AI law and the newsroom, September 30–October 1: a US appeals court unseals its AI fair-use reasoning, California starts paying for journalists, and Connecticut's content-marking duties begin

This issue covers 30 September and 1 October 2026: four items from three jurisdictions. A US appeals court published the reasoning behind its AI-training ruling. California's governor signed a tax credit that pays newsrooms for the journalists they employ. Connecticut's content-marking duties for generative AI took effect. And China's internet regulator opened a two-month campaign against fake personas in short video.

Three questions run through the items below, and this digest keeps 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 Third Circuit filed its opinion on 29 September under seal. It became public on Wednesday 30 September, and it says why the court rejected the fair-use defence that ROSS Intelligence raised for copying Westlaw's headnotes — the editorial summaries of points of law — to build a competing legal search engine. 12

"Thomson Reuters' materials possess a creative spark, and ROSS aspired to be a direct competitor by using them for a highly similar purpose," Judge Tamika Montgomery-Reeves wrote for a unanimous three-judge panel. The court affirmed the District of Delaware on the fair-use question and on ROSS's argument that headnotes are not copyrightable. 1

The panel narrowed its own holding in a footnote. ROSS's search engine did not use generative AI, and the court said the concerns the US Department of Justice raised in its copyright case against OpenAI "do not apply here". 1 "Under ROSS's framing, this case appears to concern the future of AI legal technology," Judge Montgomery-Reeves wrote. "But appearances can be deceiving. In truth, this is no more than an ordinary copyright case."

For a newsroom, the reasoning is the part that travels. A federal appeals court has now held that copying a publisher's editorial product into a rival product is not fair use, and it has left open what training a generative model on the same material counts as. The case is Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 25-2153. 1 Thomson Reuters, which filed the 2020 suit, is the parent company of Reuters News. 1

Gov. Gavin Newsom signed AB 2222 on Wednesday, creating refundable tax credits for California news organisations based on how many journalists they employ. 3

A qualifying organisation can claim $20,000 for each of its first five full-time journalists and $15,000 for each one after that, $7,500 for each part-time journalist, and a further $15,000 for each full-time journalist in a position that is new. The retention and new-hire credits stack, so a newsroom that adds a reporter can claim both for that position in the same year. Because the credit is refundable, an organisation that owes less tax than the credit is worth is paid the balance from the state's Tax Relief and Refund Account or its Corporation Tax Fund. 4

The credit applies to taxable years beginning on or after 1 January 2027 and before 1 January 2032. 3

Print publications, digital outlets and broadcast stations all qualify, whether they are for-profit, nonprofit or sole proprietorships. Each must have been organised or registered in California for at least 12 months before the tax year, disclose its beneficial owners or its board of directors, publish an error-correction policy with a route for the public to complain, carry media liability insurance, and not be controlled by a political action committee or a 501(c)(4). Digital outlets must publish at least monthly and show that at least a third of their audience is in California. Print outlets need court-adjudicated general-circulation status, a USPS periodicals privilege, or the same audience test; broadcasters need an FCC licence to a California community of licence, or public-broadcaster status. 5

A qualifying journalist has to be a California resident whose duties are in the state. Full-time means at least 30 hours a week for more than 26 weeks of the year and at least $35,000 a year; part-time means 20 to 30 hours a week over the same period and at least $25,000. Freelancers are outside the scheme, and an organisation cannot stack this credit with another California hiring credit for the same journalist. 5

California's Department of Finance opposed the bill because it caps nothing, which the department called unlimited fiscal liability for the state, and business groups objected to the tax increase that pays for it: the bill ends the state's corporate deduction for executive pay above $1 million a year, matching federal law. 6 Rebuild Local News, which sponsored the bill, projects more than $200 million reaching newsrooms over five years and says the first money may not arrive until 2028, because the Franchise Tax Board still has to build a route for nonprofits that owe no state tax. 5

Newsom signed it anyway, and said what he thought of its reach: "It does subsidize those that don't need to be subsidized. We have hedge funds in this space. We have billionaires in this space." He said he hoped the Legislature would narrow who benefits. 6

Connecticut's AI statute started binding people on 1 October. The Connecticut Artificial Intelligence Responsibility and Transparency Act, Public Act 26-15, was signed on 27 May 2026 with staged dates, and the first set has arrived. 7

The part that reaches published content is the provenance rule. A provider whose generative AI system has more than a million monthly users and is available to consumers for personal use — the Act's "covered provider" — must, so far as is commercially and technically reasonable, embed provenance data in any audio, image or video its system creates or materially alters. The provider must also use reasonable methods to make that data hard to remove or tamper with, and embed it so a consumer can judge whether the content is synthetic. Government agencies, business-to-business systems, video games and some interactive experiences sit outside the definition. 8

The same date carries three other duties. A frontier developer — one using more than 10^26 operations of computing power to train a model — may not retaliate against an employee who reports a catastrophic risk, and must tell its employees about that protection. A company selling an AI subscription must disclose the key terms in writing and take the consumer's written acceptance before charging or renewing. And the state now has statutory definitions of what counts as AI. 89

Later dates on the same statute: large frontier developers must run an anonymous internal reporting process by 1 January 2027; companion chatbots must say they are not human from 1 January 2027; employers using automated tools in employment decisions must have a framework for disclosing that use by 1 October 2027; the social media rules start on 1 January 2028. The attorney general enforces all of it, as unfair or deceptive trade practices. 810

For a newsroom, the duty sits with the model provider rather than with the desk using its output. What changes is the label stream: synthetic audio, image and video from a large generative provider shown to a Connecticut audience is now supposed to carry provenance data that a verification desk, or a reader, can look for.

China's internet regulator said on 30 September that it had issued a notice launching a two-month national campaign against fake personas in short video. 11

One of the five categories the campaign targets is impersonation using AI: synthesised audio and video deployed to pose as public figures for fake endorsements and marketing, accounts posing as people caught up in a news event or as their relatives, and "shadow accounts" that repost a public figure's material in a way that implies authorisation. The other four are posing as members of vulnerable groups, inventing professional credentials, promoting luxury and superstition, and fabricating conflict plots. 11

The labelling instruction inside the notice is the part to read twice. Videos in these categories that lack the required label are to be handled strictly, and platform operators must not let a content label — fictional performance, AI-generated, marketing, repost, personal opinion — become a way around the regulator's content requirements. The notice also tells platforms to keep such videos out of rankings and recommendation slots. 11

For a newsroom publishing short video for a Chinese audience, the label is part of the publishing duty. The notice treats it as the floor, and expects platforms to keep labelled videos that carry false personas off recommendations and out of rankings.

The nearest one is California's. SB 1000, AB 2713 and AB 1609 were still on the governor's desk as this issue went out, with a deadline of midnight in Sacramento on 30 September. The legislative record showed SB 1000 enrolled and awaiting his action, and no signature or veto on them appeared in the bill actions his office published during the day. 12

After that: 3 November 2026, when the Commission's copyright consultation closes. Then 2 December 2026, when the EU's prohibitions on AI systems that generate sexual abuse material and non-consensual intimate images start to apply. Then 1 January 2027, when Connecticut's chatbot disclosure duty begins, California's AI package takes effect if the governor lets it through, and the first tax year the California newsroom credit covers begins.