The EU Unwrites Its AI Act While the US Drafts Pre-Approval: A Regulat – Good Machine

TL;DR

EU ambassadors agreed May 6 to exempt industrial machinery from the EU AI Act, delay high-risk AI restrictions by more than a year, and grant grace periods for AI watermarking compliance.
Germany led the push with France's backing; five EU diplomats confirmed the deal to POLITICO. If finalized, it is the first significant rollback of EU digital rules.
Meanwhile, the Trump administration is considering a 16-page executive order that would create a pre-deployment vetting regime for frontier AI models and prohibit private companies from "interfering" with government AI use.
Microsoft, Google, and xAI have already agreed to give the U. S. government early access to new models for national security testing—spurred partly by concerns over Anthropic's Mythos AI capabilities.
84% of surveyed U. S. business leaders now rank AI policy changes as their top regulatory concern, up from 42% in 2025.

What Happened

Two parallel regulatory earthquakes struck on May 6. One in Brussels. One in Washington. Both concern the same technology. They are moving in opposite directions.

The EU retreat. EU ambassadors voted to support a German proposal that would remove industrial machinery from the scope of the EU AI Act, the bloc's flagship artificial intelligence regulation. The exemption would shift industrial AI compliance to existing machinery-safety rules rather than the AI Act's high-risk framework. The broader review also delays restrictions on high-risk AI applications by more than a year and grants companies additional time to comply with watermarking requirements for AI-generated content. Germany secured France's support before the ambassador-level meeting, according to a senior French official speaking at POLITICO's AI & Tech Week summit. Five EU diplomats confirmed the agreement. (POLITICO)

The U. S. advance. The Trump administration is actively discussing a slate of executive actions on frontier AI models, including a vetting regime that would require government review before deployment. According to seven tech industry representatives and policy advisers granted anonymity to discuss sensitive deliberations, the White House has floated a 16-page executive order that would, among other provisions, prohibit private companies from "interfering" with the government's use of AI models. (POLITICO)

Industry acquiescence. Microsoft, Google, and xAI separately agreed to provide the U. S. government early access to new AI models for national security testing. The Center for AI Standards and Innovation at the Department of Commerce announced the agreement on May 5, stating it would allow evaluation of models before deployment and research into capabilities and security risks. The move fulfills a July 2025 Trump administration pledge to partner with technology companies to vet AI models for "national security risks." (Insurance Journal, Bloomberg Law)

The business mood. A Littler survey of 300 U. S. business leaders found 84% expect AI policy changes to affect their operations—up from 42% in 2025. AI policy concerns have now leapfrogged DEI regulatory worries, which dropped to 38% from 84% over the same period. (Bloomberg Law)

What It Actually Means

The EU is performing a controlled demolition of its own regulatory architecture.

The AI Act took three years to negotiate. It was hailed as the world's most comprehensive AI regulation. Eighteen months after passage, its core mechanisms are being unwound—not by populist revolt, but by industrial lobbying amplified by Franco-German agreement. The machinery exemption matters because Germany is Europe's manufacturing heartland. Siemens, Bosch, and automotive suppliers argued that dual-classifying industrial AI systems under both machinery directives and the AI Act created overlapping, contradictory compliance obligations. Their argument won.

The delay in high-risk restrictions and watermarking grace periods are similarly pragmatic concessions. The EU faced a genuine problem: the AI Act's implementation timeline was outrunning the technical standards needed to comply with it. Rather than pause standard-setting, the Council is pausing enforcement. This is regulatory surrender dressed as regulatory flexibility.

The U. S. is building a pre-approval regime it has never had.

American technology governance has historically been ex-post and responsive. A company ships a product; if it causes harm, regulators act. The proposed executive order would invert this tradition. Frontier AI models—undefined in the leaked drafts, but understood to mean systems above a compute or capability threshold—would require government testing and sign-off before release.

This is not export control (which regulates hardware). It is not sectoral regulation (like FDA drug approval). It is a general-purpose technology review board for software. The constitutional and commercial implications are enormous. The provision barring companies from "interfering" with government AI use suggests the administration wants to prevent vendors from restricting how their models are deployed in federal contexts—effectively overriding standard licensing terms for government customers.

The two moves are not opposites. They are convergent competition policy.

The EU is retreating because strict AI regulation was disadvantaging European industry against American and Chinese competitors. The U. S. is advancing pre-approval because it wants to control the strategic capabilities of the most powerful systems. Both governments have reached the same conclusion: first-mover advantage in frontier AI now outweighs precautionary principle. They differ only in mechanism. The EU is deregulating to compete. The U. S. is nationalizing the vetting layer to secure competitive advantage.

Stakeholder Landscape

Party	Position	What They Gain / Lose
German manufacturing sector	Lobbyist for machinery exemption	Gains streamlined compliance. Loses the AI Act's liability shield—machinery rules offer different, potentially weaker, product-liability frameworks.
French AI ecosystem (Mistral, etc.)	Backed Germany	Gains regulatory breathing room. France's support signals that Paris now views strict AI regulation as a growth inhibitor, not a competitive differentiator.
EU AI Office	Implementer of AI Act	Loses authority and credibility. First major rollback of an EU digital rule on its watch.
U. S. frontier AI labs	Subjects of vetting regime	Lose autonomy over release timelines. Gain potential competitive moat—only well-funded labs can navigate a federal pre-approval process.
Microsoft / Google / xAI	Early-access volunteers	Gain favorable positioning with administration. Risk normalizing a compliance burden that smaller competitors cannot afford.
Anthropic	Implicit target of concern	The government's urgency is partly driven by Anthropic's Mythos AI. Anthropic gains neither from EU rollback nor U. S. vetting—it faces regulatory scrutiny on both continents.
Enterprise AI adopters	End users	EU: Lower compliance costs, shorter timelines. U. S.: Uncertainty—pre-approval could slow availability of frontier tools or create two-tier access (government-approved vs. commercial).
Chinese AI sector	External competitor	Gains relative advantage if Western regulatory fragmentation slows deployment. China's regulatory model is opaque but generally permissive for domestic players.

Cross-Layer Implications

The standards divergence problem. If the EU delays watermarking requirements while the U. S. develops its own provenance standards through the Commerce Department's AI Standards Center, enterprises may face incompatible content-authentication regimes. A watermark that satisfies U. S. federal procurement may not satisfy future EU liability rules. Compliance teams should plan for dual-track metadata strategies.

The compute-regulation feedback loop. The same week that Anthropic secured 300 megawatts of SpaceX compute, the U. S. government signaled it wants pre-approval authority over frontier models. These two stories connect: the administration's concern is not abstract. It is driven by the capabilities that massive compute investments are now enabling. As compute scales, regulatory ambition scales with it.

The IPO pipeline. SpaceX is preparing an IPO. Its AI infrastructure ambitions—Colossus, orbital data centers—are central to that offering. If the U. S. imposes a pre-approval regime on frontier models, the demand for SpaceX-class compute may concentrate among a smaller number of approved labs, potentially improving utilization rates but increasing customer concentration risk.

The transatlantic dialogue. The EU sent a fact-finding delegation to Washington to discuss AI policy synergies. (Washington Post) The timing is awkward: Brussels is unwinding its rules while Washington is tightening theirs. Synergy, in this context, may mean little more than mutual recognition that neither side wants to cede the AI industry to the other.

Recommendations

For compliance and AI governance teams at enterprises:

Do not assume the EU AI Act is stable. The machinery exemption is likely the first of several carve-outs. If your organization built compliance roadmaps around the Act's original scope, reassess them now. Industrial AI systems may soon fall under machinery directives instead—understand the difference in liability, documentation, and CE-marking requirements.
Prepare for U. S. model vetting, even if it does not become law. The executive order is under deliberation, not finalized. But Microsoft's, Google's, and xAI's early-access agreements suggest industry is pre-emptively complying. If you source frontier models through U. S. providers, ask your vendors: will their federal review process affect commercial release schedules? Will there be a lag between government-accessible and commercially available versions?
Separate your watermarking and provenance strategy from any single jurisdiction. EU watermarking rules are delayed. U. S. standards are emerging. Build content-authentication pipelines that can accommodate multiple metadata schemas. Do not hard-code compliance for a regime that may not survive the year.
Audit your AI governance policies against the Littler survey findings. If 84% of U. S. business leaders expect AI policy to affect operations, your organization is probably underprepared. The 68% of employers who have implemented formal AI governance policies (up from 38% in 2025) is a floor, not a ceiling. If you are in the remaining 32%, you are now in the minority. (HR Dive)

For AI labs and vendors:

The pre-approval dynamic favors incumbents. If you are not Microsoft, Google, or xAI, your ability to influence the vetting protocol is limited. Engage with the Center for AI Standards and Innovation now, before the rules crystallize. The window for comment is narrow and closing.

Uncertainty Ledger

Will the EU machinery exemption be finalized? Ambassador agreement is a political signal, not law. The European Parliament must still approve. Opposition from consumer protection groups and some MEPs is likely.
What is the U. S. executive order's timeline? POLITICO's sources suggest deliberations are active but no release date is set. The 16-page draft may be revised substantially before publication.
Which models would the U. S. vetting regime cover? The threshold—compute-based, capability-based, or both—is undefined. This uncertainty itself is a market risk.
Will Anthropic's Mythos AI trigger specific restrictions? The Commerce Department's announcement cited national security testing concerns that appear linked to Anthropic's recently revealed Mythos system. Whether Mythos becomes a named trigger for broader policy is unclear.
How will the EU-U. S. delegation's talks affect convergence? If the EU is deregulating while the U. S. is regulating, "synergy" may mean competitive tension, not harmonization.

Bottom Line

The regulatory equilibrium of 2025—EU precautionary, U. S. permissive—just inverted. The EU is choosing industrial competitiveness over regulatory purity. The U. S. is choosing national security oversight over innovation laissez-faire. Both are rational responses to the same underlying pressure: AI capability is scaling faster than any single jurisdiction's comfort level. For enterprises, the practical effect is a compliance landscape that is simultaneously loosening and tightening, depending on which border you cross. The winners will be the organizations that build regulatory agility into their AI operating models, not the ones that optimized for a single regime.

Sources

POLITICO: "Berlin's big AI win" (EU machinery exemption, five EU diplomats); "White House mulls tight new controls on advanced AI" (U. S. executive order deliberations) (Tier 1)
Washington Post: "AI & Tech Brief: The E. U.'s AI delegation" (Tier 1)
Insurance Journal: "Microsoft, Google and xAI to Give Government Early Access to AI Models" (Tier 2)
Bloomberg Law: "AI Policy Concerns Leapfrog DEI Worries in Survey of Executives" (Tier 2)
HR Dive: "Employers 'still playing catch-up' on AI risk management, Littler report finds" (Tier 2)