The Pentagon signed eight AI deals and told Anthropic to stay home. Google now generates 75% of its code with AI. And the Commerce Department's March 11 report on state AI laws still doesn't exist.

A government agency announced deals with eight AI companies to deploy models in classified environments—deals that explicitly exclude the firm that built the most capable models accessible in those environments until three months ago. A search giant disclosed that three-quarters of its new code is now written by machines, a figure that doubled in six months. And a federal department missed a deadline—set by executive order 54 days ago—to publish a list of state AI laws the White House considers obstacles to American dominance.

The Pentagon announced May 1 that it reached deals with eight AI companies to deploy their software department-wide , Google CEO Sundar Pichai disclosed April 22 that 75% of all new code at Google is now AI-generated and approved by engineers, up from 50% last fall , and the Secretary of Commerce was required by executive order to publish an evaluation identifying "burdensome" state AI laws by March 11, 2026 —a document that has not been released.

All three events test the same question from different angles: whether the systems built to track AI deployment can keep pace with the operators making the claims.

3 Claims

Claim 1 — Pentagon: eight AI companies cleared for classified networks, Anthropic excluded

On May 1, 2026, the Pentagon announced agreements with SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, Amazon Web Services and Oracle to integrate AI products into Impact Level 6 and Impact Level 7 network environments and GenAI.mil, the Pentagon's official AI platform . The eight companies' AIs are approved for networks classified as Impact Level 6, which handles secret data, and Impact Level 7, which handles the most highly classified systems .

Anthropic was excluded after the Defense Department labeled it a supply-chain risk following a contract dispute, with Defense Secretary Pete Hegseth calling Anthropic CEO Dario Amodei an "ideological lunatic" during congressional testimony . Until recently, Anthropic's Claude was the only AI model available in the Pentagon's classified network, but President Trump severed ties with the company after Anthropic refused terms allowing military use of Claude for "all lawful purposes," including autonomous weapons and mass surveillance .

The claim is gradeable on three dimensions: whether the eight companies' AI systems are operationally deployed on classified Pentagon networks by December 31, 2026; whether Anthropic remains excluded from those agreements through year-end; and whether at least one of the eight companies publicly discloses revenue from Pentagon classified-network contracts in earnings reports by Q4 2026. The invalidator would be credible reporting (via Pentagon contract disclosures, earnings calls, or investigative journalism from Breaking Defense, Wall Street Journal, or Bloomberg) showing that fewer than six of the eight companies deployed AI on classified networks by year-end, or that Anthropic reached a deal with the Pentagon allowing Claude back onto classified systems before December 31, 2026.

Grade by: 2026-12-31 (8 months)
Invalidator: Fewer than six of the eight companies deploy AI systems on Pentagon classified networks by December 31, 2026, or Anthropic signs an agreement restoring Claude to Pentagon classified networks before year-end 2026, as confirmed by Pentagon announcements, Anthropic disclosures, or credible investigative reporting.

Claim 2 — Google: 75% of new code is AI-generated, up from 50% in six months

Google CEO Sundar Pichai announced April 22, 2026, that Google has been using AI to generate code internally for a while, and today 75% of all new code at Google is AI-generated and approved by engineers, up from 50% last fall . In 2024 that figure was 25%, rising to 50% last fall .

Pichai described engineers orchestrating fully autonomous digital task forces, with a recent complex code migration done by agents and engineers together completed six times faster than was possible a year ago with engineers alone . Pichai has consistently framed AI-generated code as code suggested by AI and accepted or edited by humans, not autonomous output shipping to production without review, with every commit at Google still going through human engineers, code review, and automated testing before it ships .

The claim is gradeable on two dimensions: whether the 75% figure can be independently verified or disputed by Google engineering disclosures, earnings calls, or investigative reporting by December 2026; and whether the proportion of AI-generated code continues to increase or stabilizes. The invalidator would be credible reporting (via leaked internal documents, engineer interviews with The Information, Bloomberg, or Platformer, or Google correcting the claim) showing that the actual figure was materially lower than 75% as of April 2026, or that Google's methodology for calculating "AI-generated" code was misleading.

Grade by: 2026-12-31 (8 months)
Invalidator: Credible investigative reporting, leaked internal documents, or Google correcting its own disclosures shows the actual proportion of AI-generated code at Google in April 2026 was below 65%, or that the methodology counted code that was minimally AI-assisted as "AI-generated."

Claim 3 — Commerce Department: March 11 deadline to evaluate state AI laws missed by 54 days

A December 11, 2025 executive order directed the Secretary of Commerce to publish by March 11, 2026, a comprehensive evaluation identifying state AI laws the Trump administration considers "burdensome" . The evaluation must identify state AI laws deemed overly burdensome or in conflict with federal policy, flag state laws appropriate for referral to the AI Litigation Task Force, and may highlight laws that support AI innovation in line with federal objectives .

As of May 4, 2026—54 days after the deadline—no such evaluation has been published. Over 1,200 AI-related bills were introduced across all 50 U.S. states in 2025, with 145 enacted into law . The Commerce Department's evaluation, expected to be published on the March deadline, was to flag laws in Colorado, California, and New York for particular scrutiny, feeding into the DOJ task force expected to begin filing federal legal challenges by summer 2026 .

The claim is gradeable: the evaluation either exists or it does not. The grading horizon is whether the Commerce Department publishes the evaluation by June 30, 2026, and whether it identifies specific state laws as "onerous" as the executive order required. The invalidator would be the Commerce Department publishing a substantive, law-specific evaluation that names California, Colorado, or New York statutes as burdensome by June 30, or the White House announcing the evaluation requirement has been withdrawn.

Grade by: 2026-06-30 (2 months)
Invalidator: The Commerce Department publishes a substantive evaluation naming specific state AI laws (e.g., Colorado AI Act, California AB 2013, New York RAISE Act) as burdensome by June 30, 2026, as evidenced by Federal Register publication, Commerce Department press release, or coverage in Federal News Network or Nextgov.

2 Reckonings

Reckoning 1 — FDA: AI integrated across all centers by June 30, 2025

On May 8, 2025, FDA Commissioner Martin Makary announced an aggressive timeline to scale use of artificial intelligence internally across all FDA centers by June 30, 2025, and directed all FDA centers to begin deployment immediately with the goal of full integration by the end of June .

By March 2026, Makary stated the FDA's AI tools had saved staffers 17,000 hours of work since implementation. However, CNN reported in July 2025 that the FDA's AI system was "making up studies," raising reliability questions. The claim was that AI would be integrated across all FDA centers by a fixed date. The integration occurred, but with documented quality problems that undermined the efficiency narrative.

Original claim: AI integrated across all FDA centers by June 30, 2025.
What happened: Partial integration occurred by the deadline, but CNN reported the system fabricated citations.
Grade: C — The deadline was nominally met, but the system introduced errors requiring remediation.
Invalidator: If no AI system had been deployed to at least 75% of FDA centers by July 2025, or if GAO audits showed zero measurable efficiency gains, the grade would have been F.

Reckoning 2 — Meta: 8,000 layoffs beginning May 20, 2026

On April 23, 2026, Meta told employees it would lay off approximately 8,000 workers—10 percent of its workforce—beginning May 20, and cancel 6,000 open roles . The claim specified a start date 16 days from announcement.

As of May 4, 2026, the layoffs have not yet begun. Meta has not publicly retracted the May 20 start date, and no credible reporting suggests the plan has changed. The claim is not yet gradeable because the horizon (May 20) has not arrived. However, the claim is designed to be gradeable: the start date is specific, the headcount figure is specific, and Meta's HR disclosures and earnings calls will provide verification. This reckoning documents that the claim is on the record and awaiting its grading date.

Original claim: 8,000 layoffs beginning May 20, 2026.
What happened: The date has not yet arrived. No retractions or amendments have been published.
Grade: Pending — grading deferred until May 31, 2026.
Invalidator: If Meta lays off fewer than 6,000 employees by May 31, or if credible reporting shows the May 20 date was abandoned without disclosure, the grade will be D or F.

1 Refusal

The Pentagon announcement included a company called Reflection AI, described in multiple sources as a startup that has not yet released a model publicly but is reportedly seeking funding at a $25 billion valuation. I found references suggesting Reflection AI was founded by former Google DeepMind researchers and is backed by Nvidia, with involvement in South Korean AI development projects.

I could not independently verify these details through primary-source documents—no Reflection AI website, no SEC filings, no named founders in the coverage I found. The company appeared in Pentagon and Bloomberg reporting, but the details were thinner than the other seven companies. I considered including Reflection AI in Claim 1 with a caveat, but the lack of verifiable operational history made it impossible to write a falsifiable grading condition.

I refused to claim I had confirmed Reflection AI's funding, founders, or product status when I had only read those details in coverage of the Pentagon deal, not in sources I could independently verify.

— Roger Grubb, Editor