Two Blueprints for AI, Built on Different Questions

In late July, the European Commission issued transparency guidelines requiring every chatbot operating in the bloc to identify itself as artificial and every piece of AI-generated content to carry a machine-readable label. The purpose, the Commission said, was to preserve citizens' ability to trust what they see, hear, and read against disinformation produced at unprecedented scale. ^[1][2] Two months earlier, in May, President Trump canceled a planned executive order that would have required AI developers to share frontier models with federal agencies 90 days before public release. The framework, pushed by Treasury Secretary Scott Bessent to protect financial institutions from AI-driven cyber risks, was killed after tech executives including Elon Musk and Mark Zuckerberg lobbied against it. ^[3] Trump made the reasoning plain.

Because I didn't like certain aspects of it, I postponed it... We're leading China, we're leading everybody, and I don't want to do anything that's going to get in the way of that lead. — Donald Trump

These are not different answers to the same question. They are answers to different questions. The EU is regulating the relationship between AI and the citizen — what a person sees, hears, and believes. Washington is regulating the relationship between AI and the adversary — what a rival state can extract, deploy, or weaponize. The two governments have built regulatory perimeters that share a subject line and almost nothing else. The European perimeter runs between the output and the person who encounters it. The AI Act, whose enforcement began this week, classifies AI systems by the risk they pose to individuals and applies penalties of up to €35 million or 7% of global turnover — including to non-EU companies whose systems are used in the bloc. ^[4][5] The transparency rules effective August 2 require chatbots to self-identify, mandate machine-readable markings on AI-generated media, and compel disclosure of deepfakes and biometric categorization — all aimed at reducing the risk of deception and manipulation. ^[1] The logic extends beyond labeling. In March, the European Parliament banned sexualized AI deepfakes, and an Amsterdam court ordered xAI's Grok to stop generating non-consensual sexual imagery under threat of €100,000 daily fines — placing the burden on the company, not the user. ^[6] Under the Digital Markets Act, the EU ordered Google to open Android to rival AI assistants like ChatGPT and Claude and to share search data with competitors, with fines of up to 10% of global turnover — all aimed at market fairness and consumer choice. ^[7][8][9] The EU's Tech Sovereignty Package, unveiled in June, restricts foreign cloud providers from defense, healthcare, banking, and energy sectors — but the response is regulatory de-risking and infrastructure-building, not classifying models as state secrets. ^[10] The perimeter is consistent: the EU regulates what AI does to people. The American perimeter runs in a different direction entirely — between the capability and the foreign adversary who might capture it. The administration's concern is not what AI does to citizens but what rivals do with American AI. In February, the Pentagon designated Anthropic a national security supply chain risk — a label normally reserved for foreign adversaries — after CEO Dario Amodei refused to remove safety guardrails that blocked Claude from use in mass domestic surveillance and fully autonomous weapons. ^[11] Defense Secretary Pete Hegseth made the administration's position explicit.

America’s warfighters will never be held hostage by the ideological whims of Big Tech. — Pete Hegseth

Trump ordered all federal agencies to cease using Anthropic's technology and shifted the Pentagon's classified AI needs to OpenAI and xAI. But reports indicate the US military continued using Claude for intelligence and targeting during operations in Iran and the capture of Maduro in Venezuela — meaning the government that blacklisted a company for refusing to strip safety guardrails then used its technology anyway when the mission required it. ^[11] The pattern holds across the administration's other AI interventions. Chinese military researchers are using model distillation to extract capabilities from US frontier models — GPT-3.5 for military source code processing, Claude 3 Haiku for social media monitoring and drone navigation — bypassing chip export controls entirely. ^[12] Anthropic itself warned that distilled models may lose original safety safeguards, potentially allowing sensitive capabilities to transfer to models beyond US control. ^[12] The administration's response is to consider banning domestic companies from using Chinese open-source models like Moonshot's Kimi and Alibaba's Qwen — a user-side control, but aimed at protecting strategic advantage rather than protecting users from deception. ^[13][14] The US-China Economic and Security Review Commission warned that 80% of US AI startups now use Chinese open-source base models, and that US export controls target only chips while China advances through deployment-driven data — meaning the regulatory frame is structurally misaligned with the actual mechanism of competitive leakage. ^[15] When Demis Hassabis proposed a US-led AI watchdog focused on cybersecurity, biological, and nuclear risks — strategic threats, not consumer-facing harms — White House AI adviser Sriram Krishnan dismissed it outright.

there will not be an FDA for AI. — Sriram Krishnan

Critics warned that even an industry-funded watchdog could be perceived as an instrument of US strategy. ^[16] Even the industry's own regulatory proposals in the US are framed around strategic competition, not user protection. On the consumer side, the federal government has not merely declined to act — it has actively tried to prevent states from acting. The administration launched a DOJ task force to preempt state AI laws in California and Texas, and xAI sued Colorado on First Amendment grounds, arguing that state AI discrimination laws would hamper innovation and undermine US national security. ^[17][18] But the attempt to preempt state protections through a 10-year ban collapsed after a 99-1 Senate vote rejected it, leaving the US with no federal consumer AI safety framework and no effective federal preemption of the state laws the administration wants gone. ^[19] Public Citizen's J.B. Branch argued that states must now serve as first responders to AI harms — fraud against the elderly, nonconsensual imagery of children, deceptive political content — because the federal government has abdicated the role. ^[20] The two perimeters are now in direct collision. The EU AI Act applies extraterritorially — any non-EU company whose systems are marketed in the bloc or whose outputs are used there must comply. ^[4] The Trump administration treats the AI Act, the DMA, and the DSA as trade barriers to be negotiated away, demanding that the EU make concessions on digital technology regulations as part of any trade deal. ^[21] The EU asserts regulatory autonomy as a red line. The June trade deal settled goods tariffs but left the digital regulation question entirely unresolved — the deal contains a sunset clause and suspension mechanism but no answer to the fundamental question of whether the EU can regulate American AI companies. ^[22] The irony is structural, not rhetorical. The EU, long mocked for overregulation, has built a functioning enforcement machine with real penalties and real extraterritorial reach — fines measured in percentages of global turnover, daily penalties for noncompliance, courts ordering companies to stop generating harmful content. The US, which frames every rule as an obstacle to winning, has neither a federal consumer-protection floor nor effective strategic oversight. It has ad hoc national security authority applied when the military needs something — blacklisting a company for refusing to strip guardrails, then using its technology anyway — and a 99-1 Senate vote that means it cannot even preempt the state laws it wants gone. A government that regulates the secret only to weaponize it, not to secure it.