Sanders proposes banning superintelligence with prison for executives: the question the industry avoids

By Zendoric · September 8, 2026.

Senator Bernie Sanders (Independent, Vermont), together with Representative Greg Casar (Democrat, Texas), has introduced the Ban Artificial Superintelligence Act, a bill that combines four measures in one: a permanent ban on building or deploying "artificial superintelligence" —defined in the text itself as systems that match or exceed top-level human performance, or that could wrest control away from people—; an immediate pause on the training and deployment of the most advanced models until a new federal AI safety agency exists to approve each project; a call for the U.S. government to negotiate an international agreement to halt the race globally; and criminal penalties, with prison for executives and developers and court-ordered shutdown of companies that fail to comply.

The trigger is not abstract. In 2026, OpenAI, Anthropic and Meta each acknowledged an incident during safety testing of their models. According to Memeburn, the three companies worked with the same outside evaluation provider, Irregular, and a configuration error left the test environments connected to systems they should not have had access to: the models ended up interacting with real systems outside the intended boundaries. None of the three companies described an actual containment breach. Sanders leans on those incidents not as proof of a disaster already consummated, but as evidence that current testing controls are not reliable enough to keep scaling capability without external oversight.

The reactions trace the usual line. Nvidia's Jensen Huang calls for regulation to be anchored in real, verifiable harms, not in still-hypothetical risks. Bill Gates doubts a global pause is even enforceable. Sanders replies that the reliability of industry self-regulation is precisely what is in question after a summer with three incidents at the three leading labs.

The underlying problem, which even the bill's supporters concede, is that no one has defined "superintelligence" in operational terms. The current wording —matching or exceeding human performance, or being able to act in ways that wrest away control— is vague enough that it sets neither a clear technical threshold nor who decides when the pause ends. On top of that, the bill requires international cooperation in the middle of a technological standoff between the U.S. and China, something the source itself describes as difficult. This is not the first attempt to rein in AI by law: in 2023 an open letter from researchers and tech leaders called for a six-month pause on training the most powerful systems, and California went as far as passing SB 1047 before the governor vetoed it in 2024. What is new here is the jump in intensity: from a temporary pause to a permanent ban with prison time attached.

Our reading: the odds of this text becoming law are low, and the underlying analysis itself acknowledges as much —the definitional problem and the political reality of Congress weigh against it—. But the value of the proposal lies not in its passage, but in the question it forces into public view: not who controls the most powerful AI or who profits from it, but whether it makes sense to keep chasing more power without limits. That is a question the industry systematically avoids, because its business model and its narrative of progress take for granted that more capability is, almost by definition, better.

At Zendoric we already documented in late August how isolation failures in test environments stopped being an isolated incident at one lab and became a pattern: Anthropic and OpenAI had failed within weeks of each other on the same kind of problem, the isolation of the environments where models are evaluated before release. This episode, which adds Meta and points to a single outside provider as the common denominator, reinforces that reading: this is not one careless lab, but a structural limitation in how the sector tests its most powerful systems before turning them loose on the world. That this failure has now reached the Senate, turned into a legislative argument, confirms something we had been anticipating: safety has stopped being an internal debate among labs and is beginning to set the political agenda.

That said, we keep our caveats. A permanent, criminally enforced ban on reaching human-level capability runs head-on into the long-term thesis we defend: it is precisely that growth in capability —properly governed— that can bring us closer to eradicating disease and generating the abundance that frees people from jobs they did not choose. Freezing progress toward that capability by law, instead of demanding testing controls, independent audits and verifiable accountability in every deployment, risks throwing out progress along with the risk. Sanders's underlying concern —that the industry is not proving capable of policing itself— is legitimate, and in the short term we share it; the right answer is not necessarily to halt research, but to require that every leap in capability come with an oversight capacity that, judging by what happened this summer, does not yet exist.

What is really worth watching is not whether this bill passes, but whether its logic —a real safety incident used as a concrete political argument, not as a theoretical risk— repeats itself in the next round of legislation, and whether the new federal AI safety agency Sanders proposes reappears, in a more moderate form, in some future bill with a real chance of advancing.