The U.S. Senate prepares a federal veto against unsafe AI

Towards a federal regulatory framework: The new oversight paradigm in the U.S.

The race for leadership in artificial intelligence has entered a phase of unprecedented regulatory maturity. A bipartisan group of senators, led by John Thune, Ted Cruz, and Amy Klobuchar, is drafting a bill that aims to radically transform the oversight of the most advanced AI models. This initiative not only seeks to standardize the rules of the game in a highly fragmented ecosystem but also grants the federal government the explicit capacity to block the deployment of any model deemed unsafe. Historically, the United States has opted for a technological 'laissez-faire' approach; however, the scale of frontier models has forced the legislature to abandon this stance in favor of active surveillance similar to that applied in sectors like aviation or the pharmaceutical industry.

The pillars of the bill: A preventive approach

The draft, according to Reuters reports and parliamentary sources, rests on three fundamental pillars that seek to mitigate systemic risks before they occur, a direct response to concerns about national security and social stability:

• Duty of care: Obligates developers to design their products under strict protocols aimed at preventing catastrophic risks. This includes the mitigation of capabilities for the creation of biological, chemical, or nuclear weapons, as well as the prevention of large-scale cyberattacks.
• Federal veto power: The Executive branch would have the legal authority to stop the release of a model if it fails to pass relevant safety evaluations. Although the process would include judicial appeal instances to protect the right to innovation, the measure represents a drastic change: technology no longer enjoys a presumption of automatic innocence.
• Preemption (Federal supremacy): The law would override state AI regulations, establishing a single national standard. This is a strategic move to avoid the market fragmentation currently faced by companies like Google, Anthropic, and OpenAI, who view with concern a legal patchwork where states like California or Colorado attempt to impose their own rules.

This move responds to growing concern in Washington. As researcher Evan Hubinger of Anthropic points out, the debate over existential risk is no longer theoretical, but a variable that institutions are beginning to quantify through standardized safety evaluation metrics.

The friction: Self-assessment or public audit?

The most critical point of contention lies in the technical validation process. While the proposing trio contemplates a model of collaboration where companies lead governance with oversight, Senator Maria Cantwell insists that validation cannot be left in the hands of the companies themselves. Cantwell proposes that models be evaluated directly by national laboratories, ensuring that technical scrutiny is independent and rigorous. This debate resonates with the history of nuclear energy regulation, where self-assessment was discarded in favor of mandatory external inspections.

AI regulation in the U.S. is at a crossroads: the balance between fostering innovation from giants like OpenAI, Google, and Anthropic, and the need to safeguard public safety. Unlike the European Union's AI Act, which is more prescriptive and horizontal, this U.S. project focuses specifically on frontier models with advanced capabilities. The strategy is clear: prevent excessive bureaucratic burden from stifling emerging startups while placing a legal 'handbrake' on corporations that, due to their scale, could disrupt public order.

Context, impact, and speculation

While the project has unusual bipartisan support, its implementation faces considerable logistical challenges. The government's technical capacity to evaluate models that evolve weekly is, as of today, a subject of skepticism. Analysts warn that overly strict regulation could incentivize brain drain or development in jurisdictions with laxer legal frameworks. It is important to note that, at the time of this analysis, specific details regarding fines for non-compliance and the exact definition of an 'advanced model' remain under negotiation and are subject to drastic changes before reaching the Senate floor.

Compared to previous events, such as the regulation of the internet in the 90s, this law marks the end of the era of voluntary self-regulation. The industry, which until now was governed by 'voluntary commitments' to the White House, faces an enforceable legal mandate for the first time. The market will react with caution; companies that already have robust 'Safety' and 'Alignment' teams could benefit from a barrier to entry that prevents competition from less prepared actors, further consolidating the power of the current sector leaders.