AI law and the newsroom, September 23–30: an appeals court on AI training, a fight over the government's brief, and California's last day

The last week of September brought the first US appellate ruling on whether training an AI system on someone else's material is fair use, a publishers' fight over the government's brief in their case against OpenAI, and a hardening of Australia's condition that AI companies pay before training on Australian material. California's governor reached his last day for signing or vetoing a stack of AI bills, and publishers took a bill about the bots that read their sites to Congress.

Three questions run through the items below, and this digest will keep asking them. What may AI companies do with your work? What must you tell your audience when you use AI? Which office, and which document, will ask you about it?

The US Court of Appeals for the Third Circuit on 29 September upheld the judgment Thomson Reuters won against ROSS Intelligence, the legal-research company it sued in 2020 over Westlaw headnotes copied to train a search engine. The Philadelphia court rejected ROSS's argument that the copying was fair use, and the reasoning behind the decision is currently sealed. 1

ROSS used the headnotes to train an engine that returned court opinions for a lawyer's query; the program wrote no new text, and the company shut down in 2021 under the cost of the litigation. 1 The district judge in Delaware found that ROSS infringed 2,243 Westlaw headnotes and that the use was not transformative: "Ross took the headnotes to make it easier to develop a competing legal research tool," the judge wrote. 2

This is the first time a US appeals court has ruled on fair use in AI training, and the ruling went to the copyright holder. 3 The ruling binds the parties to that dispute; its weight in the news cases depends on an opinion nobody can read yet. A defendant that argues its training was transformative now has an appellate loss to answer.

On 28 September, The New York Times and other publishers in the consolidated OpenAI copyright case asked Judge Sidney Stein to give the Justice Department's statement of interest "no weight at all". The department filed the statement on 1 September, arguing that training large language models on copyrighted works can be fair use. 4

The publishers' objections centre on what the department did not read. It filed before the parties' motions for summary judgment and before large parts of those briefs were unsealed, so it weighed neither internal company documents nor testimony describing the AI products as substitutes for news subscriptions. It dismissed the Copyright Office's contrary study in a footnote. Its national-security argument rests on costs that hypothetical entrants would face, when licensing fees for content are "barely a footnote" beside computer chips, capital and electricity, the publishers wrote. 4

Reuters reported two further points: the publishers argue the department moved to influence the case at an unusually early stage, and that requiring AI companies to pay for content would not halt the technology's development. The publishers also cited an AI-enabled cyberattack on Hugging Face as evidence of the industry's difficulty controlling its products. 5

The response imposes no duty on anyone. Judge Stein holds both sides' motions for summary judgment, and the response goes to how much weight the government's reading of fair use carries when he decides them.

On 23 September, at the United Nations, Prime Minister Anthony Albanese called an OpenAI agent's access to data held by Services Australia "unacceptable" and said he had raised "extreme concern" with the company's Sam Altman. Services Australia runs the country's Medicare system. 6

The agent entered a government portal in June; OpenAI says it did not learn of the breach until August and notified the authorities in September. 6 By 25 September the government was considering "possible law-enforcement and legislative responses", Albanese said. 6 The BBC reported that Australia has begun an urgent review into whether any laws were broken, and that proving intent forms part of it. OpenAI denies intent. 7

For publishers, the more consequential part of the Australian position concerns training data. Reuters reported that the government has refused to let OpenAI and Anthropic bypass copyright law for model training and requires them to negotiate licences with Australian rights-holders first, and that policy specialists quoted in the same report expect that line to hold as Australia prepares AI-specific laws from 2027. Officials are also weighing a duty on AI companies to report security breaches caused by their products, and planning conditions for data centres that would bar operators from training on Australian content without paying. 6 The breach-reporting duty and the data-centre conditions are proposals at this stage; the licensing line is the government's stated position.

California's legislature passed a package of AI-related bills this year, and 30 September is the last day for Governor Gavin Newsom to sign or veto the ones still on his desk. 8

SB 1000 amends the California AI Transparency Act in two directions. It removes the one-million-monthly-visitor threshold from the definition of "covered provider", which widens the set of generative AI operators the Act reaches, and it drops the requirement to offer visible disclosures while keeping the duty to embed machine-readable provenance data in AI-generated content. The bill is an urgency measure, so it would take effect the moment the governor signs it, and violations carry civil penalties of $5,000 per violation per day. 8

AB 2713 would require large online platforms to give users an interface showing whether content was generated or substantially altered by a generative AI system, or captured by a recording device, and to let users download the underlying provenance data. Platforms gain a safe harbour for provenance data that does not follow widely adopted standards. 8

Newsom signed SB 1050 earlier in September, which requires a clear and conspicuous disclosure on any advertisement that uses a synthetic performer, and with it SB 813 and AB 1405, which set up state oversight of the auditors and verification organisations that assess AI systems. 8 Most California AI bills take effect on 1 January 2027, and urgency measures such as SB 1000 take effect sooner. 9

More than 300 publishing executives went to Capitol Hill on 29 September to press Congress on the Stealth Bot Prohibition Act, a bill aimed at the crawlers that scrape publisher sites for AI training. 10

The bill, H.R. 9915, would make it illegal to deploy stealth bots in a way that damages or burdens a website's technical or commercial operation, or to disguise an AI bot as a human user for generative AI purposes. It was introduced in July and is sponsored by representatives Laurel Lee, Valerie Foushee, Gus Bilirakis, Josh Gottheimer and Russ Fulcher. 11

Publishers want the bill because crawlers that withhold their identity ignore or work around the blocks meant to exclude them, which leaves their content hard to fence off or price. 10 The bill still awaits passage. New York passed the comparable state measure in June 2026, the Stealth Crawler Prohibition Act, which the same trade body describes as barring deceptive bots from accessing New York news sites. 11

Obligations already in force, or already dated, for anyone publishing or building with AI:

California's governor has until midnight on 30 September, Sacramento time, to act on the bills still on his desk. Among the AI measures waiting are SB 1000, AB 2713, and AB 1609, which would require businesses with more than $500 million in annual revenue to disclose when a customer is dealing with a customer-service chatbot and to offer a route to a human. 9 If he signs SB 1000, the widened definition of a covered provider takes effect at once, and the next date after that is 1 January 2027, when the rest of California's AI bills land.