Sanders, Casar Seek Superintelligent AI Ban, Frontier Pause

Most of the debate about AI regulation in the United States has focused on disclosure rules, liability questions and voluntary commitments from developers. A new bill from Bernie Sanders and Greg Casar goes considerably further. It would stop the most compute-intensive AI development for a period and permanently outlaw one category of technology.

Sanders and Casar, a Democratic member of the House, presented the proposal on 23 September under the name Ban Artificial Superintelligence Act. It rests on three measures: a permanent prohibition on artificial superintelligence, a pause on the largest AI models, and a new federal department dedicated to AI.

A compute threshold decides who has to stop

The pause would take effect as soon as the law is enacted. It would not cover AI in general, only models whose training requires at least 10^25 operations. The bill measures the size of a model by the raw compute used to train it. The new department could revise that threshold every year to keep pace with technical progress.

For developers above the line, the consequences are direct. They could not continue training or modifying such models. They also could not release any model in this class that has not already been published.

The pause has no fixed end date. It would last until the new department is fully staffed and has published rules for the safe development and release of these systems. After that, developers would face a licensing process: they would have to report their plans, submit their models for review and obtain approval before any release.

How the bill defines superintelligence

A permanent ban needs a definition, and the draft offers two. An AI counts as superintelligent if it outperforms humans at most cognitive tasks, or if it could destroy humanity or strip it of power.

The ban does not stop there. It also covers models that show certain "precursor capabilities", meaning abilities the authors see as early warning signs. The bill lists three:

• improving themselves without human direction
• gaining unauthorised access to secured computer systems
• helping to develop biological, chemical or nuclear weapons

Under this approach, a developer would not need to build a system that surpasses humans to break the law. A model that could, for example, assist with weapons development would already fall under the prohibition.

Testing during training, not only before release

Much of the enforcement burden would fall on the new department. It would be required to examine advanced models for prohibited capabilities during training and while they are in operation, not only at the point of release.

If the department identifies a banned capability, it would have to cut the model off from the internet and halt its development. A system found to be superintelligent would have to be taken out of service immediately.

That is a notable shift in timing. Oversight would follow a model through its whole lifecycle, rather than acting as a single gate at launch.

Prison terms and loss of assets

The penalties are among the most severe parts of the proposal. Responsible decision-makers and independent actors who break the rules could face up to 20 years in prison. Companies could lose their development licence as well as their assets.

Sanders and Casar also recognise that a national ban has limits. The bill calls for international agreements aimed at preventing superintelligence from being developed outside the United States as well.

What the proposal would change

For companies training frontier models, the bill would bring three changes. Development above the compute threshold would pause for an open-ended period. Afterwards, releases would depend on government approval. And certain capabilities would be off limits no matter who builds them or how.

Much would depend on how the new department reads its mandate: where it sets the compute threshold each year, how it tests for precursor capabilities, and how quickly it can be staffed and issue rules. As long as that work is unfinished, the pause stays in place.

The proposal shifts the question lawmakers are asking. The focus is less on how AI systems should be used once they exist and more on whether some of them should be built in the first place.