A view from DC: A bipartisan blockbuster bill on AI

For many weeks now, the artificial intelligence policy community in Washington has been anticipating — and engaging vocally around — an expected legislative framework in the U.S. House of Representatives. It was known that this effort would be led by Rep. Jay Obernolte, R-Calif., long the leading Republican voice on AI regulation. It is rumored that the White House and Obernolte have worked together to think through the priorities that would be reflected in this bill, though it has been unclear the extent to which the effort is coordinated, or which version of the administration's priorities might be reflected. 

Recently, it became clear that Rep. Lori Trahan, D-Mass., was negotiating with Obernolte to bring bipartisan legitimacy to the legislative framework. Both are prominent members of the Subcommittee on Commerce, Manufacturing, and Trade within the House Energy and Commerce Committee, though neither have leadership roles on the committee.

Obernolte and Trahan released a discussion draft of their framework for public comment, calling it, with a level of pomp and circumstance befitting the current political moment, the Great American AI Act. Also signing onto the release were Reps. Scott Franklin, R-Fla., Suhas Subramanyam, D-Va., Erin Houchin, R-Ind., and Scott Peters, D-Calif. Of these, only Houchin and Peters are part of the Energy and Commerce committee.

A federal response to foundation model oversight

It is probably not a coincidence that this draft comes so soon after the Illinois legislature passed SB 315, the state's response to the similar foundation model transparency laws in California and New York. It also comes the same week that President Donald Trump signed an executive order signaling a new pivot toward oversight of these most powerful dual-use models, at least when it comes to their national security implications. 

Notably, the federal framework builds on the main distinguishing characteristic of the Illinois bill, a requirement for independent third-party audits. The Obernolte-Trahan framework would dramatically expand this concept, requiring large developers to undergo semi-annual audits conducted by specialized, state-licensed Independent Verification Organizations, and providing a liability shield for these IVOs.

In other ways, the foundation model regulations are scoped remarkably similar to all three state frameworks, including transparency obligations, incident reporting requirements and whistleblower protections, meaning it would nationalize the model that is being hammered out in the states. But it would vastly expand the liability caps in these laws, authorizing up to USD1 million per violation, per day, rather than capping the total liability for incidents at between USD1 and USD3 million, respectively, as the state laws do.

Before Trump's recent pivot on foundation model oversight, this primary focus area of the discussion draft would have seemed discordant. After all, it was less than three months ago that the White House's National Policy Framework for Artificial Intelligence was clear to say Congress must "not create any new federal rulemaking body to regulate AI," instead supporting oversight through existing, sector-specific agencies and decentralized, industry-led standards.

Though it would not create a new federal agency, per se, the draft formalizes and prioritizes the role of the Commerce Department's Center for AI Standards and Innovation, which has been serving as the federal government's voluntary testing ground for foundation model developers. As envisioned, the bill would formally authorize the agency's existence and role while providing USD100 million in annual funding to continue standard-setting and build and oversee the IVO auditing program.

And though this aligns strongly with the administration's new push for vetting foundation models, it is worth noting that the proposed regime, though arguably stronger than the current state frameworks, is not one of pre-market review or licensing. It relies entirely on post-training, private-sector audits rather than granting a federal agency veto power over a model's release.

Preemption for me, but not for thee

The framework will likely get plenty of attention for its approach to preemption, showcasing the continued interest from U.S. legislators to stop the spread of state AI laws. This time, however, the preemptive effect is limited to state laws that target the development of AI models, which appears to allow for the proliferation of state laws regulating AI use or deployment. The preemption provision also would sunset after three years.

A grab bag, not an omnibus

Beyond its core foundation model requirements, the legislative framework reflects a variety of legislative priorities that Obernolte has previously supported, most notably in the Bipartisan House Task Force Report on AI released at the end of last term. But this draft differs wildly from that report in many ways.

On the proactive side, the draft bill would include a few additional notable legal changes. It would increase the maximum fines for mail, wire and bank fraud if committed with the assistance of AI and dramatically heighten penalties for the algorithmic impersonation of federal officials.

Reflecting a major conservative priority, and the White House framework, the bill mandates a comprehensive study of "jawboning," the informal pressure federal agencies exert on AI platforms regarding content moderation and output generation, calling for federal oversight to seek out any such practice.

One major omission is the lack of reference to data scraping and copyright infringement in the training of AI models, which was a major theme in the bipartisan report and remains a priority for the White House. The White House framework goes so far as to suggest Congress should enable "licensing frameworks or collective rights systems for rights holders to collectively negotiate compensation from AI providers," and to create explicit federal protections against unauthorized digital replicas. The draft is silent on these issues.

Less surprising is the lack of mention of discrimination or bias by AI models, which in the current legislative cycle became anathema to Republican regulators, despite being mentioned in the bipartisan report last term.

Nor is the lack of kids' safety provisions a surprise, given the existing major effort to include AI-related youth safety and privacy protections in the package of bills already under consideration in the Energy and Commerce Committee. The draft framework would, however, include educational grant funding and support for K-12 AI literacy initiatives.

The lengthy framework does include many other ideas, including lots of initiatives around workforce development and the ongoing tracking of data related to AI's impact on labor, AI testbeds and standards development, and general support for AI research and development initiatives. It also throws in reauthorization of the Cybersecurity Act of 2015 to enable critical information sharing between firms, plus a few AI-related cybersecurity studies and reports, focused on open-source models. 

Overall, the new framework reaffirms the importance of catastrophic risks from foundation model development in the current U.S. policy debate, one of the only flavors of AI governance legislation to gain steam in 2026. This signals how the laissez-faire approach that characterized the early phase of the 119th Congress is starting to meet resistance due to national security concerns. Though the populist and economic nationalist arguments for additional AI regulation from some conservative corners may not have yet caught on, the door appears to be open for Republican-supported AI legislation.

Whether this idea remains focused only on the most catastrophic risks, and whether even those will be a priority for the federal Congress, will depend on the reactions to this new legislative development. Discussion draft season has begun, albeit somewhat belatedly.

Please send feedback, updates and jaw bones to cobun@iapp.org.