Anthropic、オープンウェイト禁止に反対し安全テストを提案

Anthropic does not want open-weight AI models banned. CEO Dario Amodei said so plainly today after a week of criticism. Instead, he proposes mandatory safety testing for every sufficiently capable model before release. The catch is that nobody has defined where “sufficiently capable” begins, and that determines whether tomorrow’s next open-weight models stay available.

“Anthropic has never advocated for a ban on open-weight models,” Amodei writes in this letter, calling models without dangerous capabilities “a public good” that cost nothing beyond the compute to run them.

His proposal has three parts: tighten export controls on advanced chips and chipmaking equipment, crack down on industrial-scale model distillation, and require safety evaluations for every sufficiently capable model before release.

The first two are aimed at Washington. The third would affect developers preparing models for release.

Amodei calls the testing idea close to consensus, noting the Trump administration has moved that way and that recent industry proposals would apply testing regardless of a model’s origin or license, “while exempting less capable models, such as those from startups and academia, entirely.” The exemption is critical, though; again, the term startup can be hard to pinpoint. Once some models are exempt and others are not, the policy becomes a capability threshold by another name.

So where does the line fall? Nobody has said.

The letter does not name who would set the number or who would run the tests. Model capability could be measured by metrics like benchmark results, training compute, or scores evaluated against defined risk categories. And the chosen metric would dictate the model type. Whoever picks the metric is also setting release timing for every lab that publishes weights.

This is a fast-moving story. The “Open Weights and American AI Leadership” letter so far counts 50 signatories, with OpenAI and Google joining Microsoft, Nvidia, Meta, AMD, Cisco, Cloudflare, GitHub and Ollama. Anthropic and Amazon remain notable holdouts.

On Monday, Nvidia launched the Open Secure AI Alliance with more than 35 founding members, including Microsoft, IBM, Hugging Face, Palantir, Red Hat, CrowdStrike, and the Linux Foundation. Anthropic is not among them. Nvidia’s argument is that defenders need open frontier models to investigate incidents, citing the recent case in which closed tooling blocked forensic work on a Hugging Face incident and Hugging Face ran the analysis on an open-weight model instead.

To his credit, Amodei addressed the obvious criticism directly. A U.S. ban, he said, would mostly shield domestic AI companies from competition while doing little to stop bad actors. “It would protect U.S. AI companies from competition, but that has never been my goal.” That’s fair. It’s also true that Anthropic does not release open-weight models, so the statement feels a touch empty.

Amodei concedes the testing “would need to be global, which means even the CCP would need to be on board,” and figures that may be possible. This could be the largest unanswered question around this story. Export controls and anti-distillation measures are policies Washington can impose on its own. A global capability-testing regime is different. The proposal that most directly affects open-weight models depends on broad international agreements that do not yet exist (and may never exist).

The post Anthropic wants tests, not bans, as OpenAI and Google back open weights appeared first on The New Stack.