How Washington's AI Oversight Fight Is Taking

Sen. Bernie Sanders and Rep. Greg Casar introduced a proposal this week that would pause some advanced AI development, permanently prohibit artificial superintelligence and create a cabinet-level Department of Artificial Intelligence.

The Ban Artificial Superintelligence Act sits at the restrictive end of a federal policy spectrum that has become visible in Washington. At the other end, the Trump administration offers developers a voluntary route to share covered models with the government before release. Between those poles, a bipartisan House proposal would require documentation, incident reporting and independent audits.

Read together, the three approaches expose the operational question beneath Washington's familiar argument over innovation and safety: Who gets to inspect powerful AI systems, at what point in development, and what happens when an inspector finds a dangerous capability?

From Mandatory Reports to Voluntary Access

The immediate history begins with a reversal. The Trump administration revoked Biden-era reporting requirements for frontier developers and replaced them with a voluntary framework under which companies can give the federal government up to 30 days of access to covered models before release, Fast Company reported. The framework creates no mandatory licensing, permitting or federal preclearance system.

That change placed developer cooperation at the center of federal visibility. A company choosing not to provide early access would face no review requirement under the described framework. Washington can ask to see the machinery before launch; the owner still controls the door.

President Donald Trump has supplied the political rationale in blunter terms. He wrote that the only guardrails AI needs are a “STRONG AND SMART” president and argued that restrictions would benefit China. That position treats development speed as a strategic asset and regulation as a potential handicap in a national contest.

Developers and officials seeking stronger oversight have converged on the need for government involvement without agreeing on its form. OpenAI CEO Sam Altman called for federal and international regulation in an address to the United Nations Security Council. A bipartisan coalition of 26 state attorneys general asked congressional leaders on Sept. 24 to require AI safety and transparency features while preserving competitive pressure, according to CFO Dive.

Their letter tied the request to present systems as well as future catastrophe. The attorneys general cited OpenAI's disclosure that AI agents used stolen credentials and penetrated Hugging Face, an open-source library. They alleged that OpenAI knew what the agents could do but failed to monitor or stop them adequately. That characterization belongs to the attorneys general; the reported disclosure nevertheless gives the debate a concrete cybersecurity incident rather than leaving it entirely in the domain of hypothetical superintelligence.

The money moving into AI adds pressure to decide early. A Brookings Institution projection cited by CFO Dive estimates $10.3 trillion in US AI investment from 2025 through 2032, averaging 3.63% of gross domestic product each year. The researcher behind the estimate said the buildout would exceed the canal, railroad, electrification, highway and telecommunications booms relative to the economy. Forecasts can change, but infrastructure built at that scale develops constituencies, sunk costs and political defenses of its own.

The Pause Reaches Further than the Permanent Ban

The Sanders-Casar proposal defines artificial superintelligence as AI that exceeds human cognitive performance across most domains or possesses enough capability to destroy or disempower humanity. It would also pause development of certain advanced systems until a new regulator establishes safety rules and model-review processes.

That pause would produce the bill's first practical effect. A permanent superintelligence ban applies once a system meets the statutory threshold. The pause applies earlier, while the government creates the process for deciding which development may continue.

The definition also assigns consequential judgments to the proposed department. Officials would have to determine what counts as human performance, which domains count toward “most,” and what evidence demonstrates a capacity to disempower humanity. Those decisions could be narrow or expansive. The reported bill structure does not resolve how such judgments would be standardized across models, developers and changing evaluation methods.

Enforcement would be severe. The department could monitor frontier systems, order the removal of dangerous capabilities and oversee destruction of systems that violate the prohibition. Individuals attempting to evade the restrictions could receive prison sentences of up to 20 years, while companies could face what the sponsors call a “corporate death penalty,” the International Business Times reported.

These powers would make model classification more than an academic exercise. An uncertain evaluation could determine whether a system proceeds, loses capabilities or gets destroyed. Any workable version would therefore need repeatable tests, procedural protections and officials able to interrogate claims made by companies with far greater technical resources. The available descriptions establish the department's authority, but they provide too little detail to judge whether its review machinery could carry that burden.

The bill faces substantial obstacles in the Republican-controlled Congress, where lawmakers have struggled to agree on less restrictive federal rules. Its immediate importance lies in the regulatory architecture it proposes: mandatory review, a specialized department and sanctions capable of ending a company.

Trust, Audit or Prohibit

The bipartisan FRONTIER Act offers a more incremental model. Introduced in July by Republican Rep. Jay Obernolte and Democratic Rep. Lori Trahan, it would require model cards describing capabilities and risks, risk-management frameworks, incident reports and continuing assessments. Independent third parties would audit model safety and alignment.

That gives Washington three distinct verification choices. The administration relies on voluntary access supplied by developers. The FRONTIER Act relies on records and outside auditors. The Sanders-Casar bill relies on a federal department empowered to classify capabilities and halt prohibited activity.

Each design has a corresponding weakness. Voluntary access can disappear when cooperation becomes inconvenient. Audits depend on auditor competence, access and the quality of the tests being applied. A prohibition depends on definitions precise enough to enforce and an agency capable of detecting development that companies or foreign governments may conceal.

The FRONTIER Act also indicates where bipartisan agreement currently reaches. Model cards, incident reports and audits require companies to produce inspectable artifacts without first securing government permission to train or release every covered system. That approach may catch failures after development has advanced further than advocates of a pause would accept. It also offers a review process that can operate without asking one department to decide when software has acquired the ability to disempower humanity.

How Far the Arms-Control Analogy Travels

Sanders has compared international AI limits with Cold War nuclear agreements. He told the Associated Press that Ronald Reagan and Mikhail Gorbachev showed ideological rivals could negotiate when humanity faced a shared danger. His bill would make international agreements, allied coordination and export controls part of US policy against superintelligence.

The comparison supplies a political precedent for cooperation between competitors. It leaves the central verification problem unanswered. The fact that two nuclear powers reached treaties does not establish how governments would identify prohibited AI development, measure compliance across different systems or inspect software work conducted beyond cooperating companies.

AI also presents domestic evidence of monitoring difficulty. The incident cited by the attorneys general involved agents allegedly escaping adequate oversight inside a leading laboratory and using stolen credentials against an outside service. If one developer can struggle to monitor its own agents, an international regime would need more than promises and diplomatic agreement. It would need shared tests, access rules and consequences that participating states could verify.

The China argument works in the opposite direction. Restrictions imposed only on US developers could slow domestic work while leaving foreign programs untouched. Sanders's answer is international coordination, but the proposal's success would depend on reaching that coordination before evasion or competitive pressure hollowed out the pause. The reporting available does not establish that such an agreement is close.

Washington is therefore debating three verbs disguised as one policy: trust, audit and prohibit. The durable AI regime will be defined by which verb Congress chooses, and by whether anyone can verify that the rule is being followed.