Gates demands an AI law, Sanders wants 20 years in prison for superintelligence

The American debate over AI regulation is no longer about pace. It is about instruments. On 25 September Bill Gates told NBC News' "Meet the Press" that Congress should pass a law regulating the development of artificial intelligence. Asked whether legislation was needed, he said yes. The requirements, he added, had to be mandatory.

Gates left no doubt about how he sees industry self-regulation. "Nobody thinks self-regulation alone is enough," he said. Law enforcement and politicians, he argued, had to be part of any discussion about safety standards and who oversees them. New requirements would add to companies' workload, he said, but would not slow their work in any significant way. He also warned that bad intentions combined with the latest AI tools create a threat on a mass scale, and that no weapon had ever worked so powerfully before.

"Ban Artificial Superintelligence Act 2026"

The bill known as the "Ban Artificial Superintelligence Act 2026" goes much further. It defines superintelligence as a system that surpasses humans in most tasks, or one that can easily be modified to reach that state. The definition also covers a system capable of planning and carrying out the destruction of humanity or the overthrow of a government. Such systems could not be created, deployed, acquired, possessed, financed or imported.

The bill would set up a Department of Artificial Intelligence as a supervisory body. If the department identifies superintelligence, the system is to be stopped immediately and destroyed within 30 days. Anyone who detects such a system must report it within 24 hours. A development pause would also apply to "advanced AI systems" trained with at least 10²⁵ operations, until the department sets clear rules. The threshold would be adjusted every year.

The bill's realism is limited. The 10²⁵ FLOP threshold covers a large share of today's models, so the effects would be broad rather than targeted. Still, the two statements are worth reading together. Gates represents the minimal demand: mandatory, enforceable standards. Sanders represents the maximal one. In the American system it is not a lack of arguments but a lack of a majority that keeps AI regulation a draft.

Sanders' bill is harsher than its name suggests. It bans not only the creation of superintelligence but also its use, acquisition, possession, financing and import. It establishes a new Department of Artificial Intelligence, and an identified system of that class is to be stopped immediately and destroyed within 30 days. Anyone who comes across such a system must report it within 24 hours. The development of "advanced AI systems" trained with at least 10^25 operations would be halted until the department is ready, and the threshold itself would be adjusted annually. Breaking the ban carries up to 20 years in prison.

Gates and Sanders represent the two ends of the same dispute. One wants mandatory requirements that will not stop work on models. The other believes that at this level of capability the pace is a problem in itself. The common ground is political: both accept that voluntary commitments by companies are no longer enough.