# Ban Artificial Superintelligence Act: 20 Years Jail [2026]

> Sanders and Casar's Ban Artificial Superintelligence Act would jail AI executives for up to 20 years and pause frontier development. Here's what it actually says.

- Canonical: https://www.kunalganglani.com/blog/ban-artificial-superintelligence-act
- Author: Kunal Ganglani
- Published: 2026-10-13 · Updated: 2026-10-13
- Category: Technology · Tags: ai-safety, ai-regulation, ai-policy, artificial-superintelligence, ai-security

## TL;DR

Two U.S. lawmakers, Bernie Sanders and Greg Casar, introduced a bill that would make building "artificial superintelligence" a crime punishable by up to 20 years in prison, and would freeze advanced AI development until a new federal watchdog exists. No AI today actually meets the bill's definition, so it bans a destination rather than a product. It almost certainly won't pass as written. But it sets the extreme end of the debate, which is how milder rules eventually get framed as the reasonable compromise. If you build apps on top of AI, you're not the target. The labs training frontier models are.

The Ban Artificial Superintelligence Act is a U.S. federal bill that would make developing artificial superintelligence a crime punishable by up to 20 years in prison, and would force a pause on advanced AI development until a brand-new federal agency exists to police it. Most coverage frames it as a prison bill. That misses the point. The prison clause is the headline, but the definition of what you'd actually go to prison for is where the whole thing lives or dies.

The timeline is tight. The Ban Artificial Superintelligence Act was announced on September 3, 2026, and formally introduced in the U.S. Senate on September 23, 2026. Its lead sponsors are [Senator Bernie Sanders](https://www.sanders.senate.gov/), the Independent from Vermont, and [Representative Greg Casar](https://casar.house.gov/), the Texas Democrat who chairs the Congressional Progressive Caucus. This is not a fringe filing from a backbencher. It is a coordinated bicameral move from two of the loudest voices on the American left.

## The 20-year prison clause everyone fixated on

The penalties are genuinely unusual for tech legislation, and they are what made the bill go viral.

![black and gray code padlock anchored on chain-link fence selective focus photo](https://cdn.sanity.io/images/vzekdneq/production/09a54c9b20d53039be322e5722fa578fe862bc73-1200x675.webp)

For companies, the maximum sanction is judicial dissolution. A court can order the business shut down. Not a fine that gets expensed as the cost of doing business. Dissolution. For individuals deemed responsible, meaning executives and in principle the researchers who build the systems, the bill authorizes up to 20 years in prison.

Put that next to how the U.S. normally handles corporate wrongdoing and the contrast is stark. The EU AI Act, which I've written about in the context of shipping EU AI Act compliance for [AI agents](/blog/eu-ai-act-compliance-ai-agents), caps penalties at 35 million euros or 7% of global turnover. That is a cost. Twenty years in prison is not a cost. It is a deterrent aimed at a specific person's life.

> A law that jails researchers for building something nobody can yet define is a statement of fear, not a framework for safety.

That is the tell. Fines treat a harm as something you pay down. Criminal liability for individuals treats the activity itself as inherently dangerous, closer to how we handle weapons development than how we handle product defects. The sponsors chose that framing deliberately. Whether you think that is proportionate depends entirely on the next question.

## How the bill defines artificial superintelligence — and why that's the whole fight

Artificial superintelligence, under the bill, is AI that either exceeds human ability across most cognitive tasks, or is capable of causing human extinction or catastrophic destruction. Read that twice, because the word "or" is doing an enormous amount of load-bearing work.

![A padlock and dollar bills rest on a computer keyboard](https://cdn.sanity.io/images/vzekdneq/production/44d0a59a4f18944180be89990a07f6c266764ef7-1200x675.webp)

The first branch, exceeds humans on most tasks, is a capability bar. The second branch, capable of causing catastrophe, is a risk bar. A system could trip the second without touching the first. A narrow model that is not generally smart but can, say, design a pathogen would arguably qualify. That is a feature, not a bug, if you are a safety hawk. It is a nightmare if you are a general counsel trying to tell your engineers what is legal to train.

Here is the thing nobody building the compliance process is saying out loud: nothing on the market today meets this definition, and that makes enforcement a forecasting problem. Based on the benchmark data I maintain at [kunalganglani.com/llm-benchmarks](/llm-benchmarks), no current model, not even open-weight frontier systems like the one I covered in the [Kolibri 1M-context benchmark](/blog/kolibri-open-weight-benchmark), shows anything close to general superhuman performance, let alone extinction-level capability. The best [large language model](/glossary/llm) today still fails tasks a competent intern handles. So the law is not banning a thing that exists. It is banning a destination, and then criminalizing the road toward it.

That is why the definition is the fight. Draw the line too wide and you outlaw ordinary frontier research. Draw it too narrow and the law catches nothing until it is too late.

## Precursor systems and the recursive self-improvement trap

The bill does not stop at superintelligence itself. It also regulates "precursor" systems, models with characteristics treated as stepping stones to ASI. Explicitly named: recursive self-improvement, the capability for a system to improve its own design without human intervention.

![An open padlock surrounded by scattered black computer keyboard keys under red and green light](https://cdn.sanity.io/images/vzekdneq/production/36a9b7592300095aa1822626d260acbf3d7bfbc7-1200x675.webp)

This is the clause that should worry working engineers more than the prison sentence, because "precursor" is a far squishier category than "superintelligence." Recursive self-improvement is not a binary switch. Automated hyperparameter search is a weak form of it. A coding agent that writes and runs its own evaluation harness, then rewrites its own prompts based on the results, is arguably a weaker form still. I run a multi-agent pipeline that does exactly this kind of self-correction on this blog. Agents that critique and revise each other's output against a deterministic quality gate. Nobody would call that a precursor to extinction. But the statutory language does not obviously exclude it.

That ambiguity is the real regulatory surface area. If you build [agentic AI](/pillars/ai-agents) that plans, executes, and refines its own behavior in a loop, you are operating in territory the bill gestures at without precisely bounding. The safe interpretation is that regulators would target training runs of frontier-scale foundation models, not application-layer orchestration. But "the safe interpretation" is not a defense you want to rely on when the downside is criminal.

## A pause until a new federal agency exists

The mechanism is a moratorium. Development of advanced AI models pauses until Congress stands up a dedicated federal agency to monitor AI companies and enforce safety compliance. No agency, no development. The agency is the condition precedent for the entire industry to resume frontier work.

This is the part that makes the bill a long shot, and it is worth being honest about why. The U.S. has not created a major new regulatory agency from scratch in a generation. The Consumer Financial Protection Bureau in 2011 was the last one of real consequence, and it took years to staff and survive its first legal challenges. Standing one up with the technical staff to audit training runs, people who can actually read a model card and a compute budget, is a multi-year hiring and funding problem. A pause "until" that exists is, functionally, an indefinite pause. Supporters would say that is the point. Critics, including most of the labs, will say it hands the frontier to jurisdictions that did not pass the law.

There is good reason to be skeptical of "pause until the perfect regulator exists" as an operating model. Running this blog's own agent pipeline taught me that [deterministic gates applied before the expensive step](/pillars/ai-engineering-production) catch more than doubling the size of the model you run afterward. Safety that ships beats safety that waits for an institution that may never be funded. A narrow, enforceable disclosure rule today will prevent more harm than a comprehensive regime that arrives in 2031. The institution-first approach gets the sequencing backwards. It withholds every rule until the perfect enforcer exists, when the enforceable rules are exactly the thing you can build now.

## Where this fits: SB 1047, SB 53, and the 2024-2026 safety wave

The Ban Artificial Superintelligence Act did not appear in a vacuum. It is the loudest entry in a legislative wave that has been building since 2024.

California's SB 1047, the Safe and Secure Innovation for Frontier AI Models Act, would have imposed safety-testing and shutdown-capability requirements on the largest models. [Governor Gavin Newsom vetoed it](https://www.gov.ca.gov/) in September 2024, arguing it regulated by model size rather than actual risk. A year later a narrower successor, SB 53, the Transparency in Frontier Artificial Intelligence Act, was signed into law in 2025, focusing on transparency and incident reporting rather than hard liability.

That trajectory matters. The state-level trend has been toward disclosure and away from bans. SB 53 passed precisely because it dropped the parts of SB 1047 that scared off moderates. The federal Ban Artificial Superintelligence Act runs the opposite direction: it is maximalist where the surviving laws went minimal. On the risk spectrum, SB 53 says "tell us what you're doing." This bill says "stop doing it."

| Law | Year | Approach | Max penalty |
| --- | --- | --- | --- |
| California SB 1047 | 2024 (vetoed) | Safety testing + kill switch for frontier models | Civil, size-based |
| California SB 53 | 2025 (signed) | Transparency + incident reporting | Civil fines |
| EU AI Act | 2024-2026 rollout | Risk tiers, conformity assessment | 35M euro / 7% turnover |
| Ban Artificial Superintelligence Act | 2026 (introduced) | Moratorium + criminal liability | 20 years prison / dissolution |

Seen in that table, the bill is an outlier. It is the only one of the four reaching for criminal penalties against individuals, and the only one proposing a flat pause.

## The enforcement problem nobody has solved

Set aside whether the bill should pass and ask a harder question: how would anyone actually enforce it? You cannot run a conformity assessment against a system that does not exist yet. There is no test that returns a clean "this crosses the superintelligence line" verdict, and the bill's two-pronged definition, exceeds humans or is capable of catastrophe, makes the judgment call murkier, not sharper.

So enforcement collapses onto the one proxy everyone in frontier-AI policy eventually reaches for: compute. The 2023 Biden executive order drew its reporting line at training runs above 10^26 floating-point operations. The EU AI Act flags general-purpose models trained above 10^25 FLOPs as carrying systemic risk. A moratorium with criminal teeth would almost certainly borrow the same mechanism, because a FLOP ceiling is the rare thing a regulator can measure without reading a model's mind. Cloud providers already meter it, and the chips are physical objects you can count.

That is also the proxy's weakness. Compute thresholds are a blunt instrument, and they go stale fast. [Epoch AI](https://epoch.ai/), the research group that tracks compute and scaling trends, has documented algorithmic efficiency gains that roughly halve the compute needed for a given capability every eight months or so. Under that curve, the capability you got from 10^26 FLOPs in 2024 might need an order of magnitude less by 2027. A fixed legal threshold either freezes in place and starts catching ever-smaller actors it was never meant to touch, or it has to be re-legislated constantly to keep pace.

This is why the precursor language and the agency matter far more than the headline number. Policing [production AI](/pillars/ai-engineering-production) at the frontier is a moving-target problem that a static statute handles badly. A standing regulator with authority to adjust thresholds is the only structure that survives contact with how fast the field moves. That is exactly the institution the bill refuses to let development proceed without. The design is circular, and that circularity is the real reason it stalls.

## Will the Ban Artificial Superintelligence Act actually pass?

Short answer: almost certainly not in this form, and not this Congress.

The votes are not there. A moratorium on an entire category of U.S. technology research, enforced with 20-year sentences, has no realistic path through a Senate that could not even pass the far milder provisions of SB 1047 at the state level. The economic argument against it is overwhelming for most members: a unilateral U.S. pause while other nations keep training is a hard sell, and I've written before about why treating adversarial [AI military strategy](/blog/iran-ai-military-strategy-redefine) as a serious variable changes the whole calculus.

So why introduce it? Because bills like this move the Overton window. Sanders and Casar are not naive about the vote count. They are establishing a pole, the maximalist position, against which a future compromise gets measured. If the "reasonable" outcome two years from now is a licensing regime for frontier training runs, this bill is part of why that counts as the moderate option. That is how the SB 1047-to-SB 53 shift happened, and it is the most likely legacy here too.

The doom-and-boom framing also matters. Researchers like [Yoshua Bengio](https://yoshuabengio.org/), a Turing Award winner who has become one of the most prominent voices warning about catastrophic AI risk, have given this kind of legislation intellectual cover it did not have in 2023. The bill borrows that credibility even if it overshoots on remedy.

## What it means if you build with AI today

If you are shipping [AI security](/pillars/ai-security-safety) controls, fine-tuning open models, or running agent frameworks, here is the honest read on your exposure.

1. **Application-layer builders are not the target.** If you are calling an API, orchestrating tools, or running a [local LLM](/pillars/llm-hardware-local-ai) on your own hardware, you are not training a frontier foundation model. The bill aims at the handful of labs doing billion-dollar compute runs, not at your RAG pipeline.
1. **The precursor language is the one to watch.** Recursive self-improvement is defined loosely enough that aggressive self-modifying agent architectures could, in a maximal reading, draw scrutiny. Keep a human in the loop on anything that rewrites its own weights or core objectives.
1. **Compliance muscle transfers.** The documentation, eval logging, and incident-reporting discipline that SB 53 and the EU AI Act already demand is the same muscle any future federal regime will want. Building it now is not wasted.
1. **Criminal liability changes the risk math for leadership.** Dissolution and prison are personal in a way fines are not. If a version of this ever passes, expect labs to route decisions through named accountable officers. A compliance pattern worth designing for early.
1. **Watch the agency fight, not the prison clause.** Whether a federal AI regulator gets created at all is the question that decides everything downstream. The sentence length is theater. The institution is the substance.
The bill will not become law as written. But the fact that two sitting members of Congress put "20 years in prison" and "artificial superintelligence" in the same sentence, with a straight face, tells you where the political baseline is heading. Five years ago this was a thought experiment on a forum. Now it has a bill number. The next version will be narrower, more enforceable, and much harder to dismiss, and that is the one builders should actually be reading.

Photo by Sasun Bughdaryan on Unsplash.

## FAQ

### Who introduced the Ban Artificial Superintelligence Act?

Senator Bernie Sanders, an Independent from Vermont, and Representative Greg Casar, a Democrat from Texas who chairs the Congressional Progressive Caucus. They announced it on September 3, 2026, and it was formally introduced in the Senate on September 23, 2026.

### What counts as artificial superintelligence under the bill?

The bill defines it as AI that either exceeds human ability on most cognitive tasks, or is capable of causing human extinction or catastrophic destruction. Because the two conditions are joined by "or," a narrow but dangerous system could qualify even if it isn't generally smarter than humans.

### What are the penalties under the Ban Artificial Superintelligence Act?

Companies found in violation can face judicial dissolution, meaning a court can order the business shut down. Individuals deemed responsible, including executives, can face up to 20 years in prison. These criminal penalties are far more severe than the civil fines used in most AI laws.

### Does the bill ban all AI development?

No. It targets frontier-scale development of advanced models and their "precursors," and imposes a pause until a new federal AI agency is created. Application-layer work like building on top of existing APIs, running local models, or orchestrating agents is not the intended target.

### How does it compare to California's SB 53?

SB 53, signed in 2025, focuses on transparency and incident reporting with civil penalties. The Ban Artificial Superintelligence Act goes much further, adding a development moratorium and criminal liability. SB 53 asks labs to disclose what they're doing; this bill tells them to stop.

### Is the Ban Artificial Superintelligence Act likely to become law?

Not in its current form. A full moratorium enforced with 20-year prison sentences has no realistic path through Congress. Its more likely role is to anchor the extreme end of the debate, making narrower future regulation look moderate by comparison.