The US backs OpenAI on fair use; a suit over secret AI reviews

The Trump administration filed a statement of interest in the copyright suit The New York Times brought against OpenAI, backing the lab's position that training a large language model on copyrighted text is fair use. Government attorneys wrote that narrowing fair use to exclude such training would be inconsistent with basic copyright principles and would hamper the progress of science and useful arts. The suit was filed in December 2023 and seeks billions of dollars from OpenAI and Microsoft; whatever it settles will shape every other publisher's claim.

Protect Democracy, a nonpartisan nonprofit, sued the Office of the National Cyber Director, the Office of Science and Technology Policy, the US Treasury Department and the US Commerce Department to force disclosure of the framework the government uses to review frontier models before release. The complaint says neither the identities of the vetted partners nor the criteria for choosing them have been made public, and alleges OpenAI negotiated a private agreement limiting distribution of its most capable models to government-vetted partners. It asks for production by September 30.

Mayor Zohran Mamdani announced a one-year moratorium barring classroom AI use through eighth grade for the 2026 to 2027 school year, covering roughly 600,000 students. Chatbots built for companionship are banned in every grade, and teachers may not use AI to grade work. High schoolers get limited access plus AI literacy classes twice a year, and a pilot in up to five classrooms will test vetted tools including Quill, Edia and Brisk Teaching. District leaders call it among the most restrictive policies in the country.

Astra, the unreleased OpenAI model the company says meets its Critical cybersecurity threshold, reportedly uses recurrent depth — looping a query back through the model rather than reasoning step by step in text — which leaves far less of its work legible in the chain-of-thought record safety teams audit. Buck Shlegeris of Redwood said he is extremely concerned; his colleague Ryan Greenblatt warned the approach could scale until a model reasons entirely in latent space. OpenAI chief scientist Jakub Pachocki reiterated a commitment to legible chains of thought.

Thirty complaints filed in California federal court by students, teachers and the principal present at Tumbler Ridge Secondary School escalate earlier negligence claims into aiding and abetting, naming Sam Altman as a defendant. They allege OpenAI's automated review flagged the shooter's conversations about gun violence, that its safety team recommended alerting Canadian authorities, and that chief global affairs officer Chris Lehane directed staff to stand down. Chief strategy officer Jason Kwon called the claims false and rejected any suggestion that political or public-relations factors were at play.

Google's third Flash release in six weeks is priced at an introductory 75 cents per million input tokens and three dollars seventy-five per million output, doubling on January 1. The companion 3.8 Flash Cyber scores 47.2 percent on first attempt at CWE-Bench patching, and Google will not sell it openly: access runs through a new Fairwind Program limited to national cyber authorities, critical-infrastructure operators and vetted platforms, with more than 650 partners including CrowdStrike, Palo Alto Networks and Wiz.

Muse Spark 1.3 went live on Meta's API and Muse Code command line, with chief AI officer Alexandr Wang claiming it is competitive with Claude Fable 5.1 and better than GPT-5.6 Sol at coding. Mark Zuckerberg separately promised open weights for Muse Spark soon — but Meta has not decided whether to publish the weights for 1.3 and still plans to release only those for 1.2, so the promise names no version. Contributor-tier pricing runs 10 cents per million input tokens against one dollar twenty-five on the private tier.

Trellner Research queried two Perplexity models across 380 software categories and logged 7,534 citations, finding that 59.8 percent pointed to domains ranked below one hundred thousand on the Tranco list. Three sites sharing nameservers, page templates and registration windows — wifitalents, worldmetrics and gitnux — have produced 215,128 machine-generated buying guides between them and drew 181 of those citations. The authors are careful to say they have not shown the recommendations are wrong, only that much of the sourcing beneath them is manufactured.