Three weeks ago the visible half of the largest artificial intelligence copyright case in the United States belonged to the defendants. OpenAI and Microsoft moved for summary judgment on September 4, 2026, and the publishers' answering papers arrived the same day with the traffic figures and an entire category of alleged copying blacked out. What changed on September 17 is that the publishers refiled their brief with most of those redactions lifted. The numbers underneath were Microsoft's own.

That is the thread running through an unusually legal weekend. A court record turned a company's internal measurement into a public one. A federal appeals court told a different set of plaintiffs that the statute they had chosen does not reach what they were complaining about. A district judge in Los Angeles declined to let a settlement decide a question he considers his own. And a tribunal in London approved a payment that ends a trial without ever establishing what the trial was about. In every instance the question is the same: which institution gets to say what happened, and on what evidence.

Microsoft's internal figures put a number on the substitution argument

The combined brief was filed on September 17, 2026 at docket 1977-1 in multidistrict litigation 25-md-3143, before Judge Sidney H. Stein in the Southern District of New York. The plaintiffs are The New York Times Company, eight Daily News titles including the Chicago Tribune, the Orlando Sentinel, the Denver Post and the Mercury News, Ziff Davis, the Center for Investigative Reporting and The Intercept. The defendants are OpenAI and Microsoft. The document that became public last week is the version with substantially fewer redactions, and the material it uncovers is internal measurement rather than expert reconstruction.

Microsoft compared click-through rates on Bing web search against click-through rates in its Bing Chat product, domain by domain. For New York Times properties, clicks fell between 87 and 93 percent. For Daily News domains the range was 83 to 91 percent. For Ziff Davis properties it ran from 51 to 94 percent. These are not modelled estimates published by a measurement vendor with an interest in the answer. They are a platform's own instrumentation of its own products, produced in discovery.

Alongside them sit behavioural figures. The brief states that 87.78 percent of ChatGPT users visit no external website during a search, against 26.91 percent for Google. A Microsoft survey found 37.2 percent of Copilot respondents would otherwise have used conventional web search. Among New York Times subscribers who use ChatGPT for news, 36 percent said the product means they no longer need news from the Times at all. Cloudflare data cited in the filing puts OpenAI's crawl-to-referral ratio at 1,500 pages scraped per visitor sent, against roughly two to one for Google in 2015 and eighteen to one by June 2025.

The internal communications are what turn the numbers into an argument about fair use. Nick Turley, who runs ChatGPT, described publishers as facing an existential threat and wrote that the products are largely substitutive, period, and will get more and more substitutive. Satya Nadella testified under oath that chatbots had substituted giving you the information right there on the website on the AI platform versus needing to go to the underlying source, and that anything that is paywalled should be licensed by anyone who wants to use it. Brent Hecht, a director of applied science at Microsoft, wrote that companies hoovering up content would be regarded as an astonishing theft of unprecedented proportions, and that a fair use victory would make a complete mockery of the idea of fair use. Another Microsoft document warns that the company's AI content strategy has started a doom loop that will hurt the performance of our models and the entire web at the same time.

Substitution is the fourth statutory fair use factor. Alan Chapell, a privacy specialist quoted in Digiday's account of the unsealing, put the point plainly: an admission that the products are substitutive likely eviscerates the fair use defence. David Buttle of the publisher licensing coalition Spur noted in the same piece that OpenAI, Microsoft, Amazon, Google, Meta, Perplexity and Mistral have all signed publisher licensing deals covering the same category of content, which is itself evidence that a market exists. Danielle Coffey of the News/Media Alliance expects a ruling that AI companies must pay to push publishers toward collective negotiation.

The scale allegations are separate from the traffic ones. The brief alleges 3,924,653 copies of Times material covering 944,655 unique works, and 7,387,394 copies across 1,388,384 unique works for the Daily News plaintiffs. WebText and WebText2 are said to contain 6,552 Times articles, 18,609 Daily News articles and 66,780 Ziff Davis articles; Common Crawl archives are said to have supplied 2,182,079 Chicago Tribune articles and 2,064,805 Times articles. The New York Times Annotated Corpus, 1.8 million articles from 1987 to 2007, was obtained from the Linguistic Data Consortium under a licence restricting use to non-commercial linguistic education, research and technology development. Under Project Mango, a Microsoft crawler is alleged to have collected 160,903 unique publisher works for OpenAI's training.

Then there is grounding, which matters more for advertising than training does, because it describes live retrieval rather than a historical copy. The brief says at least two million Copilot conversations copied and grounded outputs on publisher content. Microsoft's own expert, Dr John Lafferty, conceded 88,178 instances in which Microsoft copied at least six non-overlapping sixteen-word sequences from plaintiff works. One retrieval is said to have grounded on 1,820 words from a single article. In another, Copilot retrieved 340 words from a Times piece on presidential power and returned 270 of them to the user; in a third it retrieved roughly 340 words from a Denver Post investigation and returned nearly all of them verbatim.

OpenAI's counter-evidence is a rate. Its expert reviewed 20 million ChatGPT conversation logs and found 24 instances of verbatim regurgitation, which works out at 0.00012 percent. The company also argues that copies made before robots.txt files were updated were impliedly licensed, and that its text extractors stripped copyright notices as a side effect of cleaning markup rather than as an act of removal. Both defendants moved for summary judgment on September 4. No hearing date has been set, and a pending sanctions motion has to be resolved before the grounding and output claims against OpenAI can be decided.

One detail deserves its own sentence. After the suits were filed, OpenAI built what it called the Giraffe Bloom Filter to suppress output of content the plaintiffs had identified, covering 140,097 Times copies and 270,250 Daily News copies. Hecht, at Microsoft, called the approach an accidental cover up, on the grounds that it left rights holders with less visibility into what was used for training. Suppressing the evidence of memorisation is not the same as undoing the memorisation, and the brief uses the filter as proof that the underlying copies exist.

The Ninth Circuit tells programmers the wrong statute was chosen

Two days before the publishers refiled, the Ninth Circuit affirmed dismissal of a parallel set of claims against GitHub, Microsoft and OpenAI. Case number 24-7700 was decided on September 16, 2026 by a panel of Judges Sidney R. Thomas, Eric D. Miller and Stanley Blumenfeld Jr. Miller wrote the eighteen-page opinion. The underlying action, docket 4:22-cv-06823-JST before Judge Jon S. Tigar in the Northern District of California, was brought by anonymous programmers who published code under open-source licences requiring attribution, and who argued that GitHub Copilot and OpenAI's Codex reproduced their work without the notices those licences require.

They pleaded Section 1202(b) of the Digital Millennium Copyright Act, which prohibits removing or altering copyright management information. The panel held that a generative system does not remove anything. Copilot infers statistical patterns and produces new work through probabilistic processes, and, in the opinion's formulation, one who creates a new work and fails to include CMI cannot be said to have removed or altered anything.

The plaintiffs had a second theory available and lost it procedurally. Their attorneys conceded at early hearings that the complaint was not about training, so the appeals court declined to decide whether stripping attribution during data preparation violates the statute. That is the precise theory the news publishers are running in New York, and it remains open.

The panel was careful about what it did not hold. Standing survived: GitHub's own duplicate-detection filter, which flags verbatim snippets of 150 characters or more, was treated as evidence that reproduction happens, alongside examples of Copilot emitting the plaintiffs' code. Nor did the court require literal identity between works. Where two works share the same composition and one lacks copyright management information, the opinion says, infringement might be inferred, and cosmetic changes will not shield a defendant. The distinction is between a stripped copy and a generated output, not between an exact match and an approximate one.

The money explains why the claim was framed as it was. Section 1203(c)(3) sets damages at up to $25,000 per violation, while Section 504(c)(1) caps statutory copyright damages at $30,000 per work. A per-violation measure multiplies across every output, which is how the plaintiffs reached a figure above $9 billion, though that number appears nowhere in the appellate opinion. Two contract claims resting on the open-source licence terms survive and remain before Judge Tigar, which leaves the attribution requirement to be tested as a breach of licence rather than as a statutory violation.

Read the two filings together and the boundary becomes visible. The publishers allege that copyright notices were absent from 99.9 percent of Times copies, 99.8 percent of Ziff Davis copies, 99.99 percent of Mother Jones copies and 100 percent of Intercept copies, with at least 3,002,796 Times copies, 2,638,497 Daily News copies and 3,462,149 Ziff Davis copies stripped by extractors named Dragnet, Newspaper and Gutentag. That is a claim about the input pipeline. The Ninth Circuit's holding is about the output. For anyone assembling an AI product on third-party text, the operative question after September 16 is no longer whether the model's answers carry a notice, but what happened to the notice on the copy the model was built from.

Judge George H. Wu issued a sixteen-page tentative ruling on September 18 indicating that he would deny the Justice Department's motion to vacate the 2019 child privacy consent order entered against Musical.ly, the service that became TikTok. The hearing is scheduled for 8.30 a.m. on September 21 in the Central District of California, case 2:19-cv-01439-GW-RAO.

The structure of the deal is what drew the court's attention. The Justice Department announced a $400 million settlement on August 21, 2026 in the separate COPPA action it brought against ByteDance and the TikTok entities in August 2024. Its own Office of Public Affairs described the terms as $300 million payable immediately and a further $100 million upon entry of an order vacating a prior consent decree. The prior decree is the March 27, 2019 order entered by Judge Otis D. Wright II alongside a $5.7 million penalty. In other words, a quarter of the settlement was contingent on a judge agreeing to dissolve an injunction he had not been asked to review on the merits.

Wu's tentative answer was that the condition is the government's to write and the decision is his to make. The government may condition its settlement on requesting vacatur, the ruling states, but that condition cannot displace the Court's independent obligation to decide whether vacatur is warranted. He also rejected the argument that an unopposed motion earns deference, citing Rufo v. Inmates of Suffolk County Jail for the proposition that no deference is involved in the threshold question of whether modification is appropriate, and distinguishing the Ninth Circuit's 1984 decision in SEC v. Randolph.

The ownership argument fared no better. TikTok's United States business was restructured in January 2026 into TikTok US Joint Venture LLC, with ByteDance retaining 19.9 percent, following President Trump's approval of a divestiture plan by Executive Order 14352 in September 2025 and the Supreme Court's decision upholding the divestiture deadline on January 19, 2025. Wu found that COPPA applies without regard to whether the relevant operator is foreign- or domestically owned, and that the 2019 order certainly considered the possibility of a change in corporate structure or ownership. On age-gating, he separated capability from compliance: a promise to comply, he wrote, citing a July 31, 2026 decision in United States v. Lakeland Bank, is not the same as having demonstrated actual compliance. A declaration from TikTok US employee Camerin Hunt describing current controls did not close that gap.

The ruling also notes an ambiguity the parties left in their own papers. The government wrote that TikTok US paid a significant sum of money as part of the settlement, which the court found unclear as to whether payment had been completed or remained conditional. The denial is without prejudice. The government can return with the settlement agreement itself, with evidence of actual compliance, or with a narrower request. What it cannot do, on this reading, is treat the consent order as an asset the parties may trade between themselves.

London approves 260 million pounds and the ten-week trial disappears

The contrast arrived from the Competition Appeal Tribunal, which approved a collective settlement of 260 million pounds on September 16 between Professor Barry Rodger of the University of Strathclyde, as class representative, and seven Alphabet and Google entities. Case 1673/7/7/24 was twelve days from a ten-week trial listed to begin on September 28, 2026. That trial will not happen, and the tribunal made no finding on whether Google's Play Store commission structure breached competition law. Recital C of the agreement records that Google denies the conduct alleged and believes it has strong defences; clause 10 specifies that entering into the agreement is not an admission of liability or wrongdoing by any party, and that neither the agreement nor any supporting document may be relied on as evidence of the truth of any allegation.

The money splits in two. A class pot of 160 million pounds goes to eligible United Kingdom app developers for the period from August 22, 2018 to July 31, 2026, with developers whose Play Store revenue was 7,500 pounds or less able to elect a fixed 200 pound payment instead of an individual calculation. A stakeholder pot of 100 million pounds covers the litigation funder, the after-the-event insurers, the solicitors and counsel. The application states that 68.9 percent of the settlement sum, net of costs incurred, is available to represented persons.

The stakeholder pot does not cover the stakeholders. As of July 31, 2026, claimed entitlements stood at roughly 115 million pounds against the 100 million available, a shortfall near 15 percent, and projected to completion the gap widens to about 34 million. The funder, Bench Walk Guernsey PCC Limited, contracted under a December 6, 2023 agreement amended four times, had a maximum outlay of 27,671,000 pounds and a tiered profit share: 45,297,150.59 pounds plus 400 percent of capital outlay if proceeds arrived before the liability trial, or 81,894,301.17 pounds plus 600 percent if they arrived after. Settling before September 28 kept the claim in the lower tier. Bench Walk agreed not to insist on its contractual rights where doing so would jeopardise settlement approval, which is how the arithmetic was reconciled.

Rodger faced real costs on the other side of the ledger: an unbudgeted 3.35 million pounds through trial, more than 750,000 for an appeal, more than a million for a possible second hearing on intra-group arrangements, and more than a million for disclosure compliance with a deadline extended to October 2. The precedent hanging over the decision is Kent v Apple, which went to trial on January 13, 2025 and produced judgment in late October 2025; more than ten months after that judgment, and nineteen months after the trial opened, the class had received nothing while Apple's appeal waits for a May 2027 hearing.

Distribution here will take its own time. Angeion Group, appointed administrator, opens the claim period six to ten weeks after approval, runs it for four months, reviews claims over thirty to sixty days, and pays four to six weeks after that, with no payment issuing until a further tribunal order. Angeion projects take-up at 50 to 80 percent by value and 8 to 15 percent by number, on the reasoning that value concentrates in a small number of large developers and that Geradin Partners has already engaged developers representing about 80 percent of the claim. American comparators support the split: Cameron v Apple distributed $100 million to 67,440 developers with take-up of 13.21 percent by account and 32.10 percent by value, and the Google Play developer settlement in the United States ran to $90 million across 47,972 accounts. Both capped eligibility by revenue. The United Kingdom class does not.

Nothing in the commercial arrangements changes. Commission rates, billing rules and distribution terms are untouched, and the regulatory track runs on separately: the European Commission's 890 million euro fine of July 23, 2026 included 430 million for Play steering restrictions, the Court of Justice dismissed Google's appeal against the 4.125 billion euro Android fine on July 2, 2026, and the Competition and Markets Authority designated Google with Strategic Market Status on September 30, 2025. A consumer claim brought by Helen Elizabeth Coll, case 1408/7/7/21, is still listed for trial on October 5. And on August 5, 2026 the tribunal certified a separate opt-out claim of roughly 5 billion pounds on behalf of United Kingdom advertisers who bought Google search ads between 2011 and 2025, which is the proceeding the advertising industry will be watching next.

Pinterest puts its ad load behind the camera

Away from the courts, Pinterest introduced a set of advertising products on September 21 that all point at the same behaviour: searching without words. Visual Search Ads place sponsored products inside search results and pin close-ups, so a query such as statement handbags returns shoppable placements that route to merchant sites. Restyle, a camera function built on Pinterest Intelligence, lets a user photograph a room, render it in a different aesthetic, and buy the products that produce the look. Performance+, the automated suite that entered beta in 2024, gains manual A/B testing and controls for prioritising particular products during seasonal launches, along with new-customer and app-install objectives. And following the acquisition of tvScientific in February 2026, connected television campaigns can now be bought as a standalone line, with the Roku programme Bring My Pinterest to Life returning for a second season in 2027.

The platform figures behind the pitch: 640 million monthly users, 80 billion monthly searches, 96 percent of them unbranded, paid clicks up fivefold over three years, second-quarter sales up 18 percent to $998.2 million, users up 11 percent year on year, and more than half of the base in Generation Z. Chief executive Bill Ready said Pinterest is now a performance platform. Chief business officer Lee Brown framed the position as arriving before the decision is made. Chloe Wix, vice president of go-to-market, described the tools as a step toward measurable lower-funnel outcomes using campaign assets brands already hold.

The unbranded share is the number that matters commercially. When 96 percent of searches name no brand, the query itself carries no attribution signal, and whatever ranks becomes the recommendation. That is the same structural fact the New York filing describes from the other end: a surface where an intermediary's answer stands between intent and destination. In one case the intermediary is being asked in court what it owes the sources it summarises. In the other, it is selling the position outright.

Four documents, four jurisdictions, one recurring difficulty. Microsoft measured the effect of its own product and the measurement became evidence. The Ninth Circuit held that a statute written for stripped copies does not describe generated text. A judge in Los Angeles refused to let $100 million decide a question of compliance. A tribunal in London accepted 260 million pounds to close a case that will now never establish what happened. Evidence and settlement are not substitutes, and the week made the difference legible: 94 percent is a finding, and 260 million pounds is a price.

Also noted

  • September 20 - Spain's competition authority published a 126-page study of infrastructure and platform cloud services, finding that only 2.3 percent of sampled business customers switched provider between 2022 and 2024, that the top three providers hold 50 to 70 percent of combined revenue on a market worth about 1.9 billion euros in 2024, and that Spanish business cloud adoption of 44.3 percent ranks 22nd of 28 European countries. PPC Land
  • September 20 - Bannerflow analysed 2.33 million creative assets, 155.1 billion impressions and 166.8 million clicks from more than 300 brands between January 1 and August 31, 2026, finding 78 percent produced video advertisements while only 46 percent served a video impression, with video accounting for 10.5 percent of assets against a blended click-through rate of 0.108 percent. PPC Land
  • September 20 - Governor Gavin Newsom signed Executive Order N-9-26 on September 18, giving California's Government Operations Agency until November 16, 2026 to recommend whether frontier AI developers should host independent verification organisations onsite, submit safety frameworks to third-party review, build a kill switch subject to ongoing efficacy checks, and report loss-of-control incidents. PPC Land
  • September 20 - Commission Implementing Regulation (EU) 2025/1960 takes effect on September 27, requiring sellers across the European Union to display a harmonised notice explaining the two-year legal guarantee on product pages, in checkout flows and in order confirmation emails, with an optional GARAN label for producers offering longer durability coverage at no extra cost. PPC Land
  • September 20 - Google Search Console's crawl stats report is missing September 15 across all profiles, a gap that has recurred repeatedly since November 2021 and has always been backfilled, with no statement from Google this time. Search Engine Roundtable