- Policy Position: Anthropic rejects a categorical open-weight ban while proposing capability-based safety tests for powerful open and closed models.
- Testing Plan: Demonstrated capabilities would determine which models face pre-release evaluations, but no government has enacted the proposal.
- Targeted Controls: Anthropic also favors tighter China chip controls and action against industrial-scale model distillation.
- Compliance Stakes: Testing costs could favor well-funded labs, while lawmakers still need to assign thresholds, testing, payment, and enforcement.
Dario Amodei, CEO of AI developer Anthropic, signalled opposition to an open-weight ban and proposed capability-based tests for open and closed models.
His clarification came four days after more than 20 companies backed a letter opposing broad restrictions. Amodei reinforced Anthropic’s opposition to a ban, saying: “Anthropic has never advocated for a ban on open-weights models.” Anthropic however stayed outside of the new AI Security Alliance.
No government has enacted the proposal so far, while its cost and enforcement details remain undefined. Open-weight AI models expose learned parameters that developers can download, modify, and run privately; closed model like those used by OpenAI, Anthropic and for Google Gemini are available through its provider’s hosted services. Anthropic’s proposal could affect frontier labs, smaller model builders, researchers, enterprises, and policymakers differently, depending on the capability threshold and who pays to demonstrate compliance.
In Amodei’s assessment, open weights are only a partial factor in cyber and biological misuse risk. Distribution format may also be irrelevant to whether authoritarian governments build superior military AI.
A Threshold Instead of a Ban
Anthropic says downloaded weights create a control problem that hosted access does not. Once copies leave the original developer, it cannot centrally withdraw them or add safeguards to every private deployment. Recent open-weight cybersecurity testing gives that concern practical weight because capable downloadable systems can spread beyond a provider’s monitoring and update channels.
Regulators could leave lower-risk downloadable models available while demanding stronger evidence before more capable systems reach users, although Anthropic has not specified who would set the threshold, conduct the tests, or enforce the results.
Equal capability triggers would place hosted and downloadable systems under the same standard even though their control mechanisms differ. Hosted-model providers can limit access and update a central service, while open-model developers cannot update every private copy. Basing intervention on demonstrated risk could prevent distribution format from deciding the regulatory outcome by itself while retaining extra scrutiny for releases that cannot be recalled.
Anthropic’s policy also targets advanced computing for AI inference that runs existing models or is used for training new ones. The company favors tighter controls on powerful chips and chipmaking equipment sold to China, including action against smuggling and workarounds. It also wants action against industrial-scale model distillation, which uses a stronger model’s outputs to train another model with less computing power.
Anthropic, OpenAI, and Google have coordinated against unauthorized distillation, providing a prior example of the conduct the proposal targets.
The Cost and Competition Test
AI Security Alliance coalition members urge policymakers to avoid broad restrictions and argue that open models “make advanced AI more accessible, adaptable and widely available.” They also favor targeted legal and commercial consequences for unlawful extraction from proprietary models rather than sweeping limits on open-model techniques. The approach concentrates enforcement on conduct, while Anthropic’s proposed capability testing adds obligations before a powerful model reaches users.
Pareekh Jain, CEO of Pareekh Consulting, warned that mandatory evaluations could create an advantage for labs with larger compliance budgets.
“Testing is expensive and time-consuming, and so, giant, well-funded companies like Anthropic, Google and OpenAI can afford it.”
Pareekh Jain, CEO of Pareekh Consulting (via Computerworld)
Testing costs may create an uneven market effect rather than a guaranteed reduction in model choice. Smaller teams may have to fund evaluation work before earning revenue from a release, while established labs can spread that expense across hosted products and enterprise contracts. Higher compliance costs could reduce the supply of advanced open-weight models and narrow access to lower-cost systems, vendor independence, and community-led development.
Researchers and enterprises face the other side of the tradeoff. Comparable pre-release tests could provide stronger evidence about dangerous capabilities before adoption, even if fewer downloadable choices increase dependence on large providers. Regulators would need to decide whether that assurance justifies the access cost and how to prevent the testing regime from favoring companies with larger budgets.
Federal legislation proposals associated with US Senator Mark Warner asks for mandatory safety testing of advanced AI models. Any such legislation would need to identify who sets the capability threshold, who conducts and pays for tests, and which authority enforces the results.
