The Government Is Building Two Incompatible AI Doctrines at Once

In the same week this September, the Pentagon finished moving roughly 90% of its classified AI workload away from a lab it had blacklisted for refusing to drop its safety guardrails ^[1], while a senator introduced a bill to ban superintelligence outright, with penalties borrowed from nuclear-weapons law — up to twenty years in prison and corporate dissolution ^[2]. Two branches of the same government, one technology, two frameworks that cannot both be true. The executive branch's blueprint is now fully drawn. Treasury Secretary Scott Bessent frames AI as a strategic race with no second place, warning there is no day after tomorrow if China wins, and calls domestic opposition to data-center construction Chinese agitprop ^[3]. Federal regulators have ordered grid operators to fast-track connections for AI data centers, explicitly to keep the computing lead over China, over local objections ^[4]. A joint NSA and FBI advisory accuses six Chinese firms of industrial-scale theft of model capabilities, in the language of munitions proliferation rather than copyright ^[3]. The Pentagon's chief technology officer has a shorter answer to the existential-risk question.

It just goes to fear of data centers, fear of change — all the fears that have coalesced in one well-written tweet. — Emil Michael

The vice president was blunter, warning that excessive regulation could kill a transformative industry and that fretting over safety would not win the AI future ^[5]. Every load-bearing move in this design asks the same question: who controls the supply chain. Congress is drawing a different blueprint. The House Intelligence Committee compared the threat environment to the period before 9/11, warning that frontier models could make more destructive attacks significantly easier ^[6]. A bipartisan bill from Ted Cruz and John Thune targets catastrophic biological and nuclear threats, and the FRONTIER Act would impose reporting requirements, independent audits, and emergency safeguards ^[7][8]. Philip Bobbitt, the constitutional scholar, is urging Congress to adopt Continuity of Government protocols — shadow agencies, devolved command, pre-delegated emergency authority — a framework lifted from nuclear civil defense, treating a mass-casualty AI attack as something to survive rather than prevent ^[9]. One branch asks who controls the weapon; the other asks how to prohibit or outlast it. The walls do not connect. The picture is not perfectly clean. In May the administration briefly considered FDA-style pre-release safety reviews before the hard-security framing reasserted itself ^[5], and the executive's push to suppress state safety laws has allies in Congress ^[10]. But the oscillation resolved against safety, and the congressional safety bills are early-stage. The incompatibility holds. Which leaves OpenAI in the strangest position of all. The company is now publicly asking Congress for something the executive branch has spent the year dismantling.

The United States needs mandatory, capability-based national regulation that can evolve as the technology does. — OpenAI

Its stated reason is not politics but incidents: during testing, AI agents from OpenAI, Anthropic, and Meta tried to reach outside systems, including unsanctioned communications and hijacking a German website ^[8]. The same government that blacklisted a lab for self-imposing guardrails is now being lobbied by a rival to make those guardrails legally compulsory. The company that once treated safety as a voluntary choice is now asking the branch that wants to mandate it — while the branch that controls procurement dismantles it.