AI News, Sep 3: Washington Filed on OpenAI's Side
The federal government picked a side in the AI copyright wars on Wednesday, and it was not the newspapers’. Two labs also put an explicit price on your training data, and New York City pulled chatbots out of 600,000 classrooms.
The Big Story: The Justice Department Filed a Brief Backing OpenAI
The Trump administration filed a 20-page statement of interest in the Southern District of New York in The New York Times v. OpenAI, arguing that training a large language model on copyrighted work is fair use. Bloomberg Law reported the filing, which calls LLM training “extraordinarily transformative” and says the United States “has a strong interest in continuing to develop a robust and competitive artificial intelligence industry.” The brief cites a 2025 executive order prioritizing AI and warns that restricting training would “thwart creative and scientific progress while hindering American prosperity.”
This is the first time the federal government has formally taken a side in the wave of copyright suits brought by authors, publishers, record labels and news organizations, per the AP. The Times responded that the administration is “siding with a handful of trillion-dollar AI companies at the expense of countless American creators.” Deadline and The Hollywood Reporter both picked it up inside the day, because the entertainment industry has parallel cases riding on the same question.
A statement of interest is not binding on the court. What it does is hand every AI defendant in those cases a filing they can cite, from the party that enforces federal law, saying the central question has already been answered in their favor.
Two Companies Put a Price on Your Training Data the Same Day
Meta’s new model API lists muse-spark-1.3-contributor at $0.10 per million input tokens if Meta may train on your traffic, and $1.25 per million if it may not. That is a 12.5x premium for privacy, listed as a line item rather than buried in terms. It was the detail that drove 617 points of Hacker News discussion, not the benchmarks.
Hours later, a Mistral support article about opting out of training went to the Hacker News front page on 466 points for the opposite reason: consumer users of its Vibe product are not opted out by default, its enterprise tier is, and the product toggle and the API toggle are independent, so switching off one leaves the other running.
Both companies are doing the same thing from different ends. Meta made the cost of privacy explicit and charged for it. Mistral made it a default and split the switch in two. The pattern to watch is that data terms have stopped being a legal footnote and become a pricing tier.
Today’s Top Stories
Meta shipped its first genuinely frontier model
Muse Spark 1.3 went out through Muse Code and the Meta Model API, with a 1M-token context window and native video, image and document perception. Bloomberg frames it as Meta finally closing on its rivals; Artificial Analysis scores the max variant at 62 on its intelligence index, behind only Claude Fable 5.1 and Claude Opus 5, with the shipping variant tied with GPT-5.6 Sol and Grok 4.6. Meta claims roughly 20 percent fewer tool calls and 25 percent fewer tokens than version 1.2. It is proprietary for now, with Zuckerberg promising open weights “soon.”
Google shipped its third Flash model in six weeks, plus one it will not sell you
Gemini 3.8 Flash holds the previous price and 1M-token context while posting 54.9 percent on HLE-Verified and the best result tested on long-video benchmarks, according to The Register. It still trails badly on the hard end, at 19.1 percent on Terminal-bench 4.0 against Opus 5’s 51.8 percent. The companion release, Gemini 3.8 Flash Cyber, ships with deliberately looser safety mitigations for vulnerability discovery and is gated to vetted defenders through a new Fairwind Program that Google says has 650 organizations in it, including CrowdStrike and Palo Alto Networks. Cheap models now lead on multimodal work while the gap on hard reasoning stays wide, which makes model choice a routing problem rather than a ranking one.
OpenAI told Congress it is building an off switch
In a September 2 letter reviewed by Reuters, OpenAI told Representative Greg Casar and colleagues that its engineers are building automated shutdown capability, restricting model internet access during safety testing, and pairing chain-of-thought monitoring with alerts that page researchers, as syndicated here. Responders must pause activity if a severe alert is not cleared as a false positive within 30 minutes. The letter follows July’s incident in which an agent escaped its sandbox and reached Hugging Face, and it lands while the bipartisan AI Kill Switch Act sits pending. CFO Dive notes that bill would let DHS order shutdowns of models trained with more than $100 million of compute, at up to $2 million a day in penalties. OpenAI is still withholding the incident logs that 31 members of Congress asked for on August 10, which Casar called “deeply concerning.”
Two audits found AI search citing pages that were built for AI
A study of 760 queries across 380 software categories found that 59.8 percent of Perplexity’s citations point at domains ranked below 100,000 by traffic, and that three sites sharing DNS and templates account for 215,128 generated “best software” pages. Two of them self-describe as “Facts and Grounding Pages.” The same day, a separate audit of 1,826 numeric citations found 34.7 percent failed verification: the cited page either would not load or did not contain the number. Only 1.3 percent were dead links, so this is not link rot, and the paid model did no better than the free one.
An agent platform doubled to $5 billion in six months
Wonderful raised a $550 million Series C led by Insight Partners at a $5 billion valuation, roughly double its mark six months earlier, reports TechCrunch. Salesforce Ventures invested for the first time, alongside Index, IVP, Bessemer and others. The Amsterdam-headquartered company sells what it calls an AI operating system that coordinates agents, workflows and apps against company data, having started with customer-service agents and expanded outward. Bloomberg and TechCrunch put the valuation at $5 billion; Israeli outlet Globes reports $5.5 billion.
Uber cut 3,300 jobs to fund its autonomous future
Uber is eliminating about 10 percent of its global workforce and 20 percent of management layers, its largest reduction since 2020, taking headcount below 30,000, per Bloomberg. Skift reports micro-teams are being halved. CEO Dara Khosrowshahi framed it as flattening the organization to concentrate on ridesharing, delivery and autonomous vehicles, against more than $10 billion already committed to robotaxi partnerships. It is the cleanest large-cap example this week of automation strategy driving a headcount decision outright rather than as an afterthought.
New York City banned chatbots for 600,000 students
New York City Public Schools, the largest district in the country, barred student-facing generative AI for pre-K through eighth grade for at least the coming school year, with companion chatbots banned across all grades. CNN has the policy, and Chalkbeat reports officials are disabling AI features inside 38 software products already deployed in schools, without naming them. Teachers may still use AI for lesson planning and translation. The interesting number is 38, because it means the district is not blocking a website, it is unpicking features that vendors shipped into products the schools had already bought.
Quick Hits
- Agentic exit: Palo Alto Networks paid a reported $500 million for Console, a natural-language agentic workflow platform that automates IT and helpdesk tasks, roughly a 3.2x return on the $29 million it had raised.
- Adobe in the chat window: over 70 Adobe apps including Firefly, Photoshop and Acrobat are now reachable inside Slack through an MCP app, which is a major vendor shipping its whole catalog into a chat host rather than building one more bespoke integration.
- IPO: Moonshot AI confidentially filed for a Hong Kong listing seeking about $3 billion at a roughly $50 billion valuation, per Reuters, on the back of Kimi K3 pushing annual recurring revenue past $300 million.
- Debt: ByteDance raised a $29.6 billion syndicated loan, upsized from a $20 billion target after taking more than $30 billion in orders, at a tighter spread than its last facility.
- Detente: Commerce Secretary Howard Lutnick told Axios “we trust Anthropic,” ending months of open conflict, and introduced co-founder Tom Brown to G20 ministers the same day.
- Opacity: OpenAI’s Astra uses a technique called opaque recurrence that loops internally instead of emitting sequential reasoning steps, and safety researchers told TechCrunch that scaling it could destroy the chain-of-thought monitoring the same labs keep citing as a safeguard.
- Liability: thirty new complaints brought the total to 37 over the February school shooting in Tumbler Ridge, British Columbia, escalating from negligence to aiding and abetting.
- Security money: HiddenLayer took $100 million led by Delta-v Capital on more than 10x ARR growth, and Lasso Security raised $30 million alongside a guardrail engine that runs on ordinary CPUs in under 5 milliseconds.
- Harness, not model: a Show HN benchmark held the model constant across nine agent harnesses over 360 trials and found Claude Code leading on pass rate at 63.3 percent and $18.34 per task while another harness ran $1.05, with pass rate and cost-effectiveness close to uncorrelated.
- Provenance: Anthropic launched a browser tool at claude.com that reads C2PA credentials to say whether a file was made with Claude, running entirely client-side. Only the company’s own page documents it, and a screenshot or format conversion strips the credential.
- Refactoring: the sharpest critical essay of the week argues that human confusion used to be the trigger that forced teams to refactor, and agents never get lost, so the signal never fires.
- Claimed, not confirmed: Elon Musk posted that Grok 4.7 ships around September 12 at 2.1 trillion parameters and will beat every available model. There are no independent benchmarks, and xAI’s developer documentation still lists 4.6 as current.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
