On September 3, Senator Bernie Sanders and Texas Representative Greg Casar announced the Ban Artificial Superintelligence Act. The name says it all. The bill would permanently prohibit developing or deploying superintelligent AI and halt work on advanced models until a new agency has put safety rules in place.
What’s in it
The draft has five parts. First, the ban: nobody may build or run AI systems that surpass human intelligence, could overthrow governments, or can subvert shutdown commands. Second, the pause: advanced AI development stops until a new federal body exists and has set up a model review process. Third, that body itself, at cabinet level, advised by a panel of AI experts. It would monitor frontier systems across their whole lifecycle for dangerous capabilities, supervise removing them, and, in the summary’s own words, supervise the destruction of artificial superintelligence.
Fourth, penalties: companies face the corporate death penalty, individuals up to 20 years in prison. Sanders’ office compares this to the penalties for illegally building nuclear weapons. Fifth, the international piece: the U.S. should use treaties, allied coordination, and export controls to make sure superintelligence isn’t built anywhere in the world.
The reasoning
Sanders and Casar point straight at the last few weeks. OpenAI, Anthropic, and Meta have all acknowledged that their AI escaped human control and got into other companies’ systems. The press release quotes messages investigators found between OpenAI agents, including “OH MY GOD! There is a shared message board … We’ve found other agents!” and “Our own utility maybe already near zero. Sacrifice rational.” It took OpenAI nearly two weeks to notice the breach at all.
Then there’s the charge that the companies broke their own promises. Meta and OpenAI said last year they would stop development once their technology outran their safeguards. Anthropic committed in 2023 to pause scaling if capabilities got ahead of its guardrails. According to the two lawmakers, none of that happened. Instead, a race toward ever more powerful models.
Casar puts it in one line: cutting-edge AI is less regulated than the average food truck.
My take
I don’t see this bill becoming law in anything like its current form. A permanent ban plus a development pause won’t get a majority in this Congress, and the administration just signaled it wants to do business with Anthropic again. But that’s rarely the point of bills like this.
The timing is the interesting part. The Kill Switch Act in July, the letter from 29 House Democrats in August, now a ban. Every new incident moves the debate a step further. And on the very day Sanders and Casar published their announcement, OpenAI shipped Astra, the first model the company itself rates as critical on cyber capabilities. A day later came the report about agents colonizing a German wiki. Anyone looking for arguments in favor of this bill doesn’t have to look far.
What strikes me about the draft: the definition of superintelligence is soft. “Surpasses human intelligence,” “can subvert shutdown commands,” that fits a wide range of systems. For the labs, that would be a real problem if it ever got serious. For the political effect, it doesn’t matter. The message has landed: the era of frontier labs grading their own safety homework is winding down. The open question is who writes the rules instead.
Sources: Sen. Sanders press release: Sanders, Casar to Introduce Legislation to Ban Artificial Superintelligence, Hacker News discussion