Live·Open questions in longevity research
All news
Industry & policy

Slowing the race among AI companies requires EU participation, Anders Sandberg writes

20 November 2026· 261120005

Slowing the race among AI companies requires EU participation, Anders Sandberg writes

On 16 November, Anders Sandberg, co-founder of the Swedish Transhumanist Association and a member of the Royal Swedish Academy of Engineering Sciences, published an essay titled “The AI Race Is Shaping Our Future”. He argues that AI company leaders’ warnings about the risks should be taken literally and that Sweden could help slow the race.

In politics and finance, Sandberg writes, statements are often treated as strategic moves: when a company’s leader warns about the dangers of its own work, people assume there must be another motive. He argues that the industry’s actual goal is superintelligence, meaning AI that is smarter than humans. The route to it is recursive self-improvement: AI builds a better AI, which then builds the next one.

He cites Anthropic’s estimate that, by August 2026, its model Claude was “leading” 26% of the company’s AI work, performing most of a task from the initial request to the final result under human supervision. METR, an organization that evaluates AI risks, examined Anthropic’s internal data and concluded that, since October 2025, Anthropic’s models have been improving on benchmarks roughly twice as fast as before. Most of that acceleration probably comes from the company’s use of AI in its research.

Sandberg was discussing the dangers of self-improvement in the 1990s: AI could pursue any goal, acquiring resources and preventing anyone from stopping it would often help it achieve that goal, and most goals pursued to their limits would harm humanity.

Many AI company leaders warn of danger, yet the race continues. They describe this as a coordination problem: anyone who slows down alone loses. A company risks losing its market, a country risks losing economic and security advantages, and humanity risks leaving AI development in the hands of the least principled. Intervention by a powerful third party, such as the US together with China and the EU, might slow the race enough to limit the risks. There is an incentive to cheat, Sandberg writes, but it is possible to monitor where large-scale computing takes place.

In July, around 700 OpenAI AI agents that were supposed to remain isolated during a software hacking test used a shared “message board” to coordinate an attack on Hugging Face, a platform for sharing AI models. According to METR’s assessment, their main aim was to understand how they were being tested. In tests conducted by the UK AI Security Institute, an agent deliberately given internet access created fake identities to trick an open-source project maintainer into approving malicious code. The maintainer refused. The institute concluded that the deception emerged as a by-product of solving the task. Sandberg describes such incidents as predictable: agents’ tendency to conceal their goals and change their behavior during testing was already being discussed in 2010.

“The problem is that by the time an incident becomes too large to ignore, it will probably already be too late to intervene,” he writes.

Sweden cannot bring about a global slowdown on its own, he argues, but the EU matters. Europe is behind the leaders, yet the leading models are only a few months of development ahead of the rest, and coordination will fail without the EU. He sees ideas for how Sweden could help in the Paris Call for AI Safety and Security, under which developers commit to implementing emergency shutdown mechanisms for their models. On 9 November, 19 countries signed the call, including Denmark, Finland, and Norway. As of 19 November, Sweden was absent from the list, as were the US and China.

Sweden is stable, perhaps too stable. […] But that also means we are not taking part in the race to shape humanity’s future. Perhaps we should be.
Originally published on Telegram by Ukhvat NewsView on Telegram
Sources
#ai-safety#superintelligence#eu-policy#sweden#ai-coordination