Live·Open questions in longevity research
All news
Industry & policyScientific ComputingAI in medicine

Historian Rutger Bregman calls for a global halt to training more powerful AI while keeping current models in use in science and medicine

14 November 2026· 261114003

Historian Rutger Bregman calls for a global halt to training more powerful AI while keeping current models in use in science and medicine

In an essay published on 11 November 2026, he warns that as AI becomes better at helping develop new models, those models may improve ever faster. Bregman fears that people will lose control of this process and calls for independent inspections and criminal liability for AI company executives who violate the bans.

As recently as 7 June, the author of Utopia for Realists argued against halting AI development: if data centers were shut down in California, they would be built in Texas or Abu Dhabi. Even then, he warned that people could lose control, but proposed government oversight and international agreements. He urged Europe to invest in its own computing capacity: “The free world must build.”

In his November essay, he explains why he stopped trusting skeptics who believed AI capabilities had reached their limit: models kept learning to perform new tasks. He contrasts those forecasts with a 2021 scenario that predicted widespread conversations with chatbots and assistants that carry out tasks online independently. Bregman now argues that we should listen to those who correctly predicted AI's development and warn of the risk of losing control.

According to an assessment published by the research group METR on 5 November, Anthropic's models were improving at tasks with verifiable results roughly twice as fast in October 2026 as they had from March 2024 to October 2025. Based on tests of the company's unreleased models, the group considers a sustained increase in the rate of improvement more likely than a one-time jump followed by a return to the previous rate. It attributes most of the acceleration to the use of AI itself in research and development.

Bregman acknowledges that AI adoption across conventional professions has been slower than its proponents expected: hospitals and businesses have to reorganize how they work. In AI laboratories, AI helps researchers write code and run computational experiments to develop the next model. The results can be checked and used to improve the model. That improved model can then help develop its successor. He fears that development will accelerate with each cycle until people lose control of the process.

Let us start by banning AI from creating AI.

To address the possibility of development moving to another country, he now proposes an agreement with China to inventory and track chips. His plan calls for halting the production of new chips for training AI and gradually replacing existing training chips with chips specialized for running trained models. He intends this replacement to allow existing AI to remain in use during the training pause. Bregman believes such controls are feasible because the most powerful systems run in enormous data centers. Their size and electricity consumption make them difficult to conceal. Independent inspectors would verify compliance with the bans.

Bregman considers the cost of such a pause acceptable: he says we have barely begun to use the capabilities of existing models. Even if they remained as they are today, they “could already transform science and medicine for decades.”

Originally published on Telegram by Ukhvat NewsView on Telegram
Sources
#ai-safety#training-pause#chip-controls#ai-governance#recursive-improvement