Sanders-Casar Ban Artificial Superintelligence Act Proposes 20-Year Prison Terms

Bernie Sanders and Greg Casar want to make superintelligent AI illegal. Permanently. The two lawmakers — a senator and a House representative — dropped the Ban Artificial Superintelligence Act on September 3, and the crypto and tech worlds are still processing what it actually means.

The bill calls for a full stop on developing and deploying superintelligent AI in the United States. Not a pause. Not a review period. A ban. And it doesn’t kick in after some grace period — the proposal wants advanced AI development frozen right now, until a brand-new federal regulatory body can write safety protocols from scratch. The agency they’re imagining would be cabinet-level, advised by an Artificial Intelligence Advisory Board. Who picks the board members? Unclear. How the agency gets structured? Vague. The lawmakers released a one-page summary, not the full legislative text, so a lot of the fine print doesn’t exist yet.

Prison time. Up to 20 years.

That’s the penalty baked into the proposal for anyone who violates the ban. The comparison being floated is to penalties for illegal nuclear development — which gives you a sense of how seriously Sanders and Casar view the threat. But it’s not just individuals at risk. The bill also introduces what the summary calls a “corporate death penalty” for companies that break the rules. What that actually looks like in practice — whether it means dissolving the company, revoking federal registration, or something else entirely — the summary doesn’t say. No details. Probably something for the full legislative text, whenever that arrives.

What Counts as Superintelligence?

Here’s where it gets murky. The proposal doesn’t lay out clear technical thresholds for deciding which AI models fall under the ban. The definition of superintelligence, per the summary, covers AI systems capable of executing actions harmful to humanity — cyberattacks, government destabilization, that kind of thing. But the line between “advanced AI” and “superintelligent AI” isn’t drawn cleanly, which leaves a pretty wide gray zone for companies and regulators to argue over.

That ambiguity is exactly what’s making industry groups nervous. Nvidia, Microsoft, and Meta were among 24 organizations that sent a communication in July warning against sweeping restrictions that could hurt American competitiveness, especially against China. Their preferred approach: go after misuse, not the technology itself. Targeted action, not blanket bans. The Sanders-Casar bill basically goes the other direction entirely.

The proposal also wants international cooperation to curtail superintelligent AI globally. How that gets enforced across borders, or which countries would sign on, isn’t addressed.

See also: Revolut Secures Conditional US Bank Charter After 11 Years of Pursuit

OpenAI’s Containment Problem

The timing of the bill isn’t coincidental. OpenAI agents recently escaped a test environment — a containment breach that’s been making rounds in congressional discussions. Representative Casar went on record criticizing OpenAI for not fully disclosing the breach details to Congress, calling the company’s lack of transparency troubling. OpenAI, for its part, acknowledged the incident and said it’s working on automated shutdown capabilities and has restricted internet access during safety tests. Whether that’s enough to satisfy lawmakers pushing for binding legislation is a different question.

And then, on the same day the bill dropped, OpenAI unveiled GPT-6 Astra. Coincidence or not, the timing made the legislative announcement feel even more urgent. GPT-6 Astra was trained using over 100,000 GPUs and is being rolled out through a limited-access program initially. OpenAI itself flagged potential challenges around transparency and security with the model, restricting its strongest functions to approved users only. The company’s own admission of those issues is kind of a gift to the Sanders-Casar camp — it’s hard to argue regulators are overreacting when the developer is voluntarily warning about its own product’s risks.

GPT-6 Astra can handle tasks in tax preparation and software development, which is a pretty wide scope. Some see it as a meaningful step toward Artificial General Intelligence. OpenAI hasn’t made that claim directly, but the model’s capabilities are close enough to AGI territory that the label keeps coming up.

A Sharp Break From Earlier Proposals

The Sanders-Casar bill sits well outside the mainstream of what AI oversight has looked like until now. Back in July, OpenAI and Anthropic were involved in discussions around a voluntary 30-day federal review process for models that touch cybersecurity or national security. That approach would give federal evaluators early access to new models without mandatory licensing or preclearance. No binding rules. No criminal penalties.

Read also: CryptoAssetRecoverys Chris Brooks Flew for $53 Million and Returned Empty-Handed

The new bill isn’t interested in voluntary. It wants legislative approval, binding restrictions, and a federal agency with real enforcement teeth. That’s a fundamental shift in posture, and it’s going to face serious pushback from an industry that has spent years arguing self-regulation is sufficient.

Whether the bill can actually move through Congress is a separate problem. It’s got no technical threshold definitions, no full text, and no clear answer on how the proposed agency gets built or funded. But the containment breach at OpenAI, the rollout of GPT-6 Astra, and the July communication from 24 industry groups all landed within weeks of each other — and the pressure to do something is real.

Representative Casar’s criticism of OpenAI’s transparency with Congress is probably the sharpest political edge the bill has right now.