Anthropic’s Amodei Rejects Open Model Ban But Calls for Testing

(Bloomberg) — Silicon Valley leaders, including Anthropic PBC Chief Executive Officer Dario Amodei and Nvidia Corp. CEO Jensen Huang, are warning against a broad government crackdown on open-weight artificial intelligence systems as a response to a breakthrough Chinese AI model that’s challenged US leadership in the technology.

Most Read from Bloomberg

In a blog post released Monday, Amodei sought to dispel claims that Anthropic supports a ban on open-weight models, which allow users to download and customize the technology. At the same time, he called for the US to adopt policy interventions to slow China’s AI development and said that all models — both open and closed — should go through mandatory safety testing before release.

“Let me state it clearly so that there is no doubt: Anthropic has never advocated for a ban on open-weights models,” Amodei wrote, qualifying that these models pose risks because it is “difficult to apply guardrails” or “monitor their usage.”

As the Trump administration and Congress grapple with how to respond to a new open-weight model from Chinese startup Moonshot AI called Kimi K3 that rivals offerings from Anthropic and OpenAI, tech leaders are planting their policy flags.

Dozens of tech companies — including Anthropic’s main competitor OpenAI — have banded together to urge Washington to shy away from curbs on open-source models. In a missive signed on Friday, tech industry leaders, including Nvidia’s Huang and Microsoft Corp. CEO Satya Nadella, said that open systems help “make advanced AI more accessible, adaptable and widely available.”

Anthropic has not signed that letter, a move that has led some, including former White House AI Czar David Sacks, to say the company is lobbying to hamstring Chinese competitors that challenge the business model of closed-weight AI firms. In Monday’s blog, Amodei denied that allegation and defended Anthropic’s decision not to endorse the coalition’s letter, saying that he disagrees with its assertion that open-weight models make it easier to develop safeguards.

Amodei instead reiterated his long-held stance that open-weight models pose greater risks of enabling cyber or biological attacks. To reduce those risks, he is advocating for strong chip controls, policy action to crack down on alleged misuse of US models by Chinese competitors and mandatory testing of both open and closed models.

Huang, who is leading the pro open-source charge, is scheduled on Tuesday to meet in Washington with Senator Mark Warner, the top Democrat on the Intelligence Committee, according to a spokesperson for the senator. Warner last week called for mandatory safety testing of cutting-edge models as part of a framework for AI legislation, the latest in a series of proposals to regulate AI that have emerged on Capitol Hill this year. Politico reported earlier on Huang’s meetings with lawmakers.

White House Office of Science and Technology Policy Director Michael Kratsios stirred concerns of a US ban on open-weight models last week when he said in a social media post that the US government has evidence that Moonshot improperly accessed US AI models to build its latest model, a process called distillation. That same week, Treasury Secretary Scott Bessent raised the prospect of sanctions against Chinese companies found to be engaging in the practice, which involves using the outputs of a parent model as training data for a student model.

Anthropic and OpenAI have accused Chinese AI companies, including DeepSeek to Alibaba Group Holding Ltd., of distillation. The US government has agreed to share information to help crack down on these so-called attacks, which violate the companies’ terms of service. China has rejected claims that its AI developers have used distillation in an unfair manner, and authorities in Beijing warned Monday that they would meet any US government penalties against Chinese AI companies with their own reprisals.

While some AI developers in the US, like OpenAI, do offer open-weight models, much of their focus has been on closed or proprietary systems that they sell to customers. Chinese companies, including DeepSeek, have previously flooded the market with more affordable, open models that typically lag behind the most cutting-edge software from US rivals but nonetheless offer a compelling product for the price.

The debate in Washington over how to respond to Moonshot is unfolding as the Trump administration prepares to release a proposed framework for reviewing the safety of cutting-edge AI systems. Silicon Valley companies have complained about the ad-hoc nature of recent US moves to rein in new models, including its brief imposition of export controls on Anthropic that led it to temporarily disable its Fable 5 and Mythos 5 products.

One proposal taking shape, which Bessent helped develop, would create an independent regulator that would vet AI models with industry input. The proposal resembles one outlined last week by Google DeepMind CEO Demis Hassabis, who will be meeting with policymakers in Washington this week to lobby for his vision.