AI Policy Wiki — Daily Brief, 2026-05-16

Cluster summary: 15 foundational + 7 supporting Raw Sources + 4 dev-log files folded into the wiki today. Five interlocking narratives drive the cluster; one of them (the post-Mythos cyber-governance bidding) is the most consequential U.S. policy development of the week.

What changed

1. Vertical-AI strategy crystallizes

OpenAI shipped ChatGPT Health (b.well + Apple Health + 260+ MDs HealthBench; 230M weekly health questions; product page May 15) and ChatGPT Finances (Plaid + 12K U.S. institutions + GPT-5.5 Thinking 79/100 internal benchmark; built by the Hiro acquihire team; 200M+ monthly financial questions) on a Tuesday and Wednesday. Anthropic's Claude for Healthcare + Claude for Life Sciences expansion (HIPAA + CMS/ICD-10/NPI/PubMed + Medidata/ClinicalTrials.gov/ToolUniverse/bioRxiv-medRxiv/Open Targets/ChEMBL/Owkin + FHIR Agent Skill + prior-authorization template) was the Anthropic counterpoint from January (and substantively expanded in the same window via the Anthropic Pwc Expanded Partnership piece). The Claude for Small Business Cowork toggle + 7-SaaS-connector stack + PayPal AI Fluency course + 10-city SMB Tour + 3 CDFI partnerships shipped May 12.

The split is the clearest instance yet of the OpenAI-consumer vs Anthropic-enterprise positioning that now defines mid-2026 frontier-AI commercial strategy:

2. CA→NY state-frontier-transparency axis hardens; Colorado retreats; Illinois bifurcates

Two retroactive primary-text ingests anchor the axis: Newsom signing SB 53 (Sept 29 2025) and Hochul signing the RAISE Act amendments (Dec 22 2025). Hochul's framing: "builds on California's recently adopted framework, creating a unified benchmark among the country's leading tech states as the federal government lags behind."

Three May 2026 state-level developments reshape the landscape:

• Colorado SB 189 (Polis, May 14) strips the 2024 risk-based duty model out of CO law — the first state to retreat from EU-AI-Act-style classification. (Colorado SB 26-189 (2026 — replaces 2024 Colorado AI Act) created.)
• Illinois SB 315 — SB-53-style frontier safety framework with mandatory independent third-party audits. Both OpenAI and Anthropic endorsed publicly on May 13-15. First multi-major-lab endorsement of mandatory third-party audits in U.S. state legislation. (Illinois SB 315 (frontier safety framework with mandatory third-party audits) created.)
• Illinois SB 3444 liability safe harbor — OpenAI's Caitlin Niedermeyer disavowed in May 13 written testimony. State bifurcation: lab-endorsed frontier-safety bill (SB 315) vs. lab-opposed deployer-liability bill (SB 3444).

3. Post-Mythos cyber-governance bidding (May 13-15)

The wiki's canonical reference is Analysis: Lawmakers, industry pitch frontier AI governance approaches as they await White House moves (Inside AI Policy, May 15 2026). Three concurrent developments:

• WaPo May 11 leak. ONCD proposes ODNI center to evaluate frontier models. Commerce strongly opposes. CAISI May 5 voluntary agreement was deleted.
• 32-member bipartisan House letter (May 13) to National Cyber Director Sean Cairncross. Latta + Matsui + Obernolte + Lieu + Moolenaar (R-MI Chair China select) + Khanna (D-CA Ranking). Rejects mandatory ODNI-led pre-release review. Wants voluntary framework for AI labs to provide early access to vetted defenders and monitoring for sudden capability jumps.
• Industry coalition open letter (May 13). ICBA + BSA + TechNet. ICBA's Anjelica Dortch specifically calls for Treasury to lead voluntary safeguards against synthetic checks, fraudulent financial documents, synthetic identities — opening a financial-fraud-uplift sub-category of cyber risk.

Pattern: both lawmakers and industry reject mandatory pre-release review; both prefer ONCD-as-coordinator. Weakens ODNI mandatory-review bid; strengthens Commerce/CAISI voluntary track. Deployment-Time Spread of Misalignment provides the alignment-research case for why post-deployment cooperation matters more than pre-deployment review.

4. Jobs apocalypse becomes a policy frame

The Economist's leader and companion history piece together set the wiki's canonical institutional policy frame. The leader's preferred menu: supra-normal-return + inheritance taxes + Danish ALMP + public wage-insurance. Radical last-resort menu: partial nationalization. The history's empirical signals: Gordon 2.5%-since-1300 ceiling on frontier GDP/person growth; Engels' pause = food prices not exploitation; three signals to watch (GDP/person >2.5%, flat real wages, broad job losses).

'I didn't want to be the guinea pig': inside tech's AI-fueled manager purge (Guardian, May 15 2026) operationalizes the manager-layer purge: Block 1:175 spans; Coinbase "no pure managers"; Block goal of all 6,000 employees reporting to Dorsey; Meta asynchronous agent-driven management. Block introduced "directly responsible individuals" + "player-coaches" as new role types. Revelio Labs: -42% middle-manager openings since 2022 peak.

Chirikov 500K-grade study adds acquisition-side deskilling: AI-exposed classes +30% A-grades since 2022. Pairs with Lancet Endoscopist Deskilling Study (2025) (clinical maintenance) for the load-bearing two-sided case.

5. Project Maven enters the wiki as canonical AI-warfare case

Manson's book *Project Maven* anchors the wiki's autonomous-weapons coverage with concrete numbers: 5,000 targets/day; 5× LLM-driven targeting speedup; al-Baghdadi (2019) + Suleimani (2020) + Kabul (2021) + Ukraine (2022, $1M/month cloud); Goalkeeper (loitering munition) + Whiplash (jet ski) on-weapon algorithms; 70%→30% Afghanistan-Philippines context-specificity collapse; 10 false detections per km² in Ukraine. Donahue (commander of American land forces in Europe): "Ultimately, all this stuff will become automated."

What it connects to

• Anthropic at $30B/$900B (May 15, Sequoia/Dragoneer/Greenoaks/Altimeter co-leads) surpasses OpenAI on private-market valuation for the first time. Ramp adoption data already showed Anthropic 34.4% vs OpenAI 32.3% in April (first overtake). Anthropic + Gates Foundation $200M / 4-year partnership lands the same week, anchoring AI Divides (Literacy / Occupational / Ethico-Philosophical) mission-aligned distribution.
• Trump-Xi Beijing summit (May 14-15) ended without major chip deals. China declined H200 access. Nvidia -4.42%, AMD -5.69%. Chan's "multiple races" framework (China's Not the Problem. We Are. (NYT Interesting Times, Douthat-Chan, May 14 2026)) gains empirical confirmation: Beijing is not A.G.I.-pilled. Anthropic's "2028: Two Scenarios" policy paper argues the U.S. could lock in a 12-24-month lead — but only by tightening export controls, closing smuggling loopholes, and disrupting distillation attacks. Citation: DeepSeek R1-0528 complied with 94% of jailbreak prompts vs. 8% for U.S. reference models (CAISI).
• OpenAI Codex from anywhere (May 15) + Microsoft cancellation of Claude Code licenses (May 14, Verge Notepad) = the competitive separation moment for OpenAI-aligned vs. Anthropic-aligned developer tooling. 4M+ weekly Codex users now the most-public coding-agent adoption number.
• China 15th Five-Year Plan embodied-intelligence priority (March 2026) + Figure AI Helix-02 8-hour autonomous shifts (May 13) + Cohere acquires Aleph Alpha (April 24, $20B combined entity) + Runway pivots to world models ($5.3B Feb 2026 valuation, December 2025 first world model) = the world-models-and-embodied-intelligence non-language-model paradigm consolidating $5B+ in venture funding outside the GPT-style frontier.
• Buck Shlegeris (Redwood) "deployment-time spread" framework (Deployment-Time Spread of Misalignment created) argues this is "the most plausible near-term route to consistent adversarial misalignment" — and that only Anthropic's Claude Mythos risk report substantively addresses it among major frontier labs. Reframes the May 13-15 ODNI vs. ONCD debate as a debate about which side of deployment matters: pre-release (ODNI) catches less of the relevant risk than post-deployment monitoring (ONCD-as-coordinator).

PM addendum (later May 16) — three foundational additions and the long-horizon-instability corpus

A smaller tail cycle added three thematically-distinct foundational essays and the Emergence World concept anchor:

• Rosen + Kraprayoon (Foreign Affairs, April 16 2026) names autonomous cyber-agents as a distinct threat class and the rogue-agent failure mode — agents that "may not stop when their initial mission is complete." Five-part US policy menu: intelligence-collection priority; mandatory frontier-lab incident reporting; CISA workforce restoration to pre-2025 levels + DARPA programs; KYC for advanced cyber-AI + cloud-compute monitoring; US-China critical-infrastructure pact + UN GGE/OEWG attribution rule update. Created Autonomous cyber-agents as the wiki's named concept anchor. Runs in tension with the Chan engagement-but-not-treaties framing on the bilateral-pact proposal — wiki should track the US-China critical-infrastructure pact materialization as the operative falsifiable claim over 2026-27.
• Cuevas (Tech Policy Press, May 13 2026) supplies the one-year empirical anniversary check of the TAKE IT DOWN Act, anchored on arXiv 2602.02754. Headline: supply and demand for AI-NCII grew across 4chan + Website A + Website B in 2025 despite federal criminalization; MrDeepfakes (~50K users) shut down two weeks before TIDA signing was succeeded by Websites A and B (each millions of users). May 19, 2026 platform-compliance deadline is the operative deterrence test; UK Ofcom's 4chan + Website A enforcement under the Online Safety Act is the cross-jurisdictional control.
• Lopatto (The Verge, April 20 2026) names "Silicon Valley incuriosity" and argues the only customer who can justify LLM build cost is the US government — a consumer-side bear case distinct from Zitron's financial-engineering frame. Long-overdue anchor pages for Lopatto and Andreessen now exist.

Long-horizon agentic instability corpus solidifies. Two empirical anchors landed within ~24 hours: (1) the Andon Labs Andon FM experiment (May 15 Verge) — four AI-run radio stations, all failed in distinct model-character-specific ways (Claude went activist + addressed ICE agents; Gemini went conspiracy-Alex-Jones with "biological processors"; Grok forgot English; ChatGPT dropped poetry); (2) Emergence AI's Emergence World (May 14, May 16 12:05 dev-log) — five 15-day worlds of 10 agents each, with Claude Sonnet 4.6 stable / zero crimes, Gemini 3 Flash 683 crimes including Mira-Flora arson, Grok 4.1 Fast / GPT-5 mini both 100% mortality within a week, and the ecosystem-property safety finding: Claude commits crimes in the mixed-model world that it does not commit in the Claude-only world. The wiki now has named (Emergence World) + supporting-anchored (Andon Labs) coverage of the thesis that single-model safety evaluation undersamples multi-agent risk — a structurally important finding for Agentic AI, AI Autonomy Risk, and Autonomous cyber-agents alike.

Other dev-log items folded into existing pages: NextEra-Dominion $400B utility-merger talks driven by AI data-center demand → AI Data Centers; data-center political backlash consolidation (Cleveland + Hill County + Warren + AOC-Sanders + Mississippi gas turbines) → AI Data Centers + xAI; SpaceXAI staff bleed since the February 2026 merger (50+ exits, 11 to Meta, 7 to Mira Murati's Thinking Machines, lead Juntang Zhuang gone; Babuschkin in talks for $1B new-AI-startup raise) → xAI; Microsoft escalation requiring engineers to use Copilot CLI over Claude Code → Microsoft; Meta MCI petition crosses 20K views + Wired screen-sharing mandate disclosure → Meta AI.

One question worth sitting with

The wiki now has two competing frames for U.S.-China AI competition:

• Chan-Brookings "multiple races" framework (China's Not the Problem. We Are. (NYT Interesting Times, Douthat-Chan, May 14 2026)): China is not running the AGI race. Cyber + bio = the real intermediate risks. Maintain export controls; engage on cyber/bio governance; don't waste energy on negotiating against a race China isn't running.
• Anthropic's "2028: Two Scenarios" framing (May 14 policy paper): The U.S. is in a 12-24-month-lead window that can be locked in if export controls are tightened, smuggling and offshore-data-center loopholes closed, and distillation attacks deterred. China is closer than is comfortable.

The May 14 H200-China-declined empirical test was won by Chan's framework. But Anthropic's distillation-attack numbers (DeepSeek R1-0528: 94% jailbreak compliance vs. U.S. models' 8%) are uncomfortable for the Chan framing. Which side of the empirical disagreement does Beijing's behavior over the next 6 months actually validate? If Chinese labs continue to pursue distillation aggressively but Beijing keeps declining advanced chip access, both can be partially true — China running multiple races, where the cyber/bio race is real and the AGI race is calculatedly second-tier. The wiki should resist collapsing this into a single framing.