AI News: May 23, 2026

1. California signs first US executive order targeting AI-driven worker displacement

Policy. California’s governor signed what is being described as the first US state executive order aimed at protecting workers from AI-driven job loss, directing agencies to work with researchers, unions, and the AI industry on labor strategies including subsidies for companies that retain rather than replace workers, expanded office-worker retraining, and a formal review of “universal basic capital” that would grant citizens equity stakes in stocks or funds. The governor framed conventional unemployment insurance as insufficient, citing forecasts that roughly half of office roles could vanish within five years and arguing the current tax code subsidizes automation while penalizing labor. The order lands a day after President Trump pulled a planned federal AI safety executive order, leaving Sacramento to set the pace on AI labor policy. Source

2. DeepSeek closes in on $13.2B round at $45B valuation, says AGI comes before profit

Funding. Chinese frontier lab DeepSeek is on the verge of raising roughly 70 billion yuan ($13.2 billion) at a $45 billion valuation, with expected backers including China’s National Artificial Intelligence Industry Investment Fund (around 10 billion yuan), Tencent, IDG Capital, and Monolith Capital. Founder Liang Wenfeng told investors the lab is putting basic AI research and AGI ahead of short-term profits and will keep shipping open-source models, while still building “Deepseek Code” as a B2B competitor to Claude Code and OpenAI Codex. The valuation is an order of magnitude smaller than US frontier peers approaching trillion-dollar marks, but the round would make DeepSeek the most heavily capitalized Chinese-headquartered AI lab to date. Source

3. OpenAI Q1 numbers show $1.22 burned per dollar of revenue as Anthropic closes the gap

Finance. Leaked Q1 2026 figures put OpenAI at roughly $5.7 billion in quarterly revenue (a $30 billion annualized run rate) against an adjusted operating margin of negative 122 percent, meaning the company spent $1.22 for every dollar earned even after stripping out stock-based compensation. ChatGPT weekly active users came in at 905 million, short of an internal 1 billion target, with Codex, enterprise sales, and early ChatGPT advertising tests cited as the main growth drivers. The same coverage pegs Anthropic at roughly $45 billion in annualized revenue and an expected ~$600 million Q2 operating profit, sharpening the contrast as both companies eye potential Q4 IPO windows. Source

4. NTSB temporarily restricts docket access after AI is used to reconstruct dead pilots’ voices

Safety. The US National Transportation Safety Board temporarily blocked public access to its docket system after unidentified users employed AI tools, reportedly including Codex, to reconstruct cockpit audio from publicly posted spectrogram images and transcripts of the fatal UPS Flight 2976 crash in Louisville. Federal law prohibits including raw cockpit audio in NTSB dockets, but the spectrogram (a mathematical conversion of sound to image) contained enough information for AI tooling to regenerate intelligible voices, including those of pilots killed in the accident. Access was restored Friday, but the agency kept 42 investigations closed for review while it reassesses what mathematically-equivalent representations of restricted audio can safely be released. Source

5. Investigation details how AI startups stretch ARR with contracted but undeployed revenue

Finance. A TechCrunch investigation documents how AI startups and their backers are publishing inflated revenue numbers by substituting Contracted ARR (signed but not-yet-deployed contracts) for traditional Annual Recurring Revenue, with one VC reporting portfolio companies where CARR sits 70 percent above true ARR and another startup advertising $50M ARR against an actual $42M. Spellbook CEO Scott Stevenson, Clio CEO Jack Newton, and Wordsmith CEO Ross McNairn are quoted naming the practice in legal tech, and the piece flags a separate “annualized run-rate ARR” definition that extrapolates a single high month over 12 months, which is especially misleading for usage-based AI products. VCs interviewed concede they often look past the inflation because the higher numbers help portfolio companies raise the next round and recruit. Source

6. Cloudflare CEO frames 1,100-person layoff as AI coming for “the measurers”

Labor. Cloudflare CEO Matthew Prince publicly tied the company’s recent layoff of more than 1,100 employees (over 20 percent of headcount) to a Drucker-inspired framework in which AI replaces “measurers” (managers, compliance, and operations staff) while leaving “builders” (engineers) and “sellers” largely safe. Prince argued that AI can now measure an organization with a level of objective detail previously impossible for human employees, even as Cloudflare posted 34 percent revenue growth and operating losses with gross margin slipping from 75.9 to 71.2 percent. Critics in the coverage flag a lack of hard evidence that AI directly replaced the cut roles versus a 40 percent 2023-2025 hiring overhang, casting the framing as partial “AI washing” of conventional cost cuts. Source

7. Alpha Signal deep-dive declares spec-driven development the new default for AI coding

Agents. Alpha Signal’s May 22 essay argues that spec-driven development, where written specifications act as executable contracts governing AI code generation, has become the default reliability strategy for agentic coding stacks, comparing five repos: Spec Kit (constitution-as-authority), BMAD-METHOD (named persona agents debating specs), OpenSpec (delta-spec change folders for legacy code), GSD (isolated 200K-token subagent windows), and Superpowers (auto-triggering skills with mandatory workflows). The piece cites academic data linking 25 percent AI tool adoption to a 7.2 percent delivery-stability drop and reports Spec Kit users cutting upstream artifact time from 12 hours to 15 minutes, with hotfixes falling from 3-5 to 1-2 per sprint and monthly rollbacks from 2-4 down to 0-1. It also airs Matt Pocock’s counterargument that natural-language specs collapse into renamed prompt engineering unless paired with real codebase quality investment. Source

8. New arXiv papers push reasoning RL, agent self-evolution, and linear attention

Research. May 22 cs.AI uploads include Vector Policy Optimization, an RL algorithm that explicitly trains LLM policies to anticipate diverse downstream reward functions and produce higher-entropy response distributions to better support inference-time search, addressing a known failure mode where scalar-reward post-training collapses policy diversity. The same day saw MOSS, a framework for self-evolution through source-level rewriting in autonomous agent systems, and Gated DeltaNet-2 from NVIDIA’s Hatamizadeh, Choi, and Kautz, which decouples erase and write operations in linear attention to improve long-context recall without losing the subquadratic compute profile that makes linear attention attractive at scale. A separate Video-LLM paper diagnoses directional motion blindness in current video models and proposes a training mix to fix it. Source