Europe heard slow down and saw an incumbent moat

01

WHAT ACTUALLY CHANGED

Reuters reported on September 18 that several European companies, officials, and researchers opposed a laboratory-led slowdown. French developer Mistral said some incumbents were using the safety moment to consolidate their market position and push rules that favor them. That is Mistral's commercial and policy position, not proof that every safety proposal is a disguise for market control.

Raphael Auphan, chief operating officer of Swiss privacy-software company Proton, called the moves self-serving and said they could preserve dependence on United States services. Ben Brooks, public-policy head at German startup Black Forest Labs, warned that arbitrary thresholds could chill open innovation near the frontier. Neither statement establishes that all capability thresholds are arbitrary or that acceleration is safe.

Clement Delangue, CEO of French-American open-model platform Hugging Face, said it was time to accelerate rather than slow down. He also endorsed Amodei's proposal for independent evaluators embedded inside AI companies. That combination is important: European opposition is not a single demand for speed with no inspection.

Kristian Kersting, an AI professor at the Technical University of Darmstadt, described a system in which large companies define danger, choose inspectors, and seek an antitrust exemption as regulatory capture. French Finance Minister Roland Lescure pointed to the self-interest in asking companies behind the leaders to slow down. Germany's digital ministry said stopping development was not viable given Europe's digital-sovereignty goals.

Amodei's own proposal is more specific than a general pause. His three steps are embedded evaluators with employee-like access, coordination among frontier companies in democratic countries, and global coordination. He says pacing should be triggered by capabilities and safety evidence, could include checkpoints, and should not mean halting model training or technical progress.

The essay says Anthropic intends to invite an external review team with office access, company devices, relevant workspaces, and publication rights, subject to narrow redactions for security, privilege, commercial sensitivity, and third-party confidentiality. No named evaluator, executed contract, public access record, first report, or tested dispute process was included in the essay.

Europe is not waiting for a voluntary compact. The European Commission says the AI Act became generally applicable on August 2, 2026, while obligations for general-purpose AI models have applied since August 2025. The AI Office can request technical documentation, evaluate models, require corrective measures, and issue fines. This legal system is separate from Amodei's proposed company coordination and does not automatically resolve the competitive concerns raised by European firms.

02

WHY THIS MATTERS

A safety rule can be necessary and still shape a market. If compliance requires enormous private evaluations, access to rare compute, secret incident data, and a relationship with one approved auditor, the largest laboratories may absorb the cost while newer teams disappear. The answer is not to delete the rule. It is to measure the burden, publish the route through it, and make equivalent evidence available to challengers.

The party choosing the threshold can choose the competition. A capability checkpoint tied to a demonstrated ability, such as defeating a defined sandbox under controlled testing, can focus oversight on a real hazard. A threshold tied only to training spend or a fixed compute number can be easier to administer and easier to game, while catching efficient small models and missing risky downstream combinations.

Independent does not mean outside the payroll on paper. An evaluator needs secure access, technical competence, stable funding, protection from retaliation, a right to publish material findings, a route to report denied access, and limits on revolving-door conflicts. If the laboratory selects, pays, scopes, and can quietly dismiss the inspector, employee-like access becomes company-hosted theater.

Antitrust waivers deserve narrow drafting. Competitors may need to share cyber indicators, evaluation methods, and evidence about catastrophic risks. They should not gain cover to coordinate product roadmaps, prices, customers, hiring, model access, or a shared definition that conveniently excludes their own systems. Safety cooperation needs an agenda, public minutes where possible, an independent chair, and a regulator watching the watchers.

Europe's dependence creates its own bias. Mistral, Black Forest Labs, Proton, Hugging Face, politicians, and industry groups benefit from more European capacity and market share. Their warnings about incumbency can be correct without making acceleration harmless. The relevant evidence is not which side sounds purer. It is whether a proposed control reduces a specified risk without unnecessary exclusion.

The EU has a chance to separate public oversight from company permission. The AI Office can build shared evaluation infrastructure, recognize several qualified assessors, publish methods and aggregate findings, support smaller providers, and use the same capability evidence for European, United States, and Chinese models placed on its market. That would turn sovereignty from a slogan into institutional capacity.

Users and workers should not be supporting actors in a laboratory truce. Model releases affect security teams, creators, teachers, public services, energy systems, disabled people, and communities hosting infrastructure. Their experience belongs in the threshold, incident, remedy, and appeal process. A race negotiated only by the racers is not public governance.

03

WHERE IT COULD HELP

• Define every capability threshold with a testable behavior, operating conditions, model and tool access, success rule, uncertainty range, retest schedule, and public explanation of why crossing it changes the required safeguards
• Accredit several independent evaluators through a public process, disclose funding and conflicts, protect publication rights, log denied access, rotate teams, and let a regulator investigate disagreements between the evaluator and the laboratory
• Create shared European evaluation facilities and subsidized testing for startups, universities, open-source projects, and public-interest researchers so safety evidence does not require buying the same private gate twice
• Limit safety coordination among competitors to a written scope covering hazards, indicators, tests, incidents, and remedies, with antitrust supervision, participant records, sunset dates, and prohibitions on sharing commercial strategy
• Pair each release gate with a correction and appeal path: name the failed evidence, permit remediation, preserve model and test versions, publish the final decision, and prevent an incumbent from vetoing a competitor through an opaque benchmark
• Report distributional effects alongside safety results, including compliance cost, time to approval, small-provider participation, open-model access, evaluator concentration, model withdrawals, incidents prevented, false alarms, and whether affected groups received a remedy