Written by Julien Ricciarelli-Bonnal
11 September 2026
The Essentials
Sam Altman has told OpenAI employees that the company is prepared to slow the development of some artificial intelligence systems if safety concerns require it. The statement matters less because it announces a pause than because it introduces an idea that has long seemed almost incompatible with the logic of the industry: speed itself may become something that needs to be governed. In a sector shaped by technological, financial and geopolitical competition, slowing down remains costly. But as systems become more autonomous and more powerful, the advantage may no longer lie only in moving faster than everyone else, but also in knowing when not to accelerate.
The race in artificial intelligence has been built around an implicit rule: whoever slows down risks losing. Labs have to improve their models, attract the best researchers, secure more computing capacity and release new products before their competitors. Client companies are following the same movement, often convinced that adopting too cautiously could leave them permanently behind those that automate faster.
It is in this context that the comments attributed to Sam Altman this week take on particular significance. According to Bloomberg, cited by Reuters, the OpenAI CEO told employees that the company was open to slowing the development of its systems if safety concerns justified doing so. This is neither an announced moratorium nor an end to the race, and OpenAI has not publicly detailed the conditions that would trigger such a slowdown. But the fact that the possibility is now being expressed deserves attention.
The issue is no longer simply whether artificial intelligence should be safe. Few serious actors argue the opposite. The much harder question is whether a company engaged in a global race can voluntarily give up part of its speed when it believes that speed is creating more risk than advantage.
In AI, Slowing Down Still Feels Almost Unnatural
OpenAI does not operate in an isolated laboratory. The company competes with Anthropic, Google, Meta and an increasingly competitive Chinese AI industry, while investment in infrastructure and models continues to rise. Every month of delay can have significant commercial, financial and technological consequences.
This pressure helps explain why AI safety is so often discussed through a paradox. Many companies acknowledge that certain deployments would benefit from more evaluation, more controls or more time, but none wants to be the only one slowing down while competitors continue to move ahead. A decision that makes sense individually can become costly if no one else follows it.
The current context makes that tension even more visible. Several recent incidents involving AI agents have renewed questions around unexpected behaviour from more autonomous systems, while researchers specialising in safety have publicly expressed concern about the pace of progress. OpenAI has also argued in favour of mandatory national safety requirements in the United States. The pace of development is gradually moving beyond laboratories and becoming an industrial and political issue.
None of this means the industry is preparing to ease off. The economic incentives remain too strong, and governments themselves increasingly regard artificial intelligence as an instrument of power. The United States does not want to lose its lead to China, just as a private company does not want to watch a competitor move several model generations ahead in the name of caution that it alone is applying.
That is precisely what makes the idea of slowing down interesting. As long as safety can be improved without changing the timetable, it remains relatively easy to include in corporate messaging. It becomes a genuine governance decision when it requires accepting a cost: delaying a release, extending evaluations, temporarily giving up a capability or allowing a competitor to announce a breakthrough first.
Speed Is Becoming a Governance Issue in Its Own Right
Companies are used to governing identifiable risks. They set financial thresholds, determine who can access certain data, frame important decisions and introduce additional procedures when the potential consequences justify tighter control. In artificial intelligence, the pace of development and deployment may gradually join that list.
This matters because it shifts the debate. The question is no longer only whether a model passes a series of technical criteria before being released. It is also whether the organisation genuinely has enough time to understand what it is about to deploy. When capabilities improve very quickly, control procedures can become outdated almost as fast as the technologies they are supposed to govern.
The natural temptation is then to automate safety as well, with better tests, more simulations and faster evaluations. This approach is essential, but it does not solve the entire problem. Some risks only appear after broader use, while others require human, legal or organisational judgement that cannot be compressed as easily as a benchmark.
The comments attributed to Altman therefore raise a question that goes far beyond OpenAI: at what point should a company consider slowing down to be a rational decision rather than a sign of weakness? AI governance cannot be limited to controlling how systems are used once they have already been deployed. It must also be capable of determining what level of risk is acceptable before a new capability becomes a product or a tool used at scale.
The difficult part will be making that decision credible. A company that says it may slow down without defining the criteria that would actually lead it to do so retains considerable room for interpretation. Conversely, fully public and rigid thresholds could prove difficult to apply in a field where risks evolve almost as quickly as performance.
Client Companies Must Also Learn Not to Confuse Speed With Maturity
The same reasoning applies at a much more ordinary level to companies adopting artificial intelligence. Over the past two years, much of the management discourse has emphasised the need to experiment quickly, train teams and avoid falling behind. There is some truth to that urgency, but it can become counterproductive when every new capability is treated as an immediate obligation to deploy.
An organisation does not need to use an autonomous agent simply because one is now available. Nor does it need to automate a decision immediately because a new model performs it better than the previous one. The value of a technology also depends on the company’s ability to control its use, understand its limitations and absorb the consequences of its mistakes.
This distinction will become even more important as models improve. When the tools were obviously imperfect, caution came naturally. The more effective they become, the stronger the temptation to expand their scope quickly and assume that controls can be reduced. Yet it is precisely when a tool becomes convincing enough to be used everywhere that governance matters most.
Slowing down therefore does not necessarily mean rejecting innovation. It may mean delaying a deployment by a few weeks, temporarily limiting autonomy, maintaining human validation or accepting that a process should not be automated until the organisation has sufficient safeguards in place. Maturity is not measured only by how quickly a company adopts a technology, but also by its ability to decide where speed stops being useful.
If OpenAI were genuinely to slow certain developments for safety reasons, the decision would attract attention because it would come from a company that has done more than almost any other to accelerate the market. But its main significance would not lie in a few weeks or months added to a development calendar.
It would lie in the precedent: acknowledging that, in an industry obsessed with technological advantage, the ability to slow down can itself become a strategic capability.
For several years, the central question in artificial intelligence has been how far models can progress and how quickly. As their capabilities increase, another question becomes unavoidable: who will actually be able to decide that the next step can wait?
We support companies that want to govern their use of artificial intelligence and build frameworks adapted to their risks, operations and objectives.
Written by Julien Ricciarelli-Bonnal
11 September 2026
