For the past 4 years, public policy toward “AI” has been driven by two competing narratives. In the past two weeks, the clash between them has blown up in the faces of policy makers and AI service providers, especially Anthropic and Open AI. 

Narrative 1 

One camp told us that AI “models” might evolve into an autonomous, omnipotent life form and destroy us. Continued progress in the development of digital information systems, they claimed, might produce a superintelligence that was outside our control. Even those who didn’t anticipate human extinction spoke darkly of AI’s potential to create “catastrophic risks.” This threat became the groundwork for global society’s overwhelming focus on ”safety,” “guardrails,” and controls over the release of new AI models. 

Narrative 2 

The other camp told us that we were in a race to develop AI. AI will drive both economic growth and national power; thus, we should encourage and accelerate its development. The government must ensure US leadership in AI with subsidies here, protections and tariffs there, and a systematic attempt to block or handicap China’s digital capabilities. The contestants in the race, in other words, are not private firms competing for markets and profits, but nation-state-led systems of political economy. China is our opponent – our “adversary” — and must therefore be “beaten” in this “race.” 

It should be apparent that the policy implications of Narrative 1 and Narrative 2 are in direct conflict. Narrative 1 wants us to slow down development, if not abandon it altogether, and engage in global collective action to maintain control. Narrative 2 tells us to accelerate development unilaterally as a nation and dismisses ex-ante safety regulations as roadblocks. If AI really generates a race with a military and political adversary, our national security depends on running as fast and hard as possible.  

The weird thing about these two incompatible narratives is that many of the big players in the AI governance drama want to believe both at the same time.  

Converging Narratives 

What happened in the past two months (June – July 2026) was a remarkable collision between the two narratives. The Trump administration now wobbles between the two, unpredictably and incoherently. 

It started with the June 2 Executive Order (#14409) (no doubt inspired by Mythos’s ability to find software vulnerabilities). EO 14409 is a schizophrenic document. Section 1 speaks of the “enormous talent and innovation of our AI industry” and attributes that innovation to our refusal “to stifle this innovation with overly burdensome regulation.” Section 3, on the other hand, gives the Director of NSA and other representatives of the Department of War, the National Cyber Director, and the Director of the Cybersecurity and Infrastructure Security Agency the authority to review AI models before they are released. It asks the companies to “provide the Federal Government with access to covered frontier models … for a period of up to 30 days before they plan to release such models to other trusted partners.” The EO claimed that “Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models, including frontier models.” 

A few days later, on June 12, the U.S. Commerce Department did exactly what the EO promised not to do. It applied mandatory export controls to Anthropic’s Claude Fable 5 and Claude Mythos 5 models to restrict access to foreign nationals, whether inside or outside the United States. Anthropic responded by suspending access to everyone. “Because the order took effect immediately and we had no reliable way to verify nationality in real-time,” the company explained, “we suspended access to both models for all users.”  

Although the ban was lifted June 30, the incident triggered a furious debate over review and licensing of AI models, and AI governance more broadly.  

• “Voluntary” an Unstable Compromise:  Critics argued that the program was voluntary in name only. As Anthropic’s experience with the Department of War showed, the government could retaliate against firms that did not cooperate. None of the big 3 AI firms could afford to ignore it, and enterprise clients might refuse to procure any model that had not cleared federal review. 

• Lack of Transparency and Due Process: Critics pointed out that the review mechanisms were not based on any written standards or well-defined criteria. The benchmarking process called for by the Executive Order was to be “classified,” opaque, and lacking any appeal process. 

• The China Competition. While this policy turmoil was going on, China’s method of competing by promulgating open source models began to bear fruit. Zhipu AI launched GLM 5.2, a 744-billion-parameter model briefly named the top-performing open-weight model on the independent Artificial Analysis Intelligence Index. Its inexpensive pricing led to enterprise adoption. On July 16, China’s Moonshot AI released Kimi K3, the world’s largest open-weight model.  And on July 26, DeepSeek V4 replaced older APIs, offering a massive 1-million-token context window. 

Anthropic and OpenAI executives raised alarms in Washington – privately. They warned that Chinese firms were closing the gap by illicitly “distilling” advanced U.S. models to train their own systems. Amidst talk of banning open source AI models, on July 24, 2026, Nvidia CEO Jensen Huang published an industry open letter titled “Open Weights and American AI Leadership”. The letter urged Washington not to restrict or ban open-weight models, arguing that an open ecosystem is vital for American tech leadership. The letter was signed by over 100 companies, including Cisco, Google, IBM, Microsoft, and Red Hat.  

Anthropic was conspicuously absent from the signatories. 

Global “Pacing”?

But Anthropic did sign another letter that came out two days later. Pacing the Frontier a letter signed by 1,324 employees of AI companies, revived 2023-vintage fears of an out-of-control AGI. It claimed that “there is a real risk that [AI] capability development rapidly accelerates beyond our ability to understand or control the resulting systems.” The letter asked policy makers worldwide to “build on work already underway to monitor frontier model releases” (i.e., Executive Order 14409), by requesting: 

“that the U.S. government support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development.” 

Note that the letter realistically recognizes that competition in the market discourages frontier firms from slowing down development, and that any attempt to control or limit model releases would require globalized collective action, not national competition. This would require cooperation with China, the same country that Anthropic believes should be denied advanced semiconductors and is stealing their “intellectual property” via distillation. An international AI governance effort that does not include China? It’s hard to imagine a more incoherent policy stance. 

So where does that lead us? The safety paradigm asks us to slow down and control development via globalized governance. The race narrative urges American companies to stay ahead but also fosters a connection between national security and digital technology development that encourages the government to exclude foreigners from the market and even to review and license American AI development. So which is it – review and license or let her rip? All these issues came to a head in June-July 2026, when the leading AI firms and the U.S. government tried to square a circle by taking both sides. 

Narrative 3 

There is  a less ideological narrative about AI going on. Narrative 3 might be characterized as letting global digitization proceed under market constraints. This third narrative sees AI not as a special, isolated capability but as one of many expressions of a global digital ecosystem. It pays attention to news reports about the real-world digitization of the economy, finding some of them successful and others not. It consists of spotty and unsystematic reports about the actual application of AI to specific parts of the economy, as well as reports about how unrealistic predictions (such as massive job losses) are being debunked. It chronicles the capital expenditures and revenue generation of American and Chinese AI firms and the growing presence of Chinese firms as a competitive factor in the development of the industry, noting how AI capabilities are subject to market constraints just like any other industry.  

Narrative 3 is neither polarized nor polarizing. It focuses on real-world developments, real costs and benefits, not hopes and fears. One of its most interesting features is its recognition that, despite all the decoupling pressures, China and the US are, economically and technologically, interconnected parts of the digital economy. People who pay close attention to how AI is actually being used realize that both Chinese and American technologists and markets are driving progress in digital information systems and that users are picking and choosing from both ecosystems. 

Semiconductors are the third most traded item in the global economy, behind oil and autos. The supply chain is regionally centered in East Asia, and includes South Korea, Japan, Taiwan and China. Data and cloud services, too, are largely borderless. There is not really a national competition, in which the benefits and costs follow national borders, but a globalized digital information infrastructure, and a globalized market for the production and use of digital information systems. Competition, specialization and globalized market access, on net, show signs of benefiting all world societies. AI governance is, after all, just like Internet governance. And that’s because the Internet and AI are both part of the same interconnected, transborder digital information system. 

The post Colliding AI governance Narratives appeared first on Internet Governance Project.