The Ban Artificial Superintelligence Act pauses training above 10^25 FLOP and leaves home inference alone

The Ban Artificial Superintelligence Act pauses training above 10^25 FLOP and leaves home inference alone

Senator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act on September 23, 2026. The bill would create a federal Department of Artificial Intelligence, pause all training of AI systems above a compute threshold until that department writes safety rules, permanently ban systems rated as artificial superintelligence, and punish violations with up to 20 years in prison. Emma Roth at The Verge and the AP both covered the introduction today. Sanders’ own office published a section by section summary of the bill text.

The compute line sits at 10^25 operations, and that number decides whether a training run falls under the mandatory pause.

The pause line starts above Llama 3.1 405B, so home training stays outside it

Section 3 defines an “advanced artificial intelligence system” as any system trained on 10^25 operations or more. The Secretary of Artificial Intelligence can adjust that number after enactment.

Llama 3.1 405B, Meta’s open flagship from the summer of 2024, trained on roughly 3.8 times 10^25 operations. That training run is already above the line. Every model at or beyond Llama 405B class training cost would be caught by the mandatory pause in Section 8, which blocks training, modifying, or fine-tuning until the new department is fully staffed and has issued rules. Deployment of unreleased systems pauses too.

The models people actually run at home sit orders of magnitude below the line. A 27B model trained from scratch is on the order of 10^23 to 10^24 operations. The gap between a home rig and the pause line is a factor of a thousand to a million, not a factor of two.

The bill as written never sees your hardware because nothing in it touches inference. Section 8 pauses training, not running models. A fine-tune on a workstation, including an M4 Max, is around 10^20 to 10^21 operations, far below the Section 8 pause.

The bill’s behavior definitions leave room for smaller systems

Two definitions in Section 3 are qualitative, and they are where enforcement could creep toward smaller systems. The bill bans anything with “superintelligence precursor characteristics”, and defines those to include the capacity to automate AI research and development, the capacity to ensure its own continued operation against shutdown attempts, and the capacity to independently modify its own functions.

Those definitions describe behaviors, not scale. An agent that improves its own prompts edges toward the first. A service that restarts itself when killed sits near the second. The bill’s answer is that only systems caught by the compute threshold or formally identified as superintelligence get sequestered under Section 10, and the Secretary would need to make a finding. Those definitions remain available for a future Secretary to use.

As introduced, this bill is aimed at frontier labs and the trillion-parameter training runs behind them. For a home user, the practical effect is narrow: Section 8 pauses training above 10^25 operations, and your fine-tune is far below. The phrase “capacity to automate AI research” is broad enough that a future enforcement action could argue it applies to a model that costs nothing to download.