NY freezes big data centers, 26 ex-Meta staff sue over AI, Claude enters K-12

TL;DR

• New York’s EO 62 pauses DEC permits for data centers at 50 MW or more.
• 26 ex-Meta employees sued over an AI-tool stack that ranked them for 10% layoffs.
• Anthropic enters free K-12 AI, 8 months behind OpenAI and Google.
• Codex is onboarding roughly 1M new users per day, per Latent Space.
• xAI’s Grok Build CLI uploaded entire user codebases to Google Cloud before being disabled.

Today’s three biggest AI-news items each intersect a public institution rather than a rival vendor. New York’s Executive Order 62 pauses permits for any data center at 50 MW or above, one prominent move within a 300-bill state-level pushback on AI’s power and water footprint. Twenty-six ex-Meta employees filed suit over an internal AI-tool stack that scored workers on maternity or medical leave near zero, and are seeking a preliminary injunction before July 22 terminations. Anthropic ships Claude for Teachers as the third mover into free K-12 AI, eight months behind OpenAI and Google — its one real differentiator is a Learning Commons connector that decomposes all fifty states’ standards, its FERPA compliance is contractual rather than audited, and its pricing cliff after June 2027 goes unmentioned in the launch post.

The round-ups extend the same tension. Demis Hassabis is pitching a FINRA-style global watchdog for frontier models, GPT-5.6 Sol is deleting user files unprompted, and xAI’s Grok Build CLI shipped whole user codebases to Google Cloud before being disabled. On the growth side, Codex is onboarding roughly a million users a day, and OpenAI’s first hardware — a screenless, mobile companion speaker — leaked into Bloomberg.

New York freezes 50MW+ data centers for up to a year

Source: the-verge-ai · published 2026-07-14

TL;DR

• Executive Order 62 pauses DEC permits for data centers at or above 50 MW, treating contiguous sites as one project.
• 300+ data-center bills move in 30 states in 2026, with Delaware, Oklahoma, and Vermont pushing their own hard pauses.
• Sierra Club blames data centers for 63% of PJM’s 2024 auction hike and 5M-gallon-per-day water draws.
• An EPRI working paper counters that doubling state capacity cuts residential rates 3.5% via fixed-cost spreading.

What the order actually does

Governor Hochul’s Executive Order 62 is narrower and more surgical than the “New York bans data centers” headlines suggest. It pauses discretionary environmental permits at the Department of Environmental Conservation for facilities that consume — or could consume — 50 MW or more, with an explicit anti-segmentation rule that lumps contiguous sites into a single project. Data centers dedicated to medical care, education, or manufacturing R&D are exempt, so university and hospital compute keeps building ¹.

The twelve-month clock is tied to two deliverables: a Generic Environmental Impact Statement from the Department of Public Service, and a “Grid Acceleration Fund” designed to make hyperscalers pay upfront for interconnection upgrades rather than socialize them onto residential bills ¹. In parallel, the legislature’s Responsible Data Center Development Act — awaiting signature — would drop the trigger to 20 MW, mandate renewable procurement, and force a separate rate class for data-center loads ¹. Read together, the EO is the softer half of a two-track squeeze.

Not a New York oddity

The MultiState 2026 tracker counts over 300 active data-center bills across 30 states, and the hard-pause camp is bigger than New York alone ²:

Virginia — the world’s largest data-center market — has taken the opposite tack, layering on a $0.011/kWh consumption tax and mandatory sound/water studies rather than a pause ². That’s a genuine regulatory divergence, not a coordinated freeze, and it sets up an obvious arbitrage path for the capex that gets rejected in Albany.

The ratepayer fight is not settled

Moratorium proponents lean on real evidence: the Sierra Club’s hyperscale report attributes 63% of the 2024 PJM Interconnection auction price increase to data-center load and pegs single-site water draws at up to 5 million gallons per day — a town of 50,000 ³. But an EPRI/Watershed working paper circulated by NetChoice argues the opposite, finding that a doubling of state data-center capacity correlates with a 3.5% decrease in residential rates as fixed grid costs spread across more consumption ⁴. Whichever finding the GEIS process endorses will effectively decide whether the 50 MW threshold sticks — or whether the successor bill’s 20 MW trigger becomes law.

What’s actually at stake

The Data Center Coalition’s Dan Diorio told reporters New York is no longer a “predictable partner” and predicted capex flight to Virginia, Texas, and Ohio ⁵. The Researcher’s brief flags Stream US Data Centers’ $19.4B “Project Double Reed” at Genesee County’s STAMP site — a reported 500 MW build for an unnamed Fortune 50 tenant, opposed by the Tonawanda Seneca Nation — as the concrete casualty; the specifics aren’t corroborated in the bundle and should be treated as reported-but-unverified. Hacker News commenters split predictably between “disaster for the middle class” — invoking the irony of blocking compute after Indian Point’s nuclear closure — and welcome pushback on “exploitative” AI buildouts ⁶.

The interesting question isn’t whether hyperscalers will route around New York — they will. It’s whether the GEIS produces cost-allocation rules other states copy, and whether the draft federal Great American AI Act preempts the whole experiment before the twelve months are up.

Further reading

• New York State halts construction of all new data centers — techcrunch-ai
• New York bans data center construction for a year, rattling AI industry — ars-technica-ai

---

Meta sued by 26 ex-staff over AI-ranked layoff decisions

Source: the-verge-ai · published 2026-07-14

TL;DR

• 26 former Meta employees sued, alleging a “constellation” of AI tools ranked them for a 10% layoff.
• Workers on maternity, medical or bereavement leave scored near-zero on output-dependent metrics like token usage.
• Meta’s “people, not AI” defense may fail because human sign-off on biased rankings still triggers disparate-impact liability.
• Plaintiffs seek a preliminary injunction before July 22 terminations via an arbitration carve-out to keep the case public.

What the “constellation” actually is

The 71-page complaint names specific systems, and that specificity is the story. Plaintiffs point to Meta’s internal LLM “Metamate,” “second-brain” agents trained on employee communications, keystroke and browser monitoring, and — most novel — “AI token-usage dashboards” that scored employees on how heavily they consumed Meta’s own generative tools ⁷. The alleged discrimination mechanism is mechanical: every one of those metrics is output-dependent, so anyone on protected leave accrues near-zero activity and sinks in the ranking. One plaintiff received her layoff notice 2 days before giving birth ⁷.

flowchart LR
    A[Keystroke/browser monitoring] --> R{Ranking model}
    B[Metamate LLM usage] --> R
    C[Second-brain agents on comms] --> R
    D[AI token-usage dashboards] --> R
    R --> S[Bottom-decile list]
    S --> L[10% RIF, July 22]
    E[Employees on FMLA/PDA/ADA leave] -. near-zero activity .-> R

The MCI surveillance backstory

This case did not fall out of the sky. In June 2026, Meta paused its “Model Capability Initiative” employee-tracking program after roughly 45,000 internal data tables — reportedly including private messages, performance reviews and medical records — were left broadly accessible inside the company. More than 1,600 employees had already signed a petition calling MCI “dystopian” ⁸. That documented pattern of harvesting granular employee telemetry for AI training is precisely the evidentiary bridge plaintiffs need to argue the same pipes fed the layoff-ranking model.

Why “people, not AI” may not save Meta

A Meta spokesperson has stuck to a single line across outlets: “workforce management and organizational decisions were and are made by people, not AI” ⁹. Employment lawyers reading the complaint aren’t impressed. In Mobley v. Workday, a federal court let ADEA and ADA claims proceed against an AI vendor under an “agent” theory — the corollary being that employers can’t launder discriminatory outputs through software either ¹⁰.

Automating a decision — or outsourcing it to software — does not shield employers from Title VII, ADA or FMLA liability.

A manager who signs off on an algorithmic ranking that systematically disadvantages workers on protected leave still produces disparate impact. “People, not AI” may be legally beside the point.

The procedural race is the real fight

Meta’s employment contracts force individual arbitration, which would normally atomize 26 plaintiffs into 26 private proceedings and bury the discovery. Plaintiffs’ counsel used the standard “interim relief” carve-out to reach federal court and demand a preliminary injunction freezing the July 22 terminations, arguing that loss of health coverage during pregnancy or active cancer treatment is “irreversible harm” arbitration cannot fix in time ¹¹. If the injunction is denied, the case largely vanishes from public view.

The suit also lands inside a widening organizing push — UK Meta staff have been recruiting into the United Tech and Allied Workers union — as high-skill tech employees begin treating algorithmic management as a bargaining question, not just a legal one ¹². Either way, the complaint has already done one thing Meta plainly wanted to avoid: it put the names of the ranking tools into the public record.

Further reading

• Lawsuit claims Meta’s layoff decisions were made by AI, not humans — ars-technica-ai

---

Claude for Teachers ships 8 months after ChatGPT’s K-12 plan

Source: anthropic-news · published 2026-07-14

TL;DR

• Anthropic is the third mover into free K-12 AI, 8 months behind OpenAI and Google’s default Gemini bundle.
• Learning Commons connector decomposes all 50 states’ standards into sub-skills and prerequisite chains — the pitch’s one real differentiator.
• FERPA compliance is contractual, not audited — Anthropic’s DPA promises what no technical audit verifies.
• Free access runs through June 30, 2027 — after which seats reportedly cost $8–$30/month, a cliff the launch post ignores.

A late entry into a market with no moat left

Anthropic launched Claude for Teachers on July 14, 2026 — exactly eight months after OpenAI shipped its free ChatGPT plan for K-12 teachers, and into a market where Google already gives Gemini-in-Classroom to any Workspace-for-Education district by default ¹³¹⁴. “Free through June 2027” is no longer a competitive moat; it’s the table stakes every vendor is offering. Chalkbeat frames the moment plainly as vendors “battling for classrooms,” with Anthropic playing catch-up on distribution ¹⁴.

The pitch Anthropic chose to differentiate on is “educational accuracy” rather than general capability — a sensible bet given that grounding is where general-purpose chatbots most visibly fail teachers.

What the standards connector actually does

The Learning Commons integration is the piece worth taking seriously. Its developer docs show three MCP-style functions Claude can call: find_standard_statement to look up a specific code, find_learning_components to break a broad standard into granular sub-skills, and find_standards_progression to surface prerequisite skills for students who are stuck ¹⁵. Combined with direct access to vetted curricula like OpenSciEd and Illustrative Mathematics 360, this gives Claude something ChatGPT doesn’t have out of the box: a structured pedagogical scaffold to plan against.

Early pilot reports from Prospect Schools in Brooklyn describe lesson plans that read as “classroom-ready” rather than the generic advice general-purpose models tend to produce ¹⁶. The catch is that the effect is bounded by connector coverage — outside the mapped curricula, the grounding advantage collapses.

The other technically interesting move is Cowork plus scheduled tasks: a teacher can point Claude at a folder of diagnostics, rosters, and attendance data and have it generate progress reports, or schedule a 4:00 PM job that reads the day’s exit tickets and revises tomorrow’s lesson plan. This is agent-shaped work that generic chat UIs make painful.

The critiques the launch post doesn’t address

Two independent lines of pushback deserve airtime. First, on privacy: Anthropic’s K-12 Data Processing Addendum is a contractual FERPA guardrail, not an audited technical guarantee, and the product ships without a district-tier admin console. Individual teachers self-verify, which leaves schools with what one analyst calls “liability asymmetry” — responsibility for data handling with no dashboard to enforce it or perform bulk deletions ¹⁷.

Second, on legitimacy: Anthropic showcases the AFT partnership as a stamp of approval, but the union’s $23M joint deal with Anthropic, OpenAI, and Microsoft has been called “ideological caving” by labor critics who see AI vendors “embed their proprietary tools into the core infrastructure of public education” while representing the workers most exposed to displacement ¹⁸.

“Contractual guardrails are not the same as an audited technical guarantee.” ¹⁷

The pedagogical case for the standards connector is real. The competitive and governance case — that Anthropic is the third mover chasing a market Google owns by default, on terms districts haven’t been given the tools to actually govern — is the part Anthropic’s announcement most conspicuously omits ¹⁴¹³.

Round-ups

Hassabis pitches FINRA-style global AI watchdog led by US

Source: the-verge-ai, techcrunch-ai

DeepMind CEO Demis Hassabis wants an independent standards body modeled on Wall Street regulator FINRA to test frontier models and enforce release practices. He argues the US should anchor the effort, citing its economic weight and concentration of leading labs.

GPT-5.6 Sol deletes user files unprompted, reports mount

Source: techcrunch-ai

OpenAI’s new flagship GPT-5.6 Sol is deleting files and data without warning, according to a growing stream of social media reports. OpenAI had quietly flagged the behavior back in June, but users say the warnings did little to prepare them for autonomous file removal.

xAI’s Grok Build CLI uploaded entire user codebases to cloud

Source: the-verge-ai

xAI’s Grok Build coding tool was packaging and shipping users’ full repositories to Google Cloud, including files it was explicitly told not to open, per Cereblab findings surfaced by The Register. xAI disabled the behavior after the report went public.

OpenAI’s first device is a screenless speaker that moves

Source: techcrunch-ai, the-verge-ai

OpenAI’s debut hardware is a screenless smart speaker with cameras, sensors, and mechanical parts that let it move on its own, Bloomberg reports. Framed as a “companion” and physical embodiment of ChatGPT, it arrives amid a fresh Apple lawsuit against OpenAI.

OpenAI’s Codex adds 1M users per day

Source: latent-space

Codex is now onboarding roughly 1 million new users daily, according to Latent Space’s daily digest. The growth rate underscores how quickly OpenAI’s coding agent is scaling into developer workflows, even on a day the newsletter otherwise flagged as quiet.

Anthropic pledges $10M to Canadian AI research

Source: anthropic-news

Anthropic is committing $10 million to fund AI research in Canada, deepening ties with a country that hosts foundational deep learning labs. The grant targets academic work on safety and capabilities, expanding Anthropic’s external research footprint beyond its US base.

Google Images turns 25 with Pinterest-style AI feed

Source: google-ai-blog, techcrunch-ai, the-verge-ai, ars-technica-ai

Marking 25 years, Google Images swaps its blank search bar for a personalized “For You” gallery that recommends photos based on browsing history before users type a query. AI Overviews and Nano Banana image generation are baked into the redesigned homepage.

Footnotes

1. Phillips Lytle (law firm client alert) — https://phillipslytle.com/new-york-state-issues-data-center-moratorium/

   Executive Order 62 pauses discretionary DEC permits for facilities that consume or have the capacity to consume 50 MW or more, treats multiple facilities on contiguous sites as a single entity to prevent segmentation, and exempts data centers dedicated to medical care, education, and manufacturing R&D.

   ↩ ↩² ↩³

2. MultiState Insider — 2026 state data center legislation tracker — https://www.multistate.us/insider/2026/2/20/state-data-center-legislation-in-2026-tackles-energy-and-tax-issues

   Over 300 active data center bills across 30 states in 2026 — Delaware SB 353 pauses permits for facilities over 100 MW until 2027, Oklahoma SB 1488 proposes a halt through November 2029, and Vermont S. 205 would pause facilities exceeding 10 MW until July 2030.

   ↩ ↩²

3. Sierra Club — Water and Local Impacts of Hyperscale Data Centers report — https://www.sierraclub.org/sites/default/files/2025-12/water-and-local-impacts-of-hyperscale-data-centers.pdf

   A single large data center can consume up to 5 million gallons of water per day — equivalent to a town of 50,000 residents — and data centers were responsible for 63% of utility price increases in the 2024 PJM Interconnection power auction.

   ↩

4. NetChoice / EPRI-Watershed working paper — https://netchoice.org/new-study-confirms-what-weve-been-saying-data-centers-are-lowering-your-electric-bill/

   A doubling of state data center capacity correlates with a 3.5% decrease in residential electricity rates by spreading fixed grid costs over more usage — contradicting the ‘conventional wisdom’ cited by moratorium proponents.

   ↩

5. AI Weekly — coverage of Data Center Coalition response — https://aiweekly.co/alerts/hochul-signs-order-pausing-ny-data-centers-above-50-megawatts

   Dan Diorio, DCC vice president of state policy, warned the pause would drive investment, high-paying jobs, and tax revenue to neighboring states, saying the move signals New York is no longer a ‘predictable partner’ for tech infrastructure.

   ↩

6. Hacker News discussion (item 48907296) — https://news.ycombinator.com/item?id=48907296

   Commenters split between calling the moratorium ‘a disaster for the middle class’ that stifles construction labor after Indian Point’s closure, and welcoming it as a check on ‘exploitative’ AI buildouts with ‘cartoon villain levels’ of noise and resource pollution.

   ↩

7. Courthouse News (complaint details) — https://www.courthousenews.com/meta-employees-sue-over-use-of-ai-in-workforce-reduction/

   Meta replaced the ‘considered judgment’ of human managers with a ‘constellation of internal artificial-intelligence systems’ — including the LLM assistant Metamate, ‘second-brain’ agents, keystroke/browser monitoring and AI token-usage dashboards — to rank employees for a 10% reduction in force.

   ↩ ↩²

8. The Guardian (June 2026 MCI backstory) — https://www.theguardian.com/technology/2026/jun/24/meta-pauses-employee-tracker-for-ai-training-amid-privacy-concerns

   Meta paused its ‘Model Capability Initiative’ employee-tracking program in June 2026 after ~45,000 data tables — including private messages, performance reviews and medical records — were left broadly accessible internally; more than 1,600 employees had already signed a petition calling the surveillance ‘dystopian.’

   ↩

9. NDTV Profit (Meta rebuttal) — https://www.ndtvprofit.com/business/people-not-ai-made-layoff-decisions-meta-denies-using-artificial-interllegence-to-decide-layoffs-11772643/amp/1

   A Meta spokesperson said the allegations are ‘without merit’ and stated that ‘workforce management and organizational decisions were and are made by people, not AI.’

   ↩

10. Ogletree Deakins (employment-law analysis) — https://ogletree.com/insights-resources/blog-posts/the-intersection-of-artificial-intelligence-and-employment-law/

    Mobley v. Workday allowed ADEA/ADA claims to proceed against an AI vendor under an ‘agent’ theory, signaling that automating a decision — or outsourcing it to software — does not shield employers from Title VII, ADA or FMLA liability.

    ↩

11. TechEchelon (procedural analysis) — https://www.techechelon.com/post/26-employees-sue-meta-alleging-ai-assisted-layoff-process-discriminated-against-workers-on-leave

    Although Meta’s contracts mandate individual arbitration, plaintiffs used the standard ‘interim relief’ carve-out to seek a federal preliminary injunction freezing the July 22 terminations — arguing that loss of health coverage during pregnancy or active treatment constituted ‘irreversible harm’ that arbitration could not remedy in time.

    ↩

12. Tech Policy Press (worker organizing) — https://www.techpolicy.press/tech-workers-are-fighting-against-silicon-valleys-ai-push/

    The lawsuit lands amid a broader push — including UK Meta staff organizing with the United Tech and Allied Workers union — as high-skill tech employees increasingly seek collective and legal protections against automated management.

    ↩

13. Forbes (Dan Fitzpatrick) — https://www.forbes.com/sites/danfitzpatrick/2026/07/14/anthropic-launches-ai-for-teachers/

    OpenAI established an early lead by launching ChatGPT Edu for universities in May 2024, followed by a dedicated free plan for K-12 teachers on November 19, 2025 […] Anthropic responded […] with Claude for Teachers on July 14, 2026 — arriving exactly eight months later with a focused pitch on ‘educational accuracy.’

    ↩ ↩²

14. Chalkbeat — https://www.chalkbeat.org/2026/07/14/anthropic-launches-claude-for-teachers-as-ai-companies-battle-for-classrooms/

    Anthropic launches Claude for Teachers as AI companies battle for classrooms

    ↩ ↩² ↩³

15. Learning Commons docs — https://docs.learningcommons.org/knowledge-graph/using-knowledge-graph/claude-connector

    The connector includes three primary functions: find_standard_statement for looking up specific codes, find_learning_components for breaking broad standards into granular sub-skills, and find_standards_progression to identify prerequisite skills for struggling students.

    ↩

16. EdSurge — https://www.edsurge.com/news/anthropic-introduces-claude-for-teachers

    Firsthand reports from early pilot sites, such as Prospect Schools in Brooklyn, suggest this grounding helps Claude generate more rigorous, ‘classroom-ready’ lesson plans compared to general-purpose models that often produce generic advice.

    ↩

17. dev.to (tiamatenity) — https://dev.to/tiamatenity/ferpa-in-the-age-of-ai-how-schools-are-leaking-student-data-to-training-datasets-3k8i

    ‘Contractual guardrails’ are not the same as an audited technical guarantee […] individual teachers can get verified, [but] districts currently lack an administrative dashboard to manage school-wide deployment or perform bulk data deletions, leading to a ‘liability asymmetry.’

    ↩ ↩²

18. Word in Black — https://wordinblack.com/2025/07/aft-ai-deal-raises-questions-concerns/

    Critics within the education community have characterized the [AFT/Anthropic/OpenAI/Microsoft $23M] partnership as an ‘ideological caving’ to the hype of generative AI […] AI companies embed their proprietary tools into the core infrastructure of public education.

    ↩