OpenAI and Anthropic Align on Open-Weight AI Model Risks

OpenAI and Anthropic, fierce competitors in the AI marketplace, have found common ground in Washington: both are cautioning federal policymakers about the potential dangers of powerful open-weight AI models.

This alignment places the two dominant U.S. AI laboratories in direct opposition to researchers, startups, and open-model advocates who maintain that broad access to AI technology is fundamental to both competition and scientific advancement.

Why it matters

The debate over open versus closed AI models could determine which companies dominate the industry and how quickly AI innovation spreads globally. If regulators side with closed-model labs, it could cement the market position of existing leaders while limiting access for researchers and smaller competitors.

The regulatory capture argument

Critics, including Trump administration adviser David Sacks, have raised concerns about regulatory capture—the possibility that safety-focused rules could inadvertently protect large incumbents by making it harder for competitors to release models. Closed-model labs would directly benefit from increased scrutiny of open-weight alternatives.

Anthropic CEO Dario Amodei has argued that open-weight models present unique safety challenges because developers cannot revoke access, update safety features, or prevent misuse once model weights are publicly released. Open-model supporters counter that this permanence is precisely the point: no single company should control who uses AI technology or how.

An OpenAI spokesperson stated the company views open source as part of "democratizing access" but within a national framework, though specifics of that framework remain undefined.

The China distillation debate

Both OpenAI and Trump administration officials have raised alarms about distillation—a process where one model is trained using another company's model. U.S. Trade Representative Jamieson Greer characterized Chinese distillation as intellectual property theft, while Treasury Secretary Scott Bessent announced plans to examine Chinese models against U.S. standards in coming weeks.

If substantiated, these allegations would mean Chinese developers used American models to build comparable systems at reduced cost, then released subsidized versions into Western markets. However, no detailed safety standards for U.S. models currently exist, and the administration is still determining what those standards should be.

Open-model advocates warn that treating distillation as IP theft could effectively restrict independently developed open-weight models. An OpenAI spokesperson maintained that "advances in Chinese open-weight models are not an argument against openness" but rather demonstrate the need for a framework enabling quick model evaluation and risk management.

Research implications

Suresh Venkatasubramanian, former Biden White House tech advisor, expressed concern about potential access restrictions to Chinese models, calling it a "big problem for doing research." He emphasized that open-weight models are especially valuable to researchers because they can be modified and studied more freely than closed alternatives.

State-level divergence

While aligned federally, OpenAI and Anthropic have diverged at the state level. OpenAI endorsed a Massachusetts frontier safety bill that differs significantly from the version Anthropic supported. Anthropic's preferred version required independent model testing every six months for catastrophic risks, public reporting, and empowered the state attorney general to sue for non-compliance. OpenAI's endorsed version lacks these provisions.

OpenAI CEO Sam Altman is scheduled to meet with lawmakers and the White House next week to discuss upcoming model releases.

These details were first reported by Axios.