Should AI be slowed down ?
Dario Amodei's call, and the reasons of those who oppose it

1. What happened

On 12 September 2026, Dario Amodei, chief executive of Anthropic, published a text titled We Must Pace the Frontier. His thesis fits in a sentence : the industry must slow the pace at which it increases the capabilities of AI models.

Coming from the head of one of the laboratories leading that race, the call was picked up immediately. Le Grand Continent published the full text in French the same day.

2. What Amodei proposes

The reasoning runs as follows. If slowing down bought an extra year or two before models reach capability levels judged critical, that time could be spent advancing alignment — the body of methods that make a system do what is expected of it. The worry he puts forward is not the everyday use of models, but their use to build the next generation : the moment progress starts feeding itself and stops being adjustable.

He is not asking for a halt, but for a coordinated pace : a common cadence would let laboratories do that work without any one of them losing commercial advantage, and without the United States losing its lead. That is the condition he sets, and it is also what his critics attack most directly.

A three-step plan

The text describes a three-stage approach. The first is the one Anthropic commits to unilaterally : bringing third-party evaluators inside the company, with a level of access normally reserved for employees. The idea is to make verifiable from outside what until now rested on the laboratory's word.

Unexpected backing

Notable in an industry where executives rarely agree : Sam Altman (OpenAI) and Elon Musk (xAI) both lent their support. Most commentators singled out that convergence as the most surprising element of the whole sequence.

3. The case for slowing down

The arguments put forward by supporters of the call come down to four points.

Safety is advancing more slowly than capability.

The means of understanding and controlling a model are improving, but more slowly than its power. The gap is widening ; slowing the second gives the first time to catch up.

The point of no return is the automation of research.

As long as humans design the next generation, the pace remains a decision. The day models contribute decisively to it, it no longer is.

A delay is not a renunciation.

The argument concerns one or two years, not an indefinite moratorium. The economic value expected from these technologies, he argues in substance, does not vanish if it arrives a little later.

Coordinating avoids punishing the cautious.

A laboratory that slows down alone loses. That is precisely why the text insists on coordination rather than individual virtue.

4. The reasons of those who oppose it

The objections published in the following days fall into four families. They do not all come from the same side : some are put by people who want faster development, others by critics of the industry who fault the text for looking in the wrong place.

a. Whoever writes the rules of their own industry writes rules that suit them

This is the sharpest objection. An incumbent proposing a coordinated slowdown is, in effect, proposing to freeze acquired positions. Several commentators fear that a slowdown among the leaders would harden into a duopoly, and see the plan as a mechanism of power concentration as much as a safety device. New entrants and open-model projects are, mechanically, the ones a common cadence disadvantages most.

b. Evaluators chosen by the evaluated

The flagship measure — embedded third-party evaluators — raises the question of their independence. If the laboratory selects them, nothing guarantees they will do anything other than endorse the regulatory agenda of the company hosting them. The objection is not theoretical : it applies to any audit arrangement whose commissioner is also its subject.

c. A plan with no number and no sanction

The text sets no measurable limit, does not define what « slow enough » would mean, and provides no enforcement beyond the voluntary commitment of a single company. An analysis published by StartupHub.ai sums it up in its title : the plan is judged vague. Nothing at this stage suggests the industry will agree on a common definition.

d. The existential framing distracts from present harms

For another set of critics, concentrating the debate on future risks diverts attention from what the technology already produces : bias in automated decisions, the working conditions of annotators, surveillance uses, effects on employment. Those subjects, they argue, do not require waiting for a future model to be addressed.

The analyst Zvi Mowshowitz published a detailed response, point by point, which remains the most complete reading for anyone wanting to get past the summaries.

5. China, the hard part

Amodei acknowledges it himself : China is the thorniest dilemma in his plan. A slowdown followed on one side only slows nothing — it transfers the lead. His critics make it their central argument : any voluntary cadence adopted by Western laboratories would amount to handing position to those who do not adopt it.

Neither the text nor its critics offer a simple answer to that objection. It is, as of today, the open question in the debate.

6. What it changes for a company here

Let us be direct : in the immediate term, nothing. No commitment made on 12 September changes the tools available to a company in Abidjan, Dakar or Douala tomorrow morning. Three points are nonetheless worth keeping.

• The debate concerns frontier models, not everyday uses. Automating a bank reconciliation, extracting data from an invoice, sorting incoming mail, assisting an accounting team : none of that is affected by a slowdown in cutting-edge research. These uses rest on capabilities that are already available and broadly settled.
• The question of independent evaluation will work its way down to procurement documents. What laboratories are debating today will appear tomorrow in tenders : who evaluated this model, against what criteria, with what level of access. Organisations handling sensitive data have every reason to start asking their suppliers now.
• Dependence on a single supplier is a governance risk, whatever the debate concludes. An organisation that has built a critical process on one model, with no fallback and no way to export, carries a risk that has nothing to do with AI safety — and everything to do with its own business continuity.

That last point is the one that genuinely falls to a director, and the only one that can be acted on this week.

7. Sources

This article reports public positions. Each one is verifiable at the address given.

• Dario Amodei, We Must Pace the Frontier — darioamodei.com (original text, 12 September 2026)
• Le Grand Continent — full text in French
• Zvi Mowshowitz — point-by-point critical analysis
• StartupHub.ai — « the plan is vague »
• Qz France — the slowdown plan and the backing of Altman and Musk
• Mon Carnet — summary in French
• Developpez.com — coverage and commentary

Insight Smart Technologies reports a public debate here and takes no position in it. The opinions reported are those of their authors. Outbound links lead to sites we do not publish.