Anthropic proposes a "brake pedal” for artificial intelligence

The American company Anthropic, one of the main developers of generative artificial intelligence, has proposed the creation of an international mechanism that would allow major players in the industry to slow down or even temporarily suspend the development of the most advanced AI systems, to give society the time it needs to adapt to technological changes, AFP reports. The initiative comes at a time when competition between companies such as OpenAI, Google, Anthropic and the Chinese group DeepSeek is fueled by investments in the tens of billions of dollars and growing concerns about the safety and control of these systems.

• Risk of "loss of control”

In a paper published by the Anthropic Institute, the company's think tank, the Claude model developer argues that it would be beneficial for the industry to have mechanisms capable of slowing down technological progress when research on safety and society's ability to adapt no longer keeps pace with the pace of innovation. The company also invokes the risk of "loss of control” over artificial intelligence systems, warning that, without global coordination, companies and governments could be forced to make sensitive decisions under the pressure of economic competition and geopolitical rivalries.

• Model inspired by arms control treaties

According to Anthropic, a useful precedent is the Intermediate-Range Nuclear Forces (INF) Treaty, which established mechanisms to verify and limit nuclear weapons. But the company warns that artificial intelligence is much more difficult to monitor than classic military infrastructure, because software models can be developed and hidden much more easily than missile silos.

• "The AI sector has an accelerator, but no brake”

In an interview with the BBC, Anthropic co-founder Jack Clark compared the current stage of artificial intelligence development to the oil boom of the early 20th century, a period in which the emergence of regulations helped to strengthen public confidence. "At the moment, it is as if the AI sector has an accelerator pedal, but no brake pedal,” he warned.

• Difficult step to accept in the context of rivalry with China

According to AFP, the idea of a voluntary slowdown in the development of artificial intelligence could meet strong opposition in the United States, where many industry leaders believe that such a measure could give China a strategic advantage.

In this context, President Donald Trump recently said that he had discussed with Chinese authorities the possibility of cooperation in the field of artificial intelligence security. At the same time, the leader in the White House signed an executive order that allows the monitoring of the most advanced AI models in the name of cybersecurity, but the participation of companies in this mechanism remains voluntary.

• Accusations of "fear marketing”

Founded by former OpenAI researchers, Anthropic has built its image around the concepts of safety and ethics in the field of artificial intelligence. However, the company has been criticized by part of the industry and by American officials, who accuse it of exaggerating the dangers associated with AI and of practicing a real "fear marketing”. In recent months, the company has postponed the launch of its most advanced model, Mythos, to implement security updates, announcing that systems with similar performance will be launched in the coming period.

Anthropic argues its call for caution with internal data that suggests the possibility of the emergence, in a still uncertain future, of the phenomenon of "recursive self-improvement” - the situation in which artificial intelligence systems would become capable of designing and training the next generations of models, with minimal human intervention. The company emphasizes that there is no evidence that such a threshold is imminent, but it believes that, if current trends continue, such a scenario becomes plausible.

According to Anthropic, such an evolution could generate major advances in medicine, technology and economics, but could also bring the risk of a loss of control over artificial intelligence systems.