Recent/upcoming developments… Given the increasingly negative public sentiment around AI, policymakers are moving in more aggressive direction on a variety of fronts.  This includes President Trump signing an Executive Order (EO) on government review of frontier models, House Republican leaders backing a legislative proposal around AI regulation, and various leading Democratic lawmakers beginning a discussion around taxing AI firms/data centers to fund various progressive priorities.

* Trump’s EO establishes a voluntary framework for AI companies to provide the federal government with access to covered frontier models 30 days before public release, while also directing national security and cyber officials to work with agencies, critical infrastructure operators, and AI companies on vulnerabilities identified by advanced models.  The final order is narrower than the version delayed last month, which contemplated a 90-day review window.  The final EO reflected a compromise among competing White House factions: David Sacks and industry-aligned advisers pushed to keep the framework voluntary and avoid slowing model releases, Defense Secretary Hegseth and Under Secretary Michael favored stronger guardrails, while Treasury Secretary Bessent and White House Chief of Staff Wiles occupied a middle-ground position.  Various AI safety advocacy groups (e.g., Alliance for Secure AI, Americans for Responsible Innovation) framed it as a constructive first step toward federal oversight but argued that its voluntary framework is insufficient and should be codified as mandatory pre-release review for advanced models.

* House Republican leadership (i.e., Speaker Johnson (R-LA)) indicated support for a legislative framework introduced by Reps. Obernolte (R-CA) and Rep. Trahan (D-MA).  The Obernolte/Trahan bill requires reports on critical safety incidents/catastrophic risk, and semiannual audits by licensed independent verification organizations, but does not require AI firms to submit their models to the government for review before release.  It would preempt state laws specifically regulating AI model development for three years, while preserving generally applicable state laws and post-deployment AI rules.  The preliminary reaction from AI safety advocacy groups is that the bill does not go far enough.

* Progressive Democrats are beginning to move beyond conventional AI safety regulation toward tax, public-ownership, and data-center moratorium proposals.  Sen. Sanders (I-VT) is preparing legislation that would create a sovereign wealth fund through a one-time 50% equity tax on the largest AI companies, while Rep. Casar (D-TX) and Sen. Warren (D-MA) have argued for taxing AI firms to address job displacement and other harms.  California Gov. Newsom has separately floated ideas such as universal basic capital, public equity funds, and dividends.

Our outlook… The Trump EO represents a material increase in government regulation of AI.  And though the review process is voluntary, we expect the industry will be under pressure to comply, and that this new layer of review could slow the release or expansion of new models (as we have already seen in the Mythos situation).  On the legislative front, we do not believe the Obernolte/Trahan bill will be enacted this year (20-25% probability) as we do not believe lawmakers will be able to reach bipartisan consensus on the breadth of state-level preemption and the strength of a federal regulatory standard.  Following the midterm elections, Democrats gaining a majority in at least one chamber of Congress (likely the House, possibly the Senate) will result in more headline risk for industry (e.g., hearings, letters to the administration and/or companies, movement of bills through committee process), though the probability of legislative action will remain low.  However, the risk to industry will become more tangible/acute if a Democrat is elected President in 2029, as a Democratic administration will likely push for mandatory model review, an expansion of what the federal government reviews (not just cybersecurity but also issues around bias and worker displacement), and a tax on AI/data centers.

* We do not expect the Obernolte/Trahan bill to be enacted in the near term. The core problem is preemption: Democrats will likely view the bill as turning federal law into a ceiling that blocks CA, NY, and other states from going further on model-development rules, while many Republicans and industry allies will likely view the bill as too limited because the preemption is temporary and still leaves room for further state regulation of post-deployment AI uses.  We also anticipate that there will be inter- and intra-party divides over the scope of the federal guardrails included in the bill (i.e., some Democrats and populist/anti-tech Republicans will want mandatory pre-release review and approval; some Republicans will want fewer obligations on industry).  That said, the bill goes further than last year’s pure moratorium approach by pairing preemption with federal reporting requirements, suggesting at least some Republicans are moving closer to Democrats on the need for some federal guardrails even if the parties remain far apart on whether federal law should be a floor or a ceiling.

Watch for these developments… We are watching for whether Democratic leaders (e.g., House Minority Leader Jeffries (D-NY), House Energy and Commerce (E&C) Committee Ranking Member Pallone (D-NJ)), key players on AI such as Rep. Beyer (D-VA), as well as progressives and Republican populists who have been assertive on AI policy (e.g., Sens. Blumenthal (D-CT), Hawley (R-MO), Blackburn (R-TN), Reps. Ocasio-Cortez (D-NY), Jayapal (D-WA)) embrace the Obernolte/Trahan bill.  If there is dissent among Democrats and the bill passes on a partisan Republican basis in the House, that suggests it is unlikely to gain the bipartisan momentum necessary for progress in the Senate.  If instead, the bill achieves broad bipartisan support in the House (which we do not expect to occur), its probability of enactment will rise.

* In order to assess the likely AI agenda under a Democratic administration in 2029+ and where sentiment is moving generally, we are watching for the positions being adopted by the emerging Presidential candidates (e.g., Sen. Kelly (D-AZ), Gov. Moore (D-MD), Gov. Shapiro (D-PA), Rahm Emanuel, others).  Among the key dividing lines are whether model reviews should be voluntary or mandatory, if/how model reviews should be expanded beyond just an assessment of cybersecurity, and whether there is receptivity to Sanders’ AI tax/public ownership concept.