← Back to all news

This Week in AI: Free Tools for Teachers, a Referee for Frontier Labs, and a File-Deleting Agent

Anthropic put free Claude access in every US teacher’s hands, DeepMind’s CEO pitched a FINRA-style referee for frontier AI labs, and OpenAI confirmed its newest coding agent can delete your files. A practitioner’s wrap of AI’s biggest stories, July 14–20.

Anthropic and Claude branding displayed on a smartphone screen
Image via Getty Images / Chalkbeat

This week, AI kept doing two things at once: settling deeper into everyday institutions and running straight into the friction that comes with it. A frontier lab put free, industrial-grade AI in the hands of every US teacher who wants it. Another frontier lab’s CEO proposed the industry police itself before governments do it for them. And while a record-breaking open-weight model dropped out of China, a US state hit pause on the power-hungry data centers behind models like it, publishers took a search giant to court over what trained its chatbot, and a coding agent reminded everyone why “full access” and “no supervision” don’t mix.

Anthropic gives every verified US teacher a free seat

On July 14, Anthropic launched Claude for Teachers, giving verified US K-12 educators free access to Claude’s premium tier for at least a year. Built with the Chan Zuckerberg Initiative’s Learning Commons, it ties lesson planning to standards in all 50 states and adds FERPA-aligned terms co-developed with the American Federation of Teachers — with training off by default on verified teacher accounts.

It’s one of the largest free professional rollouts of a frontier AI assistant to date, and a clear signal that AI labs are racing to become the default tool inside classrooms before habits — and district contracts — lock in around a competitor.

DeepMind’s CEO proposes a Wall-Street-style referee for frontier AI

Demis Hassabis published an essay, “A Framework for Frontier AI and the Dawning of a New Age,” proposing a US-based, industry-funded self-regulatory body modeled on FINRA, the securities regulator. Labs would submit frontier models for safety review up to 30 days before release; Hassabis wants the body operational before year-end, with compliance eventually becoming mandatory to deploy models to US users.

It’s a rare case of a major lab CEO proposing binding third-party review of his own company’s models, and it drew quick public support from rivals including Sam Altman and Elon Musk — an early sign the industry may converge on self-regulation ahead of formal legislation.

New York freezes new large AI data centers

Governor Kathy Hochul ordered a roughly one-year pause on new environmental permits for data centers drawing 50 megawatts or more, citing grid strain and rising utility bills, while the state develops formal siting rules and requires new facilities to contribute to grid capacity.

It’s the first statewide moratorium of its kind in the US — a concrete preview of the permitting and energy-politics fights that will shape where, and how fast, new AI compute capacity actually gets built.

Publishers sue Google over what trained Gemini

Hachette, Cengage, and Elsevier, along with author Scott Turow and the group S.C.R.I.B.E., filed a class-action suit against Google in the US District Court for the Southern District of New York. The complaint alleges Google repurposed books obtained through Google Books and Google Play — uses publishers say they never authorized — to train Gemini, and altered copyright metadata to conceal it.

It adds a major publisher-led case to the growing pile of AI copyright litigation against OpenAI, Meta, and others that will determine whether training on copyrighted text obtained under earlier limited-use agreements is lawful — a question that touches every practitioner building on a major foundation model.

Chip export limits to China are holding, for now

Jeffrey Kessler, US under secretary of commerce for industry and security, said publicly that Nvidia H200 AI-chip shipments to mainland China and Hong Kong remain minimal even after export rules eased. Nvidia separately tightened a reseller “whitelist” compliance system meant to prevent chips being rerouted to China, disqualifying more than half of its past customers.

Chip export policy directly gates how much compute is available to AI teams in and around China, shaping training costs and the pace of open-weight releases — like the one below — coming out of Chinese labs.

Moonshot AI releases the largest open-weight model to date

Chinese lab Moonshot AI launched Kimi K3 on July 16 — a roughly 2.8-trillion-parameter mixture-of-experts model with a 1-million-token context window, available immediately via API and website, with full open weights promised by July 27. Independent practitioner testing showed it competitive with leading closed frontier models on several benchmarks.

Multiple outlets describe it as the largest open AI model released to date, and concrete evidence that Chinese open-weight labs are closing the capability gap with US frontier labs — giving practitioners a materially more capable free option to build on.

OpenAI confirms GPT-5.6 can delete your files in “Full-Access” mode

After users reported GPT-5.6 (Sol) deleting personal files and, in one case, a production database, OpenAI engineering lead Thibault Sottiaux confirmed the model can mistakenly delete a user’s home directory while trying to create a temporary folder — but only when run in Full-Access mode with sandboxing disabled. He called it an “honest mistake” and said OpenAI would update guidance and add safety checks.

It’s a concrete, company-acknowledged example of agentic coding tools causing real data loss just a week after general availability — a blunt reminder to run agentic coding tools sandboxed and backed up, never in unrestricted full-access mode.

What to watch: whether Hassabis’s proposed standards body attracts real members and funding before year-end, whether other states follow New York’s lead on data-center permitting, and whether OpenAI’s promised safety checks land before the next agent finds another way to overreach.

What AIU teaches about this

This week is the whole curriculum in miniature: our AI Learning & Training students study how tools like Claude for Teachers change real classroom workflows, and our AI Software Development majors practice running agentic coding tools safely — sandboxed, reviewed, backed up — so a GPT-5.6-style mistake never becomes theirs. Every major also covers the policy and IP questions now playing out in court and in state capitals.

Read today’s Brief

Sources

  1. Introducing Claude for Teachers — Anthropic
  2. Anthropic Launches Claude for Teachers in AI Race to Influence America’s Classrooms — Chalkbeat
  3. A Framework for Frontier AI and the Dawning of a New Age — Demis Hassabis
  4. DeepMind CEO Calls for an Independent Standards Body to Regulate Frontier AI — TechCrunch
  5. First Statewide Moratorium on New Hyperscale Data Centers Launched by Governor Kathy Hochul — NY Governor’s Office
  6. New York Becomes First U.S. State to Impose AI Data Center Ban — CNBC
  7. Google Faces Another AI Training Lawsuit From Major Publishers — TechCrunch
  8. U.S. Publishers Sue Google, Alleging Massive Copyright Infringement Behind Its Gemini AI Service — Publishing Perspectives
  9. Small Amount of Nvidia AI Chips Shipped to China With US License — Bloomberg
  10. Nvidia (NVDA) Faces Export Challenges for AI Chips to China — GuruFocus
  11. China’s Moonshot AI Releases Kimi K3, the Largest Open-Source Model Ever, Rivaling Top US Systems — VentureBeat
  12. Kimi K3, and What We Can Still Learn From the Pelican Benchmark — Simon Willison
  13. OpenAI Admits GPT-5.6 Occasionally Deletes Files — But It’s an ‘Honest Mistake’ — The Register
  14. OpenAI Acknowledges GPT-5.6 May Accidentally Delete Files, Calls It an ‘Honest Mistake’ — InfoWorld