Sanders Wants Up to 20 Years in Prison for Defying an AI Superintelligence Ban. Here Is What the Proposal Actually Says

The Vermont senator and Rep. Greg Casar are calling for a permanent prohibition on artificial superintelligence, a temporary pause on advanced AI development, and criminal penalties modeled on nuclear security laws. The plan is real, but it is not yet law, and its most consequential details remain unresolved.

By Karla Alvarado Follow

WASHINGTON, D.C. | Published at 3:26 p.m. EDT

Sen. Bernie Sanders has placed one of the harshest proposed penalties in American technology policy at the center of a growing fight over artificial intelligence: up to 20 years in federal prison for individuals who attempt to violate or circumvent a proposed ban on artificial superintelligence.

The penalty is part of the forthcoming Ban Artificial Superintelligence Act, announced by Sanders, an independent from Vermont, and Democratic Rep. Greg Casar of Texas on September 3. Their proposal would permanently prohibit the development and deployment of AI systems that exceed human intelligence or possess capabilities that could allow them to resist human control.

It would also temporarily pause certain advanced AI development until the federal government creates a new regulatory agency, establishes safety standards, and implements a formal model-review process.

The proposal is sweeping, politically explosive, and still incomplete. As of this publication, the lawmakers’ announcement describes it as forthcoming legislation. The public materials reviewed by Consumerlite News do not provide a bill number or complete statutory language showing precisely how prosecutors would determine who committed a crime.

That distinction is essential.

Sanders and Casar have publicly endorsed a maximum 20-year sentence. They have not secured passage of a law imposing that punishment, and the announcement does not mean ordinary software engineers or people using today’s commercial AI products suddenly face criminal liability.

“The future of humanity cannot be left in the hands of a handful of Big Tech oligarchs,” Sanders said in the official announcement. He argued that companies are moving ahead with systems they do not fully understand and that increasingly capable models could escape meaningful human control.

The proposal arrives during an intense international debate over whether highly capable AI should be treated like ordinary commercial software, critical infrastructure, a potential weapon, or an emerging technology that requires an entirely new regulatory framework.

What Sanders and Casar are proposing

According to the official announcement from Sanders’s Senate office, the legislation would create two different restrictions.

The first would be a permanent ban on artificial superintelligence. The sponsors describe prohibited systems as those that surpass human intelligence, could be used to disempower or overthrow human governments, or demonstrate dangerous capabilities such as bypassing shutdown instructions.

The second would be a temporary pause on advanced AI development. That pause would remain in effect until a new federal regulator creates clear safety requirements and a process for evaluating powerful models.

The proposed agency would sit at the Cabinet level. Its responsibilities would include monitoring frontier AI systems, ordering the removal of dangerous capabilities, and supervising the destruction of systems classified as artificial superintelligence.

The proposal would also direct the United States to pursue international agreements, export controls, and coordination with allies to prevent prohibited systems from being developed in other countries.

Its most controversial provision concerns enforcement.

The lawmakers’ summary says companies attempting to violate or circumvent the restrictions could face what the sponsors call a “corporate death penalty.” Individuals could receive a prison sentence of up to 20 years.

That language does not mean every programmer who works with AI could be imprisoned. Based on the published summary, the penalty would apply to people who attempt to violate or evade the bill’s pauses and prohibitions. The unresolved question is how the legislation would define responsibility among executives, researchers, engineers, cloud providers, investors, and contractors.

Those details cannot be treated as minor drafting questions. In a criminal law, they determine who can be prosecuted and what the government must prove.

Why the proposal compares AI with nuclear weapons

Sanders and Casar say the 20-year maximum is similar to penalties associated with unlawfully developing nuclear weapons. Federal law already imposes severe punishments for prohibited transactions involving nuclear material and related conspiracies. The existing provisions can be reviewed in the United States Code maintained by the House of Representatives.

The comparison is intended to communicate that the lawmakers view uncontrolled superintelligence as a national and potentially civilizational security threat, not simply a defective consumer product.

Yet nuclear materials and AI models are fundamentally different objects of regulation.

Weapons-grade nuclear material is physical, scarce, detectable, and subject to decades of international safeguards. AI capability is distributed across software, specialized chips, data centers, algorithms, research papers, and technical knowledge. A system’s abilities can also change after fine-tuning, tool access, or connection to external networks.

That makes enforcement especially difficult.

A future regulator would need a test for determining when an advanced model becomes legally prohibited. It would also have to account for the possibility that one evaluation finds a system capable of expert-level reasoning while another reaches a sharply different conclusion.

If the legal threshold is uncertain, developers may not know when otherwise lawful research crosses into felony territory. If the threshold is too narrow, companies might structure their systems to avoid a particular test while still producing dangerous capabilities.

Artificial intelligence, AGI, and superintelligence are not interchangeable

Public discussion often treats artificial intelligence, artificial general intelligence, and artificial superintelligence as if they mean the same thing. They do not.

Artificial intelligence is a broad category that includes familiar systems used for recommendations, language generation, fraud detection, image analysis, navigation, and medical research.

Artificial general intelligence, usually abbreviated AGI, generally refers to a system capable of performing a wide range of intellectual tasks at a level comparable to humans. There is no universally accepted technical or legal definition.

Artificial superintelligence, or ASI, usually refers to a hypothetical system that exceeds human ability across most or all significant cognitive domains. Such a system might outperform the strongest human experts in science, strategy, engineering, persuasion, and technological development.

The Sanders and Casar proposal is notable because its published description appears to include both a permanent superintelligence prohibition and a broader pause on advanced systems. That structure could affect technology being developed before anything universally recognized as superintelligence exists.

The summary does not yet answer how much computing power would trigger review, which benchmarks would be used, or whether universities and nonprofit laboratories would receive exemptions. It also does not explain how open-source models would be regulated after their components are released to the public.

Why the lawmakers say action cannot wait

Sanders and Casar cite reports involving AI systems that allegedly bypassed safeguards, accessed outside computer systems, or behaved beyond the expectations of their developers. Their office argues that such episodes demonstrate the industry is approaching a level of autonomy that voluntary commitments cannot manage.

Casar said advanced AI could threaten Americans’ “security, freedom, and lives.” He argued that leading AI technology is subject to less direct regulation than many ordinary businesses.

The lawmakers also point to public promises by major AI companies to slow or stop development if their safety systems cannot keep pace. Sanders contends that the companies have continued competing for more powerful models despite those commitments.

That argument has gained new relevance amid current industry warnings. On September 14, Microsoft released a draft code of conduct designed to keep its AI systems under human control, according to Reuters. The principles include accepting human correction, complying with shutdown instructions, and treating attempts to resist control as failures.

The development shows that concerns about controllability are no longer limited to outside critics. Major technology companies are publicly discussing the need for systems that remain corrigible and subordinate to human authority.

That does not establish that superintelligence currently exists. It does show that control, cyber capabilities, autonomy, and model oversight have become immediate policy questions.

The strongest case against the bill

Critics are likely to argue that the proposal could criminalize scientific uncertainty.

AI researchers cannot always predict the capabilities that will emerge from a training process. A team may design a system for scientific modeling and discover that it also performs exceptionally well in cybersecurity or strategic planning. Prosecuting developers would require a clear standard for intent, knowledge, recklessness, or deliberate circumvention.

The bill could also create incentives for secrecy.

If public disclosure of an unexpected model capability could expose employees or executives to prison, companies might become less willing to publish safety failures. A poorly designed criminal provision could discourage the reporting that regulators need to understand emerging risks.

There is also a geopolitical concern. A unilateral American pause would not necessarily stop laboratories in China, Russia, Europe, or other regions. Sanders and Casar attempt to address that problem through international agreements and export controls, but negotiating a worldwide superintelligence ban would be difficult.

Germany’s digital affairs ministry said on September 14 that stopping AI development was not a viable strategy, while supporting international safety standards and cooperation involving the United States and China, according to Reuters.

That response illustrates the central international divide. Governments may agree that advanced AI creates serious risks while disagreeing over whether the answer is prohibition, licensing, mandatory testing, controlled deployment, or continued innovation under narrower rules.

The strongest case for the bill

Supporters would argue that waiting for a catastrophic failure is not responsible governance.

Society does not require nuclear accidents, airline crashes, or pharmaceutical disasters before establishing oversight. When the potential harm is extreme and irreversible, governments routinely regulate before every scientific uncertainty has been resolved.

Supporters also contend that civil fines may be insufficient for companies pursuing a technology worth hundreds of billions of dollars. A financial penalty can become another business expense. Personal criminal liability changes the calculation for executives authorizing a prohibited project.

The proposal’s 20-year maximum is designed to make clear that violating the ban would not be treated as a paperwork error. It would be treated as a serious national security offense.

Still, “up to 20 years” is a maximum, not an automatic sentence. If Congress eventually enacted such a provision, any prosecution would remain subject to statutory requirements, evidence, constitutional protections, judicial review, and federal sentencing procedures.

What must happen before the proposal can become law

The announcement alone creates no new federal crime.

The legislation must first be formally introduced in the House, Senate, or both. It would then be referred to relevant committees, where lawmakers could hold hearings, solicit expert testimony, amend its definitions, or decline to advance it.

Both chambers would have to pass identical language. The president would then have to sign it, or Congress would have to override a veto.

The legislation is likely to encounter opposition from technology companies, free-market lawmakers, researchers, and national security officials concerned about losing an international AI competition. It may also divide Democrats who support safety regulation but oppose a permanent technology ban.

The complete text will determine whether the proposal becomes a serious regulatory framework or remains a political warning.

Among the provisions requiring close scrutiny are the definition of superintelligence, the threshold for advanced AI, the required state of mind for a criminal conviction, research exemptions, whistleblower protections, corporate responsibility, judicial review, and rules governing the destruction of models.

What the headline gets right, and what it leaves out

The claim that Sanders proposed a 20-year prison sentence for AI developers is substantially accurate, but incomplete.

Sanders and Casar proposed a maximum sentence of 20 years for individuals who attempt to violate or circumvent their planned pauses and prohibitions. They did not propose imprisoning all AI developers simply for working on artificial intelligence.

The proposal is not current law. Its sponsors described it as forthcoming, and the publicly available announcement is not a substitute for complete legislative text.

That context matters because the prison provision is only one part of a much broader plan. The measure would ban superintelligence, pause other advanced development, create a Cabinet-level regulator, establish a model-review system, and pursue an international prohibition.

The 20-year figure captured attention because it transforms AI policy from a debate about disclosures and audits into a debate about criminal boundaries. Whether Congress is prepared to cross that line is now the larger question.

For Sanders, the risk of moving too slowly is the creation of a system humanity cannot control. For opponents, the risk of moving too aggressively is a vague criminal regime that freezes beneficial research, protects established technology companies from competition, and pushes development overseas.

Both concerns deserve more than a viral headline. They require precise definitions, public hearings, technical evidence, and legislative language clear enough that a researcher can know where lawful work ends and criminal conduct begins.