On October 6, 2026, U.S. Democratic Senator Adam Schiff of California said in an interview with The Verge that as AI models continuously iterate on themselves through recursive evolution, "it becomes harder and harder for us to see what they're doing, how they're making decisions." He called for a dedicated AI regulator with real authority and professional expertise—"an FDA for AI, or something like it"—capable of implementing oversight far faster than Congress can legislate.
This is one of the most concrete public statements by a federal-level lawmaker on the structure of AI regulation. Schiff also announced that he and Republican Senator John Curtis are jointly advancing the CLEAR Act (Copyright Labeling and Ethical AI Reporting Act), which would require AI companies to submit declarations to the Copyright Office detailing their use of copyrighted content in training data before releasing new models, and to build a publicly searchable database; already-released models would have to file retroactively or face civil penalties.
Why Recursive AI Breaks the Logic of Regulation
Schiff used the concept of "recursive AI" to describe the threat path he worries about most. Recursive evolution refers to models using their own outputs to train or optimize the next generation of models. When a model's weights, behavior, and decision boundaries are no longer primarily controlled by human annotators and engineers, but instead emerge through multiple generations of iteration, external observers will struggle to determine a model's actual behavioral boundaries through system cards, technical reports, and benchmarks.
Schiff put it bluntly: "We can't even control these models themselves, let alone provide them to the public, to good actors and bad actors, and expect nothing terrible to happen." When even developers struggle to predict the behavioral boundaries of recursively evolving models, administrative regulatory frameworks built on product documentation face the risk of fundamental failure.
This is precisely why Schiff invokes the FDA as a reference point. The logic of drug regulation is not to have Congress review every molecular structure one by one, but to establish an agency with professional staff, testing laboratories, and mandatory pre-market approval authority, so that new drugs must undergo independent verification before entering the market. Schiff wants a similar mechanism for AI.
The CLEAR Act: Copyright Disclosure as Regulatory Infrastructure
Schiff chose to build the information infrastructure first. What the CLEAR Act requires is not compensation but records: companies must declare which copyrighted content they used before releasing a model, and those records would enter a public database, providing a basis for future congressional judgments about "which training uses constitute fair use."
This approach sidesteps the most difficult dispute of the moment while creating conditions for future legislation. Schiff directly criticized the industry's historical practice—"they decided from the outset to ignore intellectual property law and simply take this content"—but chose to make the data visible first and discuss liability later.
For large model developers, this means training data records would shift from an internal risk management matter to a statutory compliance obligation. Companies that previously avoided copyright discussions on the grounds that "the scope of training data is hard to trace" will face new evidentiary pressure.
Where the Stakeholders Stand
For leading AI labs, the most direct impact of the "FDA for AI" proposal is the release cadence. The FDA's approval cycle for new drugs typically runs more than a year. If an AI regulator obtained similar pre-market approval authority, the release pace of large models would be forced to slow—a structural shock to an industry ecosystem whose mainstream strategy is currently "iterate fast, learn after release."
For developers and API users, the impact is felt more at the level of compliance documentation. If regulators require declarations of usage norms akin to drug labels, API callers may need to demonstrate that their application scenarios match the declared uses. This is especially sensitive in high-risk industries such as finance, healthcare, and law.
For the open-source model community, the situation is more complicated. The FDA model presupposes that approval can happen before market release, but once open-source weights are published they cannot be recalled, and a regulator's jurisdiction over open-source models has a fundamental gap in the legal framework.
Copyright holders—publishers, music companies, visual artists—are the direct beneficiaries of the CLEAR Act. Once a mandatory disclosure database is established, it will provide a reliable chain of evidence for copyright litigation.
Legal Headwinds from the Supreme Court Ruling
Schiff acknowledged a key legal obstacle: the Supreme Court's Loper Bright decision means courts will no longer automatically defer to administrative agencies' interpretations of ambiguous statutes. This poses a substantive threat to the "FDA for AI" proposal. Even if Congress authorizes the creation of a dedicated AI regulator, the specific rules that agency issues are highly likely to face a barrage of litigation.
Schiff's response is that Congress must "write it as specifically as possible when legislating," reducing the room for interpretive ambiguity. This means that for an "FDA for AI" to stand in the post-Loper Bright era, its authorizing statute would itself need to resolve a large number of technical details at the level of its provisions.
A Historical Precedent: The Costly Delay in Social Media Regulation
Schiff points to the failure of social media regulation as a negative reference: "We absolutely cannot take that many years again to even begin to figure out social media." Facebook was founded in 2004, and comprehensive federal legislation addressing platform liability, privacy, and algorithmic transparency has yet to pass.
He rejects replacing mandatory rules with industry self-regulation agreements. This position stands in contrast to the historical trajectory of social media governance: platforms repeatedly traded voluntary commitments for periods of legislative leniency, and ultimately failed to meet public expectations on issues such as content moderation and data privacy.
What Is Most Likely to Happen Next
The pace of the CLEAR Act's progress in the Senate is the most observable near-term indicator. The bill already has bipartisan co-sponsors and a certain base of political consensus on copyright issues; how quickly it moves will determine whether training data disclosure becomes the first federal AI compliance obligation to take effect.
The broader "FDA for AI" legislation faces a much longer timeline. Creating a new independent federal agency in the United States requires legislative authorization, congressional appropriations, and Senate confirmation of appointments. While Schiff's remarks have drawn a favorable response in policy circles, the specific agency's powers and jurisdiction have yet to be announced.
For large model vendors, the most pragmatic course of action right now is to proactively build training data provenance records. That capability will become a core competitive asset in both copyright litigation and future regulatory scenarios.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接