Open-Weight AI in the Congressional Hearing Record

A Radiant Intel review of 114 written witness statements found six affirmative open-model advocate slots in a corpus where 21 non-advocate slots used frontier-risk, model-access, model-release restriction, or China access-control frames.

Finding

Radiant Intel reviewed official House and Senate hearing pages and written witness statements for AI-centered congressional hearings from June 27, 2025, through June 27, 2026. The compiled corpus includes 30 hearings and 114 written witness-statement slots: 83 House slots across 22 hearings and 31 Senate slots across 8 hearings.

Only 6 of those 114 slots, representing 5 unique witnesses, made an explicit affirmative case for open-source or open-weight AI models. A broader frontier-risk, model-access, model-release restriction, or China access-control frame appeared in 22 unique slots. One of those 22 slots was also an open-model advocate, so the non-advocate frontier-risk/model-release/model-access/China access-control group consisted of 21 slots.

That is the central finding: affirmative open-model testimony was rare in the reviewed written hearing record, while adjacent risk and control frames appeared more often. The finding does not show that Congress should hear less from AI risk experts. Much of the reviewed risk-focused testimony addresses real security, cyber, biosecurity, and strategic-competition concerns. It does show that the written hearing record gives open-model arguments a small footprint relative to the policy stakes.

If Congress considers model-release restrictions, licensing regimes, export controls, procurement rules, or safety mandates that could shape the future of open models, committees need a more complete record on the strategic value and operational trade-offs of open-source and open-weight AI.

Results

The model-openness coding below is mutually exclusive. Each witness slot appears in one row.

The three openness-adjacent slots consisted of two civic open-source software references and one open-weight market discussion that did not make a model-release policy recommendation.

The risk and access-control frames below are not mutually exclusive with the model-openness coding. A witness slot can be both an open-model advocate and part of a China/frontier access-control hearing frame. The combined rows therefore count unique witness slots, not a sum of the rows above them.

The strict open-model advocate group consisted of six witness slots:

Those witnesses did not make identical arguments. Some emphasized civil liberties and transparency. Some emphasized enterprise trust, competition, and auditability. Some emphasized U.S. competition with Chinese open models. Some emphasized the likely diffusion of open-weight models and the need for American open alternatives. The common thread is narrower: each made an affirmative case for open models rather than treating openness principally as a threat.

What the finding shows

The reviewed hearing record is not an open-versus-closed referendum. Most AI testimony in the corpus did not address model openness at all. The largest groups focused on workforce, education, labor, health care, financial services, copyright, child safety, government operations, data centers, and sector-specific AI deployment.

The imbalance appears when hearings move from AI deployment to advanced model governance. In that part of the record, open weights more often appeared as part of a security, China, cyber, biosecurity, theft, distillation, or access-control discussion than as an innovation, transparency, competition, resilience, or strategic-diffusion asset.

The finding does not prove congressional intent, institutional bias, or the full universe of information available to members and staff. Congress receives information outside hearings. It also does not mean every witness in the 21-slot non-advocate comparison opposed open models. Some focused on frontier AI, China, biosecurity, or model theft without taking a direct position on open-weight release.

The narrower conclusion is more defensible: the reviewed written hearing record is a thin evidentiary base for evaluating open-weight AI policy. If lawmakers use that record to reason about model release, licensing, procurement, export controls, or safety mandates, the open-model side of the trade-off is underdeveloped.

Why it matters

Witness selection affects which policy options look natural. If Congress mostly hears about open weights in the context of Chinese model diffusion, cyber misuse, CBRN risk, distillation, or model theft, model-release restrictions can begin to look like the natural center of gravity. The reviewed record contains fewer witnesses explaining why open models may also serve transparency, auditability, competition, resilience, and U.S. strategic diffusion.

The advocate testimony supplied that missing rationale in fragments. Cody Venzke emphasized civil liberties and transparency. David Cox, testifying for IBM in two hearings, connected open models to inspection, competition, enterprise trust, and democratic values. David Bray pointed to open-weight systems with open-source code as useful for public-service workflows. Kyle Chan framed American open-source foundation models as a response to Chinese open-weight diffusion. Jack Cable argued that open-weight proliferation is likely and that the United States should support American-made open-weight systems rather than leave the field to foreign alternatives.

That is a materially different policy frame from treating openness principally as a leakage problem. It does not erase the risks. It gives Congress a fuller account of the trade-off.

What Congress should do

Congress should not exclude risk-focused testimony. It should put that testimony in proportion.

When committees convene multi-witness hearings on model release, frontier model safety, Chinese AI diffusion, AI cyber risk, CBRN risk, model licensing, procurement restrictions, or AI export controls, the panel should include at least one witness who can explain the strategic value and operational trade-offs of open-source or open-weight AI.

Congress should also separate issues that are often conflated:

1. Open-source software security.
2. Open-weight model release.
3. Theft or exfiltration of closed model weights.
4. Foreign access to closed frontier models or U.S. compute.

Those issues require different policy tools. Treating them as one generalized "open model risk" category will produce imprecise law.

Finally, Congress should require precision before considering any licensing regime or model-release restriction. A licensing regime that is formally neutral could still close the model ecosystem if it imposes compliance burdens that only the largest closed labs can satisfy. Any review process should have narrow triggers, short review periods, clear standards, and a presumption of approval unless the government can identify a specific risk that justifies delay or denial.

Methodology

This is a compiled witness-level review, not an official congressional statistic. There is no single public URL that reports "114 written witness-statement slots across 30 AI-centered hearings." The figure comes from adding the witness statements listed on the official hearing pages in the source appendix.

This analysis counts written witness statements listed on official congressional hearing pages. It does not count oral speaking time, member statements, questions for the record, support documents, bill markups, hearing transcripts, or committee press releases.

One "witness slot" means one written statement submitted or listed for one witness at one hearing. If the same person testified in two separate hearings, that counts as two witness slots because the position appeared in the written record twice.

The corpus was built from official House docs.house.gov pages and Senate committee hearing pages, then screened by title and subject matter. A hearing was included if the official title or subject matter centered on artificial intelligence and the official page listed witness statements. The first included hearing occurred on July 16, 2025, and the last included hearing occurred on June 25, 2026. The corpus should be read as a best-effort review of AI-centered congressional testimony during the period, not as a claim that no other congressional proceeding mentioned AI.

The review distinguishes model-openness arguments from ordinary open-source software references. Testimony about open-source civic software, open-source intelligence, or open-source development practices was not counted as open-model advocacy unless it made an affirmative argument about open-source AI models, downloadable models, open weights, model release, or preserving open AI.

The classification is conservative:

• "Open-source/open-weight model advocate" means the witness made an affirmative policy, strategic, or governance argument for open-source AI, open models, downloadable models, open weights, or preserving open AI.
• "Model-release/open-weight restriction concern" means the witness identified open release, open weights, or disclosure of model code or weights as a risk requiring restriction, mitigation, or special treatment.
• "Model access/weight-control concern" means the witness focused on securing model access, preventing model-weight theft, countering distillation, or restricting foreign access to frontier models, without necessarily opposing open-weight models.
• "Frontier/catastrophic or advanced-safety risk" means the testimony centered frontier AI capabilities, biosecurity, autonomous cyber capabilities, catastrophic risk, or other advanced AI safety concerns.
• "China/frontier access-control risk" means the testimony centered China, export controls, frontier model access, compute access, model-weight theft, or strategic diffusion.

Radiant Intel verified that all 30 official hearing pages in the source appendix returned HTTP 200 on June 29, 2026. One House witness-statement link listed on docs.house.gov for Dr. Andrew Ibrahim in the September 3, 2025 hearing redirected to a docs.house.gov 404 page at the time of review. That witness was counted as a listed witness-statement slot because the official hearing page listed the witness statement, but the unavailable testimony text was not used for substantive classification.