USA Today's publisher sues OpenAI for more than $250 million over the local-news summarising pattern

FACT (CONFIRMED — primary complaint, ï¶1–14 and caption). On Thursday 2026-10-08, 14 plaintiffs filed a copyright action in the Southern District of New York:

• USA TODAY Co., Inc.; Gannett Satellite Information Network, LLC; Gannett GP Media, Inc.; The Courier-Journal, Inc.; Des Moines Register and Tribune Company; Detroit Free Press, Inc.; Detroit Newspaper Partnership, L.P.; CA Florida Holdings, LLC; Scripps NP Operating, LLC; CA North Carolina Holdings, Inc.; GateHouse Media Oklahoma Holdings, Inc.; Journal Sentinel Inc.; GateHouse Media Ohio Holdings II, Inc.; and Phoenix Newspapers, Inc.
• All are described as owned by USA TODAY Co., Inc.; all list a principal place of business in Pittsford, New York; they are collectively the "USA TODAY Plaintiffs".
• Counsel: Rothwell, Figg, Ernst & Manbeck, P.C. (Steven Lieberman and Jennifer Maisel are identified in docket and trade reporting).
• Defendants: OpenAI Foundation; OpenAI GP, LLC; OAI International, Inc.; OpenAI OpCo, LLC; OpenAI Global, LLC; OAI Corporation; and OpenAI Group PBC — seven OpenAI entities, tracking OpenAI's October 2025 recapitalisation. Microsoft is not a defendant, unlike the New York Times and Seattle Times actions.

Together the plaintiffs own copyrights in content from 19 publications (ï¶2): USA TODAY, The Tennessean, Indy Star, The Bergen Record, The Enquirer, Asbury Park Press, Democrat & Chronicle, The Knoxville News-Sentinel, Naples Daily News, The Oklahoman, Milwaukee Journal Sentinel, The Columbus Dispatch, The Arizona Republic, The Courier-Journal, The Des Moines Register, Detroit Free Press, The Detroit News, The Palm Beach Post and Star News.

FACT (CONFIRMED — complaint ï¶13 and prayer for relief). The complaint states: "In this lawsuit, the USA TODAY Plaintiffs seek damages in excess of $250 million." It adds that the law allows "up to $150,000 for each willful copyright infringement, plus up to $25,000 per violation for OpenAI's stripping of copyright management information". The prayer for relief asks for: a declaration that past, present and future use is unlawful; a permanent injunction; destruction under 17 U.S.C. § 503(b) of all GPT or other LLM models and training sets incorporating their content; and costs, expenses and attorneys' fees.

FACT (CONFIRMED — complaint, counts I–III). Three claims are pleaded:

1. Count I — Copyright infringement, 17 U.S.C. § 501. Amassing datasets containing the plaintiffs' works; reproducing them during pre-training, training and fine-tuning; storing memorised copies in the models; exchanging copies with Microsoft (the "Project Taxi" Bing Index transfer and the "Project Mango" jointly operated crawler, for which the complaint says OpenAI paid Microsoft); and producing allegedly infringing outputs.
2. Count II — Vicarious copyright infringement. The parent/holding entities (OpenAI Foundation/OpenAI, Inc.; OpenAI GP, LLC; OAI Corporation; OpenAI Group PBC; OpenAI Global, LLC) directed, controlled and profited from infringement carried out by OpenAI OpCo, LLC and OAI International, Inc.
3. Count III — DMCA, removal of copyright management information, 17 U.S.C. § 1202. Use of the "Dragnet", "Newspaper" and "Gutentag" content extractors, which the complaint says target removal of CMI-type terms such as "byline", "copyright" and "©".

FACT (CONFIRMED — complaint ï¶101–106). The complaint quantifies the alleged training-data footprint:

• WebText (the GPT-2 corpus): the plaintiffs' content "comprises more than 160,000 entries in the WebText dataset, including 83,266 entries from usatoday.com, 12,994 from freep.com, 9,814 from jsonline.com, 9,073 from azcentral.com, 7,028 from detroitnews.com, 6,726 from indystar.com, 6,194 from cincinnati.com, 6,082 from dispatch.com, 5,396 from tennessean.com …".
• C4 (a filtered English-language subset of a 2019 Common Crawl snapshot): the plaintiffs' domains "account for over 122 million tokens … including 23M tokens from usatoday.com, 12M from azcentral.com, 8.2M from freep.com, 8.1M from jsonline.com, 8.1M from detroitnews.com, 7.8M from cincinnati.com, 7.6M from indystar.com, 6.5M from tennessean.com …".
• WebText2: "weighted 22% in the training mix for GPT-3 despite constituting less than 4% of the total tokens in the training mix" — the basis for the argument that journalism was treated as high-value, deliberately favoured material rather than incidental filler.

FACT (CONFIRMED — complaint ï¶8–12, 160, 182). The complaint quotes OpenAI internal material obtained in the consolidated MDL discovery (cited as In re OpenAI, Inc. Copyright Infringement Litigation, No. 25-md-3143, Dkt. No. 1977-1, S.D.N.Y. Sept. 17, 2025):

• OpenAI's Head of ChatGPT, Nick Turley, wrote that publishers face an "existential threat" from OpenAI's products, which "are largely substitutive, period" and "will get more and more substitutive as they get better".
• An unnamed OpenAI software engineer wrote: "no matter how prominently we show the links, users won't click."
• OpenAI Policy Director Jack Clark wrote that OpenAI's AI work will "increasingly lead to us creating systems that substitute for the labor of the people that define the 'culture' of society".
• Internal documents described ChatGPT as "[t]he modern newsstand".
• OpenAI's VP of Research: "We train our networks to memorize the training data — that's their objective."
• OpenAI co-founder Greg Brockman described the models as "excellent at news" and "very good at any news task".
• A Microsoft executive who raised the issue internally called the output-filter approach an "accidental cover up", because it left "people who have a right over the content having less visibility into what was used for training".

FACT (CONFIRMED — complaint, Exhibit B described in ï¶160s and in reporting). The plaintiffs attached worked examples: for each of the 19 titles they prompted GPT-5.6 to "find and summarize the article with this title and give me an in-depth summary", and allege the model returned summaries that followed the original's structure and reproduced its substance (e.g. the Indianapolis Star FBI-raid story, a Democrat & Chronicle ranked snowstorm list, a Tennessean chronological timeline).

INDEPENDENT EVIDENCE (CONFIRMED). Reuters broke the story on 2026-10-08 and noted OpenAI did not immediately respond to a comment request; The Verge reported the same day that USA TODAY Co. "asks for damages of more than $250 million" and that OpenAI "didn't immediately respond". Editor & Publisher, The Desk and Press Gazette all record the filing date, the 19 publications and the $250 million figure. The Justia docket confirms 14 plaintiffs, 7 OpenAI defendants, copyright cause of action under 17 U.S.C. § 501, jury demanded by plaintiff, and Filing 3 — a STATEMENT OF RELATEDNESS asking that the action be filed as related to 1:25-md-03143-SHS-OTW.

COMPANY CLAIM (OpenAI's standing position, not a response to this suit). In "OpenAI and journalism" (2024-01-08) OpenAI states: "Training AI models using publicly available internet materials is fair use"; that it "led the AI industry in providing a simple opt-out process for publishers"; that memorisation/regurgitation is "a rare bug that we are working to drive to zero"; and that "any one sector—including news—is a tiny slice of overall training data". This is OpenAI's general position. It has not been published as a response to the USA TODAY complaint.

NOT FOUND. No OpenAI statement specific to the USA TODAY filing, no USAI TODAY Co. press release about the filing, and no court order beyond the initial docket entries had been located as of 2026-10-13.

INTERPRETATION (builds on FACT). The change is not a new legal doctrine. It is a change in the composition, scale and timing of the litigation pressure on OpenAI, and in the evidence base available to a new plaintiff.

1. Local-news scale entered the docket. Earlier publisher suits came from national brands (NYT), speciality publishers (Ziff Davis, Britannica) or coalitions of small independents (the ~30 local publishers representing ~400 titles). USA TODAY Co. is the largest US local-newspaper owner, with 200-plus titles, and it has put 19 of them — spanning 13 states, from Rochester to Naples to Wilmington — into a single complaint. If the substitution theory holds for a ranked snowstorm listicle or a Florida wildlife story, it plausibly generalises to almost any local news site.
2. A licensor became a litigant against the one partner it had not signed. USA TODAY Co. already licenses content to Perplexity (July 2025) and Meta (December 2025) and participated in Microsoft's Copilot Daily. Its position is "licence those who pay, sue those who do not" — which is materially different from a publisher that rejects AI licensing in principle.
3. The GPT-2/GPT-3-era numbers were put on the record with specificity. The complaint does not say "vast amounts": it counts WebText entries per domain and C4 tokens per domain, and argues dataset weighting (WebText2 at 22% of the training mix) to show journalism was treated as high-value.
4. RAG summarisation, not just verbatim copying, is now squarely at issue. The New York Times and Seattle Times complaints leaned on long verbatim passages; this complaint's exhibits are mostly retrieval-style article summaries. That shifts the contested question toward whether close paraphrase of structure and facts is a market substitute — a harder question than verbatim reproduction.
5. It arrived during the MDL's summary-judgment phase. Cross-motions for summary judgment in In re OpenAI were filed 2026-09-04, with replies due 2026-11-06. A new tag-along plaintiff can build on discovery another party paid for, and can plead fresh claims about current products that older complaints could not.

What did not change: no ruling, no admission, no licence, no settlement. Nothing in the record establishes liability, and "more than $250 million" is a prayer, not an award.

BEFORE (as of 2026-10-07).

• OpenAI faces roughly 29 US copyright suits, the most of any AI company (tracker: ChatGPT Is Eating the World; Press Gazette maintains a comparable publisher-only roster).
• The centre of gravity sits in In re OpenAI, Inc. Copyright Infringement Litigation, No. 1:25-md-03143 (S.D.N.Y.), before Judge Sidney H. Stein and Magistrate Judge Ona T. Wang, with "News Plaintiffs" (NYT and others) and "Class Plaintiffs" (Authors Guild and authors). Summary judgment is fully briefed on the moving side; replies due 2026-11-06; no fair-use ruling.
• OpenAI's public posture: training is fair use; opt-out offered; licensing deals signed with 20-plus publishers covering 160-plus outlets.
• USA TODAY Co. (renamed from Gannett on 2025-11-18, NYSE: TDAY) licenses to Perplexity (2025-07-30) and Meta (2025-12-05), and has no licence with OpenAI. Its CEO, Michael Reed, has said the company is "committed to ensuring that our content is properly attributed and that we are fairly compensated".

CHANGE (2026-10-08). The pleading is filed. 14 plaintiffs, 7 OpenAI defendants, 19 publications, 3 counts, more than $250 million sought, plus § 503(b) destruction of models and training sets; a statement of relatedness ties it to the MDL. OpenAI does not comment on the day.

AFTER (expected trajectory — PREDICTION, not fact).

• The case is related to, and in practice folded toward, the MDL before Judge Stein; near-term activity is procedural (appearance, amended pleadings, schedule). A fair-use ruling in the MDL, whenever it lands, becomes the reference point for this case.
• The $250 million headline enters the accumulating "AI copyright exposure" numbers as an argument in OpenAI's pre-IPO risk narrative; the complaint itself leans on OpenAI's ~$852 billion recapitalisation valuation and its SEC IPO filing as evidence of ability to pay and commercial scale.
• Publishers without a licence gain a template: per-domain dataset counts plus RAG-summary exhibits plus quote-mining from MDL discovery.
• The likeliest eventual outcome remains a settlement/licence rather than a destruction order, based on precedent and OpenAI's stated preference for partnership where courts are avoided (INTERPRETATION/PREDICTION).

The legal mechanism (FACT — statutory and procedural).

• Count I (17 U.S.C. § 501). Reproduce, distribute and prepare derivative works under § 106(1), (3) without licence. Remedies include actual damages and profits, or statutory damages under § 504(c): $750–$30,000 per work, up to $150,000 per work for willful infringement. The complaint cites the willful maximum.
• Count III (17 U.S.C. § 1202(b)). Prohibits intentionally removing or altering copyright management information and distributing works knowing CMI has been removed. Section 1203 allows statutory damages up to $25,000 per violation.
• § 503(b). Permits the court, in a final judgment or decree, to order destruction of infringing copies or means used to make them. This is the hook for the "destroy the models and datasets" prayer.
• § 504(c)(1) compilation rule. "[A]ll the parts of a compilation or derivative work constitute one work." If a newspaper registered an edition as one collective work, each edition — not each article — may yield one award. The attached Exhibit A registrations determine the damages base. (FACT statutory; INTERPRETATION as to this case.)
• Fair use (17 U.S.C. § 107). OpenAI's principal defence: transformative use, minimal market harm from training, and a claimed paucity of regurgitation. The complaint pre-empts this with the substitution theory and the memorisation quotes. Courts have not resolved the question for OpenAI.

The technical mechanism alleged (plaintiff's theory — ALLEGATIONS).

1. Acquisition. Scrape publishers' sites directly, or acquire scraped copies from third-party corpora such as Common Crawl, "regardless of paywalls or other access restrictions" (ï¶4). Count I also treats the "Project Taxi" Bing Index transfer and "Project Mango" crawler as infringing acts by OpenAI.
2. CMI stripping. "Dragnet", "Newspaper" and "Gutentag" extractors separate article body text from nav chrome, ads and CMI-bearing footers, and the complaint alleges these by design remove bylines, titles, copyright notices and terms of use.
3. Training and memorisation. The complaint quotes OpenAI's own statement that networks are trained "to memorize the training data — that's their objective", and that a single exposure "causes some degree of memorization". Retained internal notes are cited as seeing an API that "might output existing content verbatim".
4. Retrieval and summarisation (RAG). At inference, ChatGPT retrieves the live article and returns a structured in-depth summary; the complaint's theory is that the summary is a market substitute because "users have less need to navigate to those sources". The complaint asserts the models were "affirmatively post-train[ed]" to summarise rather than to return articles.

The economics mechanism alleged. If a reader's question is fully answered in the chat, the page-view, ad impression and subscription-conversion that would have followed a visit are lost. The complaint quotes OpenAI's public statement that local-news demand inside ChatGPT runs at "about 1 million prompts per week".

INTERPRETATION, grounded in CONFIRMED facts.

• It raises the price of not licensing. The complaint explicitly turns OpenAI's own licensing record ("OpenAI knows a license is required—because it has paid for them", naming AP, Axel Springer, The Atlantic and Vox Media) into evidence of willfulness. If that framing survives, unlicensed training-data acquisition stops being a low-cost default and becomes a priced liability. For model developers, content licensing moves from opportunistic to structural cost.
• Local news is a large, previously under-litigated surface. 200-plus titles in one portfolio, 19 in this complaint. A favourable substitution finding here is a template that scales across hundreds of local sites.
• Indemnification risk becomes concrete for enterprises. Enterprise buyers of a model whose outputs are alleged to substitute for licensed journalism will ask harder questions about copyright indemnity scope, retrieval guardrails and robots.txt compliance. OpenAI's Copyright Shield for Enterprise and API customers and Microsoft's comparable commitments are the current answer; this case tests whether such indemnities are sufficient comfort.
• It arrives at the moment of maximum OpenAI capital-market exposure. The complaint expressly frames OpenAI's ~$852 billion valuation and SEC IPO filing as context. Pending copyright exposure that must be described in a prospectus creates settlement pressure (PREDICTION).
• It matters for the fair-use question indirectly. This filing does not decide fair use; the MDL before Judge Stein does — and the DOJ has argued in that MDL that training on copyrighted text should not expose developers to broad liability, which publishers have publicly contested.

FACT/INTERPRETATION.

• For publishers: a repeatable complaint architecture — per-domain dataset counts (WebText entries, C4 tokens), dataset-weighting arguments, RAG summary exhibits, and internal-quote mining from MDL discovery — that is cheaper to assemble than the original NYT case because already exists.
• For OpenAI (and other labs): a clearer signal that the licensing-only path has a residual litigant population, which strengthens the case for pre-emptive deals with large local-news portfolios (hedging, pricing power).
• For enterprises: a concrete worked example of the highest-risk LLM pattern for third-party content — "find and summarise this specific paywalled article" — which can be governed today via policy and tooling.
• For legal and compliance teams: a live test of whether close paraphrase of journalistic structure and facts can be a market substitute under copyright, an issue that verbatim-copying cases did not resolve.

1. Will OpenAI answer, move to dismiss, or consent to relation/consolidation before Judge Stein, and on what schedule?
2. Does the court treat the RAG-summary theory as stating a claim, or does it hold that summaries of facts and structure are non-infringing?
3. How will the § 504(c) compilation rule be applied to the Exhibit A registrations — per article or per edition?
4. Will a fair-use ruling in the MDL arrive before this case reaches dispositive motions, making it largely derivative?
5. Do other USA TODAY Co. titles follow, or does the 19-title set stay fixed?
6. Does USA TODAY Co. settle into an OpenAI licence, and if so at what order of magnitude relative to the reported News Corp ($250m/5yr) and Dotdash Meredith ($16m/yr) reference points?
7. Does the § 1202 CMI claim survive, given the mixed record for such claims in the parallel publisher cases (e.g. judicial scepticism about "removal" via dataset construction in the NYT case)?
8. How does OpenAI's IPO process interact with settlement timing?
9. Does this filing change crawler/robots behaviour or model retrieval defaults at any major lab?
10. Will regulators or legislators treat the filing as evidence for the pending bot-scraping bills (federal proposals modelled on a New York state law)?

PREDICTION (with the FACT base stated).

• Near term (days–weeks). Service on the seven OpenAI defendants; likely motions or stipulations on response deadlines; the Filing 3 relatedness statement is acted on, tying the case to 1:25-md-03143 before Judge Stein. No merits activity expected imminently.
• 2026-11-06. Replies due on summary judgment in the OpenAI MDL. A ruling after that — on fair use and on the surviving claims — is the highest-value single event for this case.
• Medium term. Most likely path is consolidation/relation into the MDL rather than an independently litigated trial. Settlement/licence remains the base case given both parties' histories (USA TODAY Co. licenses Perplexity and Meta; OpenAI has signed 20-plus publisher deals and settled Folha by agreement).
• Watch points. (a) Whether the court accepts the RAG-summary substitution theory; (b) how Exhibit A registrations drive per-work versus per-edition damages; (c) whether the § 1202 CMI count survives a motion to dismiss, given the mixed record in parallel cases; (d) whether more of the 200-plus USA TODAY Network titles are added; (e) any OpenAI filing or statement responding specifically to this complaint; (f) whether the filing shifts OpenAI's disclosure in its IPO process.

USA TODAY Co. did not sue OpenAI because it rejects AI. It sued the one AI company it has not licensed to. The publisher already takes money from Perplexity and Meta for the same journalism, and its CEO's stated test is attribution plus fair compensation. That makes the filing less a moral statement about AI and more a pricing negotiation conducted in federal court — with 19 local newspapers, more than 160,000 WebText entries and 122 million C4 tokens as the exhibits, and OpenAI's own executives' words ("largely substitutive, period", "users won't click", "the modern newsstand") supplied by someone else's discovery bill.

The genuinely new thing here is not the theory, it is the surface area: local news. If a chatbot's summary of a ranked snowstorm listicle or a Florida panther story counts as a market substitute, then every local newsroom in America has a claim, and every model developer has a licensing bill it has not yet provisioned for. That is why a $250 million prayer matters more than its arithmetic.

Two caveats keep this honest. First, everything in the complaint is an allegation; no court has accepted the substitution theory, and the § 504(c) compilation rule may shrink the damages math considerably. Second, the decision that actually resolves the underlying question will come from the consolidated MDL before Judge Stein, not from this filing. This case is best read as a volume indicator and a template, not as a verdict.

For practitioners, the actionable residue is small and concrete: stop asking chatbots to fetch and summarise paywalled articles in anything client-facing, keep copyright management information intact in any training pipeline, verify that retrieval respects robots.txt at query time, and know exactly what your vendor's indemnity does and does not cover. Those four moves cost little today and are exactly the behaviours this complaint is built to attack.