U.S. and China Simultaneously Tighten Regulation on Frontier AI Models
N.R. Finch
What happened in the US? What did OpenAI and Anthropic run into?
Before releasing GPT-5.6, OpenAI negotiated with US government officials for several weeks, focusing on the model's cybersecurity capabilities and safeguards. OpenAI made adjustments, but CEO Sam Altman did not disclose what changed.
Anthropic faced a blunter intervention: the Commerce Department ordered it to take down Fable 5 and Mythos 5, citing rules barring foreign persons — including some Anthropic employees — from access. Restrictions eased after Anthropic added cybersecurity safeguards.
This means → Top US AI companies now effectively need government approval before shipping their strongest models, even though no formal licensing regime exists.
The "voluntary framework" sounds gentle — what is it really?
On June 2, the Trump administration signed an executive order: frontier AI developers may give the government up to 30 days of security access before releasing a model and cooperate with officials in selecting early users. The order explicitly rules out mandatory licensing.
But Brad Carson, head of the Responsible Innovation organization, called the emerging arrangement a "de facto licensing machine" — participation is nominally optional, but skipping it makes a top-tier launch far harder.
In plain terms = The government never said "you must submit for review." But if you don't, whether your model ships becomes an open question. Voluntary has become semi-mandatory.
What risks does this create for the AI industry?
There is currently no published, quantifiable safety threshold — neither companies nor regulators have an agreed standard for what counts as "safe enough" or when intervention is warranted. Anthropic explicitly called for a universal framework to assess jailbreak severity.
This means → Regulation could devolve into case-by-case negotiation: each company, each model, bargaining with government individually — outcomes shaped by political pressure, not technical criteria.
The direct fallout: delayed product launches, higher compliance costs, and the smallest developers hurt most — they lack OpenAI's negotiating leverage.
What is China doing on its side?
China had previously treated the global spread of homegrown open-source models — DeepSeek, Moonshot AI and others — as technological soft power, actively encouraging distribution. These models have already built large user bases in Silicon Valley.
But according to the Wall Street Journal, Chinese regulators have begun discussing new measures with leading AI labs: stricter pre-release reviews, delayed public releases, tiered access restrictions for foreign users, tighter AI-technology export controls, and caps on foreign investment in specific developers.
This reflects a striking reversal: it was precisely the overseas success of Chinese open-source models that prompted Beijing to ask — are some capabilities too strategically valuable to give away freely?
The US and China are taking opposite paths — do they arrive at the same place?
The US direction: pull the "release gate" into government hands — models must clear a government checkpoint before launch.
China's direction: claw back the "export gate" — models that were freely open begin facing questions about who gets access, and how much.
In plain terms = One is "show your ID before you leave the building"; the other is "the building that never checked IDs is now installing a scanner." Opposite directions, same conclusion: the freedom to release frontier AI models is narrowing.
Until unified rules emerge, regulatory uncertainty itself is a cost the industry must bear — whether you sit in Silicon Valley or Beijing.
Content is for reference only, not financial advice.