Arthur Mensch’s AI Safety Challenge: Real Risks or Market Power?

In a September 24 interview with Le Monde, Mistral AI CEO Arthur Mensch argued that American AI giants are exploiting fears about the technology and insisted that “AI is software. It can be controlled.” His challenge is consequential, but it is not a verdict on AI safety: a warning can describe a genuine risk and still advance the interests of the company making it.29

What Mensch is challenging

Mensch’s argument concerns who gets to define a responsible pace for AI development. If the largest developers shape safety rules around requirements they are best equipped to meet, those rules could disadvantage smaller competitors. That is a competition concern, not proof that the risks are invented.29

Mistral’s position is not that deployment is risk-free. The company has warned that models become powerful actors when given autonomy and access to tools, and that deploying them safely requires care. Mensch’s claim that AI can be controlled therefore needs to be read as an argument for managing risk, not as evidence that every future system will be controllable.72

Why the warnings cannot simply be dismissed

Former Anthropic researcher Jacob Coxon, who had previously worked at OpenAI, warned in early September that AI “could kill us all by the end of the decade.” Anthropic CEO Dario Amodei subsequently called for a voluntary slowdown in the development of the most advanced models; Le Monde reported that leading American competitors supported his September 12 appeal.18

Those are serious warnings and proposals, but they are not established forecasts. Coxon’s statement describes a possibility, not a demonstrated deadline. Likewise, support for slowing development does not, by itself, show that a particular slowdown would reduce risk more effectively than targeted safeguards.18

The disagreement also extends beyond a Europe-versus-US divide. Geoffrey Hinton has urged faster action on safeguards, while Andrew Ng has challenged the evidence behind doom-oriented publicity and argued against blanket calls to pause progress. Their opposing assessments make the practical question more precise: which risks have been demonstrated, and which measures would address them?

Where the antitrust allegation fits

A lawsuit brought by AI-service subscribers alleges that Anthropic, OpenAI, SpaceXAI and Google agreed to restrain competing development efforts following the September 12 slowdown appeal. It is an allegation, not a finding that the companies reached an unlawful agreement. The distinction matters: executives expressing similar safety concerns is not, on its own, proof of coordination.

Mensch’s criticism and the lawsuit raise related but separate questions. His argument asks whether safety rhetoric can favor incumbents. The legal allegation asks whether competitors actually agreed to slow development. Neither question can be answered by treating every warning as a marketing tactic—or every call for caution as proof of an anticompetitive pact.

The most defensible approach is to assess proposed safety measures against independently examinable evidence, then ask whether their costs fall disproportionately on potential competitors. That scrutiny should apply to Mistral’s case for continued development as well as to incumbent-backed calls to slow down. A company’s commercial interest is a reason to test its argument, not a reason to reject it automatically.789