Who Gets to Decide? The AI Governance Fracture Nobody Has Answered

The global AI governance debate shattered along multiple fault lines this week. The European Commission invited AI labs to Brussels for slowdown talks; Germany — a member of that same EU — rejected an AI pause as not viable and called for U.S.-China oversight frameworks instead. China's government called for UN-led AI governance while China's spy chief simultaneously named specific American AI models as security threats. And Anthropic's CEO published a slowdown essay the same week his company reported growing profitability — a tension that was not lost on observers.

For those following the regulatory debate, a foundational legal point deserves attention: the bedrock of U.S. antitrust law, the Sherman Antitrust Act of 1890, prohibits not market dominance itself but the use of monopoly power to harm competition through exclusionary practices, predatory pricing, or acquisitions designed to neutralize rivals. A company can hold 70 percent of a market and be entirely legal if it arrived there through superior products and efficiency rather than conduct that shuts out rivals. In the AI context, scrutiny would become relevant if a major lab's dominance of foundation model compute or talent were used to prevent meaningful competition from emerging — not simply because three companies hold most frontier model capability.

That distinction matters because several congressional proposals treat market concentration in AI as inherently problematic under a standard that would require a significant departure from existing law. The kill switch bill that Rand Paul blocked was not framed in antitrust terms, but its operational effect — requiring government approval before deploying certain AI capabilities — would function as a regulatory barrier benefiting incumbents with already-deployed systems over new entrants, which is itself a competition argument.

A harder question lurks beneath this week's policy panic: whether the evidence underpinning the AI safety consensus is as robust as its policy consequences require. The threat reports driving Amodei's slowdown essay and von der Leyen's Brussels talks are authored by the same companies that would benefit from regulatory frameworks difficult for smaller competitors to satisfy. The Iran Navy targeting story appears credible. But the specific claim that current AI systems meaningfully lower the barrier for bioweapons synthesis beyond what published scientific literature already provides is an empirical assertion that has not been independently verified at the level of rigor that a consequential slowdown policy would demand. Germany's argument — that caution has its own costs if it narrows the capability gap with China — captures exactly that tension, and no governance framework proposed this week has resolved it.