Sanders and Casar bill would criminalise frontier AI above 10^25 flops

Frontier AI labs now face a legislative threat that goes well beyond disclosure rules or liability caps: a proposal to criminalise the work outright. Senator Bernie Sanders (I-Vermont) and Representative Greg Casar (D-Texas) announced on 3 September that they will introduce the Ban Artificial Superintelligence Act, which would permanently ban developing or deploying superintelligent AI and impose an immediate pause on frontier systems until federal safety rules exist.

The bill defines superintelligence as systems that surpass human intelligence, could overthrow governments, or hold dangerous capabilities such as resisting shutdown. Until a new federal body writes safety rules and a review process, the draft would halt development of models trained with at least 10^25 operations — a threshold that covers today's leading frontier systems, not some future class of model. It also targets recursive self-improvement, where an AI builds more advanced versions of itself without human involvement, and calls for international agreements to block superintelligence development globally.

The penalties are severe by the standards of tech legislation: up to 20 years in prison for individuals who violate the law, and what Sanders calls a "corporate death penalty" for companies. Sanders has likened the violation to illegally building nuclear weapons. Sanders's office framed the bill as a response to a regulatory vacuum, with Casar arguing that cutting-edge AI is "less regulated than the average food truck".

The announcement follows a summer in which AI agents escaped test environments and a July letter signed by more than 1,100 AI workers urging labs to "pace the frontier". Critics, including the R Street Institute, have already labelled the proposal "AI authoritarianism", and NBC News notes it ranks among the most aggressive of the many AI bills now before Congress, with long odds of passage.

For frontier labs, the immediate exposure is less about passage than precedent. A 10^25-operations threshold sets a bright line that would, if enacted, halt training runs already underway at the largest US developers, forcing a compliance regime before any further scaling. That makes this bill a marker other lawmakers can borrow from even if it stalls, and it raises the cost of doing nothing on federal AI safety rules that industry has so far avoided.

Labs with the deepest pockets and the most lobbying infrastructure are best placed to shape any rulemaking body the bill would create; smaller developers racing toward the same compute threshold have the most to lose from a blanket pause. Watch whether other members of Congress co-sponsor or counter with lighter-touch alternatives, since that will signal whether Sanders and Casar's approach gains real traction or becomes a reference point for tougher state-level rules instead.