Anthropic's draft AI safety bill sparks debate over high risk thresholds

A framework for intervention

Anthropic is currently circulating a draft AI safety bill with lawmakers that outlines specific conditions for federal oversight. Unlike many existing regulations, this proposal is not a formal piece of legislation yet; it is a framework designed to tell the government when it is permitted to act against a frontier AI model.

The core of the proposal is a very narrow path for enforcement. Under this plan, only the U.S. Department of Commerce would have the authority to ask a judge to suspend an AI model's development or use. Every other federal agency would be strictly prohibited from taking independent action, even if they believed a model was dangerous. This creates a single point of entry for regulation, moving away from the multi-agency approach typically seen in environmental or consumer protection laws.

To make this work, the bill would require AI developers to conduct internal testing and share those results publicly. These records would serve as a paper trail, making it easier for the Department of Commerce to prove to a judge that a specific risk was predictable rather than a surprise.

The high bar for government action

The most controversial part of the draft involves the specific numbers required to trigger a government block. As reported by shattered.io, the Department of Commerce could only seek a court order if an AI model creates an imminent and foreseeable risk in one of three areas: cyberattacks, the model escaping its own safety controls, or the creation of biological weapons.

Crucially, these actions must also result in more than $1 billion in damages or cause more than 50 deaths. If a model causes significant harm that falls below these numbers—such as large-scale fraud that costs less than a billion dollars or an incident that causes dozens of injuries—the draft provides no mechanism for federal intervention. This specific threshold has caused a split in media coverage, with some outlets viewing it as a necessary limit on government overreach and others arguing it sets the bar for safety far too high.

The framework also includes a 72-hour notification window. Independent evaluators registered with the Department of Commerce would have three days to report a qualifying risk. However, the actual judicial process for a court order would follow the standard timeline of the federal court system, which is not defined in the draft.

Strategic goals and reactions

Anthropic has spent the last two years building a public narrative that frontier AI poses catastrophic risks. This draft bill appears to be an attempt to turn those internal safety standards into a legally binding national framework. By proposing a narrow, judicially-reviewed standard, Anthropic may be trying to create a predictable environment that is easier for companies to navigate than a confusing patchwork of 50 different state laws.

While competitors like OpenAI, Google, and Meta have not officially commented on the draft, their silence may be strategic. A single federal standard that preempts state-level rules is often desirable for large tech firms because it simplifies compliance. However, critics argue that by centralizing power in one department and one court, the bill creates a bottleneck where a single agency's judgment becomes the only safety valve for the entire country.

The debate reflects a broader public skepticism regarding AI. With a majority of U.S. adults reporting low confidence in the government's ability to regulate AI effectively, the specific numbers in Anthropic's pitch are being scrutinized by both those who fear government overreach and those who fear the technology is moving too fast for current protections.

The path forward

The draft represents a bet on 'predictability.' Anthropic argues that clear, well-defined rules are safer for the industry to build around than broad discretionary authority. By setting a high bar for intervention, the company seeks to avoid the unpredictable enforcement that many tech firms have lobbied against in the past.

Moving forward, the focus will be on whether lawmakers see this as a reasonable compromise or a way for the industry to self-regulate the most dangerous aspects of the technology. Because the bill is currently a draft, the next step is to see if any committee schedules a hearing or if other labs propose competing frameworks that offer lower thresholds for intervention.

Comparison of oversight models

To understand how this differs from other global approaches, it helps to look at the current landscape. The European Union, for example, uses a tiered risk classification that treats different types of AI systems differently based on their specific use cases. In contrast, Anthropic's draft ignores these broad categories in favor of three specific catastrophic scenarios.

Several U.S. states, including California and Colorado, have already moved forward with their own safety and transparency requirements. Anthropic's proposal is a reaction to this fragmented landscape. It suggests that a single, narrow federal standard might be more durable and more palatable to the industry than trying to comply with dozens of different state-level rules simultaneously.