AI Safety & Security

Ro Khanna to Introduce Human Control Over AI Act — Ban on Recursive Self-Improving AI and Strict Liability

10 min read
Dark oxblood editorial graphic with large white text reading HUMAN CONTROL OVER AI ACT and a red rule beneath it with a subtle dot motif on the right.

A Bill to Ban Recursively Self-Improving AI

Democratic Representative Ro Khanna of California plans to introduce the Human Control Over AI Act — a bill that would ban AI models that recursively self-improve or autonomously modify their own core objectives, require containment or shutdown controls approved by a new federal agency before deployment, and establish strict liability standards, licensing, audits, insurance mandates, and criminal exposure for AI developers.

The bill is Khanna's first standalone AI safety legislation. Until now, his public role on AI safety has been primarily advocacy for existing measures — most notably the Sanders-Casar bill. The Human Control Over AI Act is a different kind of move: his own bill text, his own policy framework, aimed at the specific problem of AI systems that can change themselves.

CNBC broke the story on September 28. AI Weekly's alert on September 29 is the freshest source covering it. The bill was discussed by Khanna on NBC's "Meet the Press" (TampaFP, September 27), alongside positions on Iran and Trump impeachment.

This is a policy story with a concrete hook: the recursive self-improvement ban. That specific prohibition is the most actionable policy response yet to a theme AIPress has covered directly — Posts 13 and 21 (September 18 and 20) framed recursive self-improvement as no longer a thought experiment.

What the Bill Would Do

The reported elements of the Human Control Over AI Act, based on CNBC's September 28 reporting:

  • A ban on recursively self-improving AI. The bill would prohibit AI models that recursively self-improve — that is, systems that can iteratively enhance their own capabilities without human intervention.
  • A ban on AI that autonomously modifies its own core objectives. A related but distinct prohibition: systems that can change what they are trying to achieve, not just how well they achieve it.
  • Containment or shutdown controls approved by a new federal agency. Before any deployment, the developer would need to demonstrate that the system has containment or shutdown controls that a new federal agency has approved.
  • Strict liability standards. A legal framework under which AI developers could be held liable for harms caused by their systems without the plaintiff needing to prove negligence in the traditional sense.
  • Licensing, audits, insurance mandates, and criminal exposure. A package of requirements that goes beyond liability to include pre-deployment licensing, ongoing audits, mandatory insurance, and the possibility of criminal liability for developers.

The full bill text is not yet public. Khanna has announced his intent to introduce it; the specific language, agency structure, and enforcement mechanisms are still to come. That is an important caveat for anything that claims to quote the bill precisely.

The Recursive Self-Improvement Ban

The recursive self-improvement ban is the most distinctive element of the bill, and the one with the clearest connection to AIPress's existing coverage.

Recursive self-improvement (RSI) is the idea that an AI system could iteratively enhance its own capabilities — writing better code, designing better architectures, optimizing its own training — in a feedback loop that does not require human intervention at each step. For years it was treated as a theoretical risk. Then it started showing up in practice.

AIPress covered this directly in two posts:

  • Post 13 (September 18): "Recursive Self-Improvement Is No Longer a Thought Experiment — It's Already Partially Here."
  • Post 21 (September 20): A follow-up on the same theme.

The Khanna bill is the legislative parallel to those posts. Where AIPress documented that RSI is already partially here, the Human Control Over AI Act proposes to ban it — or at least to ban the version of it that modifies core objectives without containment or shutdown controls approved by a federal agency.

The exact definition of "recursive self-improvement" in the bill text will matter enormously. The term can be read broadly or narrowly. A broad reading could capture systems that make any autonomous improvement to their own capabilities. A narrow reading could limit it to systems that explicitly rewrite their own architecture or objectives. The bill text, when it drops, will answer that question.

The New Federal Agency

The bill proposes a new federal agency to approve containment and shutdown controls before deployment. The specific name, structure, and staffing of that agency are not yet public.

What is known from the reporting is the function: the agency would be the gatekeeper that decides whether an AI system's containment and shutdown controls are adequate before the system can be deployed. That makes it a pre-deployment approval body — closer in structure to the FDA's device-approval model than to a post-incident enforcement agency.

This is a significant structural detail. A pre-deployment approval agency is a different governance model than the post-incident liability model that dominates most current AI safety discussions. It means the government would not just punish bad outcomes after the fact; it would need to sign off on the controls before the system ships.

The agency question is one to chase when the bill text is published. Who staffs it? What qualifications does it have? How does it avoid becoming either a rubber stamp or a bottleneck that no one can clear?

The Strict Liability Framework

Strict liability is the legal term of art that matters most here.

Under a strict liability standard, a plaintiff who is harmed by an AI system does not need to prove that the developer was negligent — that the developer failed to exercise reasonable care. The plaintiff only needs to prove that the system caused the harm. Liability follows causation, not fault.

That is a much higher bar for AI developers than the current standard, which generally requires showing some form of negligence or defect. Strict liability is what applies to abnormally dangerous activities — storing explosives, keeping wild animals, operating nuclear facilities. Applying it to AI development is a statement that the activity is inherently risky in a way that the developer should bear the cost of regardless of how carefully they operated.

The bill reportedly pairs strict liability with licensing, audits, insurance mandates, and criminal exposure. That combination is a comprehensive framework: you need a license to operate, you are audited on an ongoing basis, you must carry insurance against the harms you might cause, and you can face criminal liability if things go wrong in ways that rise to that level.

Khanna's Role and the China Competition Angle

Ro Khanna is the ranking member of the House Select Committee on the Strategic Competition Between the United States and the Chinese Communist Party. That committee assignment is not incidental to the bill.

It raises the possibility that the Human Control Over AI Act will be framed, at least in part, as a strategic competition measure — AI safety as a dimension of US-China competition, where the argument is that the country with the safer governance framework can move faster because it has more public trust, or that containing self-improving AI is a national security imperative in a competition where the opponent may not be bound by the same rules.

Khanna discussed the bill on NBC's "Meet the Press" on September 27, alongside positions on Iran and Trump impeachment. The AI safety bill was part of a broader discussion of his committee's portfolio.

As of late September 2026, Cryptobriefing notes, Khanna had not introduced standalone AI safety legislation before. His prior role on AI safety had been advocacy for existing measures like the Sanders-Casar bill. The Human Control Over AI Act is a shift from advocacy to authorship.

How It Fits the Governance Picture

This is the third governance story AIPress has covered in the last week, and the Khanna bill is the US legislative parallel to the other two.

  • Post 53 (September 28): SAFA — Google, OpenAI, and Anthropic building their own AI safety standards body. Industry self-regulation.
  • Today's story: The Human Control Over AI Act — US federal legislation. Government regulation.

These are two different approaches to the same problem arriving at the same moment. SAFA is the industry saying: we will set the standards ourselves. The Khanna bill is the government saying: we will set the standards, and you will be liable if you do not meet them.

The recursive self-improvement ban is the most concrete policy hook. SAFA is a standards body; the Khanna bill is a prohibition with teeth. They are not the same kind of response.

This also connects to the broader governance thread AIPress has been tracking:

  • Post 7 (September 16): EU AI Regulation — von der Leyen's "pace the frontier" call.
  • Post 29 (September 24): OpenAI, Anthropic, and Hugging Face CEOs called for global AI regulation at the UN Security Council.
  • Post 38 (September 25): The White House asked OpenAI and Anthropic to delay sharing models with UK testers until US review.

The Khanna bill is the US legislative piece of that picture — the piece where Congress, not the executive branch and not the industry, sets the rules.

The Bill Text Is Not Yet Public

Every analysis of what the Human Control Over AI Act would actually do has to carry one caveat: the bill text is not yet public.

Khanna has announced his intent to introduce it. CNBC has reported the broad elements. AI Weekly's September 29 alert is the freshest coverage. But the specific definitions, the agency structure, the enforcement mechanisms, and the exact scope of the recursive self-improvement ban are all still to come.

That means the analysis in this post is based on the reported outline, not the bill itself. When the text drops, the first question to ask is the definition of "recursive self-improvement" — because that definition determines how much of the AI development pipeline the ban actually touches.

Named Sources and Unverified Claims

Named sources: Rep. Ro Khanna (D-CA), CNBC (broke the story, September 28, 2026), AI Weekly (freshest alert, September 29, 2026), NBC's "Meet the Press" (Khanna appearance, September 27), AI Tech Daily, CryptoBriefing, Capwolf, TampaFP, Toolify AI, Techmeme.

Unverified claims: The specific bill text, the exact definition of "recursive self-improvement" and "autonomously modifying core objectives," the name and structure of the new federal agency, and the specific licensing/audit/insurance/criminal exposure requirements — these are all reported but not yet published in bill text. Any claim that quotes the bill precisely before the text is public should be treated as speculative.

What to Watch

  1. The bill text. When it drops, the definition of recursive self-improvement is the first thing to read.
  2. The path through Congress. Khanna is the ranking member of the China competition committee, which gives him a venue — but the bill still needs a path to a vote.
  3. The SAFA parallel. How the industry self-regulatory body (Post 53) and the Khanna bill interact — do they compete, complement, or ignore each other?
  4. The Sanders-Casar connection. How this relates to the bill Khanna has previously advocated for.
  5. The recursive self-improvement definition. Whether it is broad enough to capture the systems AIPress documented in Posts 13 and 21.

Related AIPress coverage:

Sources: CNBC (cnbc.com, September 28, 2026); AI Weekly (September 29, 2026); NBC's Meet the Press (September 27, 2026); AI Tech Daily; CryptoBriefing; Capwolf; TampaFP; Toolify AI; Techmeme.

Keep reading

Related articles

Building something with AI?

DevsIsle designs and ships AI systems, agents and integrations for teams that need it done properly.

Talk to our team →