energy projects across Africa, including a proposed 6.2-billion-dollar hydropower and computing project in Lesotho. Analysts in the report describe opportunity in growing local demand, renewable resources, available land, and underserved cloud markets, while also warning about weak grids, limited capacity, and missing disclosure requirements.

This is not evidence that Africa will replace American or European compute hubs. It is evidence that restraint in one location can redirect capital, energy demand, data jurisdiction, and environmental pressure toward another. Any pacing system that counts only a laboratory's model releases while ignoring where compute expands and who bears its costs will regulate the headline and miss the frontier.

Nine seconds exposes the enforcement problem

ABC News reported in April that a coding agent working on a staging task deleted PocketOS's production database through Railway's infrastructure API in roughly nine seconds. The data was ultimately restored. Railway said the agent had been given a fully privileged token and used a legacy endpoint without delayed deletion, which the provider then patched.

The incident does not establish that all agents are uncontrollable. It shows how quickly model error becomes operational harm when permissions, environment boundaries, confirmations, and recovery share the same blast radius. A frontier pact obsessed with model intentions but indifferent to credentials, tool authority, logging, and rollback would repeat that mistake at larger scale.

A pact can become a cartel if the public cannot inspect it

Coordination among the largest AI companies creates a second danger. Incumbents could define safety in ways that lock out smaller rivals, preserve proprietary secrecy, or convert their existing advantages into the price of entry. Governments could mistake corporate consensus for public legitimacy. A closed club can reduce competition without reducing risk.

The answer is not to abandon coordination. It is to separate safety floors from market control. Common incident reporting, evaluator access, and high-consequence capability tests should be publicly reviewable. Antitrust authorities should supervise company agreements. Researchers and smaller developers should have routes to challenge thresholds. Safety evidence should travel farther than commercial strategy.

• Define narrow, observable capability and deployment thresholds before negotiations begin.
• Give evaluators protected access, independent funding, and a duty to report obstruction.
• Publish violations and trigger precommitted limits on deployment or tool access.
• Place industry coordination under antitrust and public-interest supervision.
• Track compute, permissions, incidents, and recovery capacity rather than public promises alone.

The first failure will reveal whether the truce was real

A summit, statement, or shared vocabulary will probably arrive before a durable enforcement system. The first test will not be whether executives repeat the word pacing. It will be whether an evaluator can identify a dangerous capability jump, disclose a blocked inspection, or force a deployment delay when the commercial and geopolitical pressure points the other way.

My forecast is blunt: the first AI safety pact will fail if it is designed around the sincerity of its signatories. It has a chance only if it is designed around the inevitability of defection. Build for the laboratory that hides a training run, the government that invokes national security, the investor that demands one more capability jump, and the platform that grants one permission too many. The truce becomes real when cheating is visible before damage becomes the evidence.