US introduces a federal AI regulation bill: what changes and who it affects
A bipartisan group of senators introduced a federal bill seeking to establish a single regulatory framework for AI in the United States, with transparency requirements, a federal model registry, and limits on state regulations.
Key Takeaways
The law would create a mandatory federal registry for high-risk AI models
Requires impact assessments before deploying systems that affect decisions about people
Limits states' ability to create their own stricter regulations
Establishes an 18-month transition period for compliance
Includes exemptions for academic research and open-source models below a threshold
The regulatory landscape for artificial intelligence in the United States has just taken a significant turn. A bipartisan group of senators has introduced a draft federal bill that, if passed, would establish the first binding national rules for the development and deployment of AI systems.
1What does the bill propose?
The draft bill, called the AI Accountability and Transparency Act, is structured around five main pillars:
Pillar 1: Federal model registry
Every AI model exceeding a computational capacity threshold (measured in training FLOPs) must be registered with a new federal body. The registry includes information about training data, safety tests performed, and intended use cases.
Pillar 2: Mandatory impact assessments
Before deploying an AI system in "high-risk" contexts (employment, credit, healthcare, criminal justice, education), companies must conduct and publish impact assessments analyzing potential biases and risks.
Pillar 3: Algorithmic transparency
Users affected by automated decisions have the right to a comprehensible explanation of the process. This does not mean revealing source code, but providing a clear description of the factors that influenced the decision.
Pillar 4: State limitation
One of the most controversial provisions: federal law would preempt state-level AI regulations, with some exceptions for existing consumer protection laws.
Pillar 5: Enforcement and sanctions
The FTC (Federal Trade Commission) would be the primary enforcement body, with the power to impose fines of up to 4% of annual global revenue for serious violations.
2The state limitation debate
Whether the federal government should prevent states from independently regulating AI is perhaps the most divisive aspect of the entire proposal.
Proponents argue that a patchwork of state regulations creates an impossible environment for companies operating nationally. California, New York, and Illinois already have AI-specific laws, and more states are developing theirs.
Critics point out that federal regulation tends to be more lenient than state regulation, and removing states' ability to regulate eliminates an important consumer protection mechanism.
3Who is affected?
Big tech companies
OpenAI, Google, Meta, Anthropic, and Microsoft would be most directly affected. The federal registry and impact assessments would add significant costs, but would also provide legal clarity that currently does not exist.
Startups
The impact on startups depends on the threshold. The draft includes exemptions for companies with fewer than 50 employees and models below a certain size, but details are still under discussion.
Academic researchers
The law includes a broad exemption for academic research, which has been well received by the scientific community.
Open-source models
Open-source models below the registration threshold would be exempt, but those exceeding it would have the same obligations as commercial models. This has generated debate about whether it disincentivizes large-scale open-source.
4Comparison with the European AI Act
📊 The US law is less prescriptive than the European AI Act in several ways:
- Does not ban entire categories of use (like social scoring)
- Does not require pre-market certification for high-risk systems
- Gives more weight to industry self-regulation
- Includes more exceptions for national security
💡 The US approach prioritizes innovation over precaution, while the European approach does the opposite. Both models will be closely watched by the rest of the world.
5Next legislative steps
The draft still has a long road ahead. It needs to pass through Senate committees, possibly be amended, and then voted on in both chambers. Analysts estimate that, if it advances, it could become law by late 2026 or early 2027.
Meanwhile, the industry is preparing. Several large companies have already begun voluntarily implementing practices aligned with the draft, anticipating that regulation is inevitable.