Two AI Rulebooks, Written the Same Week

One document, signed June 5, carried both rulebooks at once. The executive order made pre-release vetting of the largest frontier models voluntary — a classified benchmarking process run by the NSA, a Treasury clearinghouse, both opt-in. The national-security memorandum signed the same day — a presidential directive to the defense apparatus — gave the Defense Secretary ninety hard days to update the military's autonomous-weapons policy. Same pen, same week: the models get to choose, the war machine gets a clock. ^[1] The military book has been writing itself in dates since February. That month the Pentagon issued the one rule it has actually enforced on an AI lab: a model had to be available for every lawful purpose — mass surveillance included, fully autonomous weapons included — or its maker would be designated a supply-chain risk, a label that cuts a company out of government work. xAI's Grok agreed. Anthropic refused and lost its sole-provider position. ^[2] June added the ninety-day clock. On October 2 the Pentagon announced a new four-star command for autonomous weapons and drones, due by October 2027 — a deadline, not a suggestion. ^[3] The civilian book lost its edge in sequence. In May, hours before signing, the President canceled his own executive order establishing federal vetting of frontier models, after Musk, Zuckerberg, and Sacks lobbied against it. His reason went on the record.

I really thought that could have been a blocker. — Donald Trump

The vetting that re-emerged in June was a reduced, voluntary thirty-day process. ^[4] In September, OpenAI itself asked Congress for the mandatory version — capability-based rules, incident reporting, pre-deployment evaluation gates — on the stated grounds that voluntary commitments fall short.

we should not pursue it unless and until it can be done safely — OpenAI

This week's answer, signed the same day as that four-star command, was a "morally binding" accord with six firms, built on internal monitoring and outside auditors. The President's stated reason for a handshake instead of a law:

regulation and governance inherently impede innovation — Donald Trump

The accord carries no date and no penalty. The command does. ^[3] Where binding civilian rules are being written now, they're being written below the federal level — and Washington has moved against them there too. New York City proposed mandatory kill switches, third-party validation, and $25,000 fines, with whistleblower rewards; Governor Hochul mandated state registration and seventy-two-hour incident reporting, explicitly because federal regulation is absent. ^[5]

This is not an industry that should self-regulate. — Julie Menin

California's SB 53 arrived gutted — Newsom signed it only after lobbying stripped its safeguards and developer liability, leaving framework publication and incident reporting. ^[6] Against this, the administration drafted an executive order creating a DOJ task force to sue states over their AI statutes and withhold $42.45 billion in broadband funding from any state that refused to repeal them — then paused it after the Senate voted 99-1 against a ten-year ban on state AI regulation. ^[7] And the Pentagon's fortress competition openly favors states that cut regulatory bureaucracy in siting decisions. ^[8] The sequence leaves the ledger's one penalty-column entry exactly where it was in February. The single federal enforcement action on record against an AI company did not ask a lab to be safer. It asked the lab to strip its safety limits for military use — and the lab that said no lost the contract. ^[2]