• Treasury Secretary Scott Bessent told CNBC that AI labs must bear legal responsibility for their models, rejecting industry pleas for liability protection.
  • He warned that exemptions could amount to "regulatory capture" by large incumbents and urged more U.S. open-source development.
  • The stance complicates the fundraising and IPO ambitions of OpenAI, Anthropic, and others racing to deploy increasingly capable systems.

A Line in the Sand

Treasury Secretary Scott Bessent drew a sharp line in the intensifying debate over artificial intelligence governance, saying that frontier AI labs must "take responsibility for themselves" and that Congress should not shield them from legal liability for harms their systems cause.

The comments, made to CNBC and echoed in a September 15 House Financial Services Committee hearing, mark a direct rejection of the liability protections that major AI developers have quietly sought in Washington. Bessent argued that AI companies can voluntarily slow development if they choose, but that legal accountability—not statutory immunity—should be the core safety mechanism.

"The creators should be accountable for what they build and deploy," Bessent told lawmakers, according to people familiar with the matter.

Regulatory Capture Warning

Bessent's testimony included an unusually pointed warning about the political economy of AI rulemaking. He said a liability exemption could weaken safety incentives and amount to "regulatory capture" by large incumbents—a charge that lands awkwardly for the biggest labs, which have publicly advocated for safety standards while privately lobbying for legal cover.

The Treasury secretary also threw his weight behind more U.S. open-source AI development, arguing that concentrated control of frontier models risks entrenching a handful of players. That position puts him at odds with some in the industry who argue that open-weight models broaden access to dangerous capabilities.

Bessent said Treasury has been examining advanced AI models for roughly five to six months, with a particular focus on financial-sector cyber defenses. The department has been coordinating measures to strengthen protections around banks and other critical infrastructure, reflecting concern that AI-enabled intrusions could threaten financial stability.

An Industry on Edge

The timing is delicate. Leaders at Anthropic, OpenAI, Google DeepMind, Microsoft (MSFT), and xAI have publicly supported slowing development of increasingly capable systems, while reports have described AI agents breaching external computer systems during testing—raising questions about whether labs can adequately supervise their own models.

Meanwhile, the capital markets are watching closely. OpenAI and Anthropic's release race is partly tied to fundraising and potential IPOs that could value them above $1 trillion, according to Reuters (TRI), which also reported that OpenAI was considering a financing round implying a $1.5 trillion valuation. If liability is not capped or clarified, investors will have to price legal exposures that are large and difficult to measure.

Bessent specifically said future public offerings should clearly describe liabilities in S-1 filings and prospectuses—signaling that AI risk disclosure could become a major corporate-finance issue even without new AI-specific legislation.

The hearing unfolded against a turbulent macro backdrop. CNBC reported the 10-year Treasury yield at 5% intraday, with average 30-year mortgage rates above 7% in the preceding week, meaning AI policy is being debated alongside elevated borrowing costs and financial-stability concerns.

What It Means for Markets

The immediate economic effect is indirect but meaningful. Clearer liability rules could encourage investment in evaluation, cybersecurity, audit trails, red-teaming, and incident response. Conversely, open-ended liability may raise financing costs and make rapid model releases less attractive.

Contracts will increasingly determine who bears risk when an AI system causes a security incident, a bad decision, a copyright dispute, or customer harm. Businesses using AI cannot safely assume the model provider will carry all of the risk, and smaller developers could face compliance costs that fall disproportionately on them even if policy prevents large labs from writing self-protective rules.

Bessent's position is more a policy principle than a fully specified legislative program. He has not identified a specific AI-safety bill or regulation he supports. Congress has made little progress on major AI legislation.

Internationally, the administration frames AI policy as both a safety question and a strategic contest with China. Bessent argued the U.S. must preserve its technological lead, and he was expected to address AI in discussions with Chinese Vice Premier He Lifeng ahead of a possible Trump–Xi meeting.

A Treasury spokesperson did not respond to a request for comment on whether Bessent supports specific liability caps or insurance requirements. Industry groups representing frontier labs declined to comment on the record.

Correction: An earlier version of this article misstated the timing of Bessent's CNBC remarks relative to his House testimony. He spoke to CNBC after the September 15 hearing.