Intelligence Piracy: The Expert Who Was Never Hired

In 2020, a legal research startup called ROSS Intelligence went to Thomson Reuters and asked to license Westlaw. It was told no. Westlaw's owner was not about to hand its editorial content to a company building a rival product.

On the district court's findings, ROSS took another route. It paid a third party for tens of thousands of lawyer-written legal memos which had been compiled with reference to Westlaw's headnotes, the short editorial summaries of legal points that Thomson Reuters' editors had written and refined over decades. ROSS trained its search engine on those memos and launched a competing legal research tool ¹.

No user of ROSS ever saw a Westlaw headnote. In February 2025, a federal judge held that ROSS had nonetheless infringed copyright in 2,243 of them, and rejected its fair use defence ². ROSS disputes that finding. The Third Circuit heard argument on the appeal on 11 June 2026 and has not yet ruled, so nothing below should be read as settled law ³.

ROSS did not survive to see any of it. The company announced in December 2020 that it would close, ceasing operations at the end of January 2021, saying the lawsuit had made it impossible to raise funding or find a buyer. It called the claim spurious and accused Westlaw of using litigation to stifle competition. It went on defending the case after the product was gone ⁴.

A Westlaw headnote is not raw law. It is a piece of writing. An editor read a judgment, decided which point mattered, and wrote a summary a lawyer could rely on. That judgment is the product. The case itself is free.

So look at what ROSS actually built. Not a copy of those summaries. The ability to produce the next one — to answer the question a lawyer would otherwise have paid a Westlaw editor to answer.

The summary still appears. The editor never existed.

What a newsroom actually owns

In the earlier articles in this series, I argued that AI attacks capability rather than output, and that every substantial archive has quietly become a teacher model. In film, the capability being extracted was creative. In sport, it was tactical. In news and publishing, it is something harder to name and harder to defend: authority.

A publisher's real asset was never the sentences. It was the accumulated judgement behind them, which sources are reliable, which claims need a second call, how to explain a complex subject to a reader who has ten seconds, what a story sounds like when it can be trusted. Decades of editing, correction, litigation and reputational risk go into that. None of it appears in the text, and all of it is legible in the text at scale.

The model is not trying to reproduce the article. It is trying to become the publication.

The first sector where the harm is measurable

Film and sport have a proof problem in common: the damage is argued rather than counted. Nobody can yet put a figure on what a synthetic performer costs a working actor.

News is different, and that is why this piece matters. Here, substitution shows up in the numbers.

AdExchanger reported Google referral traffic to publishers down by around 38% year on year as of January 2026 ⁵. Search Engine Land, analysing click data, put the fall in traffic to news sites at roughly 26% in the twelve months following the launch of AI Overviews, with individual publishers reporting far steeper declines ⁶. SparkToro estimates that close to 60% of searches globally now end without a click ⁷. In one industry survey, publishers expected search referrals to fall by a further 43% within three years ⁸.

In September 2025, Penske Media, publisher of Rolling Stone, Billboard and Variety — became the first major US publisher to sue Google over AI summaries specifically, alleging that around 20% of Google searches reaching its sites now display an AI Overview that answers the reader's question on the results page. Google denies the claims and has moved to dismiss; the motion was still before the court in 2026 ⁹.

These are third-party estimates and the causes are contested. Google disputes the framing, and search behaviour is shifting for several reasons at once. But the direction of travel is not seriously in doubt, and it is the first point in this series at which the loss can be counted at all rather than only argued.

Authority without accountability

There is a second harm here that the film and sport pieces did not have to deal with.

As of June 2026, NewsGuard was tracking 3,749 AI content farm sites publishing in 16 languages, sites built to look as though journalists produced them, with little or no human oversight and no disclosure ¹⁰. That number was 49 in May 2023.

These operations are not stealing articles. They are borrowing credibility. The register, the structure, the confident explanatory tone — all of it learned from institutions that spent a century earning the right to sound that way, and none of it carrying the corrections policy, the editorial chain or the legal exposure that made the tone meaningful in the first place.

Film piracy took a performance. This takes the reason anyone believed it.

The subscription nobody renews

The sharpest commercial exposure is not in general news at all. It is in specialist publishing, where the whole business model is that someone pays for an answer.

A legal publisher sells summaries of the law. A medical publisher sells clinical guidance. A financial data house sells market analysis. In each case, the subscriber is buying a conclusion they trust, produced by people who are accountable for getting it wrong. An AI trained on that corpus can produce something that reads identically, instantly, at no marginal cost, and to a user who has no way of telling the difference. ROSS is not an outlier. It is the template.

The other reading of that case is worth holding in mind. Copyright enforced defensively by an incumbent looks a great deal like copyright used to protect a market position, and the litigation destroyed a smaller competitor years before any court decided whether it had done wrong. Both things can be true at once. Any rights holder planning to enforce in this space should expect that framing in return.

Consent and payment remain entirely possible, and much of the industry has taken that route. News Corp's 2024 agreement with OpenAI was reported at up to US$250 million over five years, and the New York Times has licensed content to Amazon while simultaneously litigating against OpenAI ¹¹.

Those are the licensed cases. They are the model, not the problem. Intelligence piracy describes the opposite situation — where no licence is sought or given, and the training proceeds regardless.

Why this is an investigator's problem

Each article in this series has made the evidence a little harder to find, and this is where it gets hardest.

In film, the theft left a face, something a biometric comparison can be run against. In sport, it left statistical fingerprints in the tactical output. In news and publishing, a well-trained model leaves almost nothing. No verbatim passage. No likeness. No single work to hold up in court. The traditional test — has a copy been made? — can simply return the answer no. In Getty v Stability AI, the English High Court found that the trained model did not store copies of the images used to build it, and so was not an infringing copy for the purposes of secondary infringement. Getty has permission to appeal that finding, so it is not the last word ¹².

So the questions change shape:

• Provenance. Was this corpus ingested, when, by whom, and in which jurisdiction? Getty abandoned its primary training claim before trial because it could not establish that the training had taken place in the UK at all — a reminder of how quickly a claim can fall away on a question that has nothing to do with the merits.

• Substitution. Can it be shown that the output displaces the publisher's product rather than merely resembling it? This is the ground on which ROSS lost at first instance, and it is where the commercial evidence, traffic, referrals, subscription behaviour, starts to matter as much as the textual evidence.

• Attribution. Who built the system, who funded it and who is monetising it. The familiar climb through infrastructure, financial and human attribution that the Intelligence Attribution Pyramid sets out.

The techniques that answer them will look unfamiliar to most anti-piracy teams: seeded and marked content placed to surface later in model outputs, structured probing of models against a publisher's proprietary corpus, editorial style fingerprinting, and referral-loss analysis tied to specific content sets. Detection alone will not carry a case. Attribution will.

One caveat is worth stating plainly, because pretending otherwise helps nobody. "Did this model learn everything my newsroom spent decades building?" describes a real commercial injury, but it does not map cleanly onto any established cause of action I am aware of. Adjacent doctrines: unfair competition, database rights, misappropriation may reach parts of it in some jurisdictions. None of them was designed for this. The gap between the harm and the law is wide, and closing it will take evidence that does not yet exist in most disputes.

The byline that belongs to no one

The first generation of piracy copied the article. The second copied the site that hosted it. The third produces the answer the article was written to give, credits nobody, and sends no one back to the source.

ROSS never published a Westlaw headnote. The content farms have no reporters. The AI Overview at the top of the page has no editor to call. In each case, the work was delivered, and nobody did it.

And the absence runs in both directions. No copy was made. No artefact was published that anyone can point to. The expert was never hired, and there is no record of the moment their expertise changed hands.

The next article turns to brands, where the same extraction is happening in plain sight, because a brand spends millions teaching the public exactly who it is, and publishes every word of that instruction on its own website.

Footnotes

1. Background to Thomson Reuters v ROSS Intelligence, including ROSS's refused licence request and its acquisition of LegalEase "Bulk Memos" compiled with reference to Westlaw headnotes. https://www.lawnext.com/2025/02/breaking-federal-judge-rules-legal-research-startup-ross-infringed-westlaws-copyrights-rejecting-fair-use-defense.html

2. Thomson Reuters Enterprise Centre GmbH v ROSS Intelligence Inc., D. Del., judgment of 11 February 2025 (Bibas J.), finding infringement of 2,243 Westlaw headnotes and rejecting fair use. https://www.reedsmith.com/articles/court-ai-fair-use-thomson-reuters-enterprise-gmbh-ross-intelligence/

3. Third Circuit oral argument in the appeal, 11 June 2026; judgment reserved. https://www.jdsupra.com/legalnews/ip-hot-topic-third-circuit-probes-fair-8976687/ (Note: no decision has issued before publication.)

4. ROSS Intelligence's announcement that it would cease operations, December 2020, closing at the end of January 2021, and its response to the litigation. https://www.lawnext.com/2020/12/legal-research-company-ross-to-shut-down-under-pressure-of-thomson-reuters-lawsuit.html and https://legaltechnology.com/2020/12/11/ross-to-shut-in-wake-of-thomson-reuters-lawsuit-but-continues-to-fight-claim/

5. Reporting on the collapse in Google referral traffic to publishers, January 2026. https://www.adexchanger.com/publishers/the-ai-search-reckoning-is-dismantling-open-web-traffic-and-publishers-may-never-recover/

6. Analysis of AI Overviews' effect on clicks to publisher sites. https://searchengineland.com/google-ai-overviews-cut-search-clicks-report-471497

7. SparkToro, "In 2026, Less than One Third of Google Searches Still Send a Click." https://sparktoro.com/blog/in-2026-less-than-one-third-of-google-searches-still-send-a-click/

8. "News publishers expect search traffic to drop 43% by 2029," Search Engine Land. https://searchengineland.com/news-publishers-search-referrals-drop-report-467408

9. Penske Media Corporation v Google LLC (D.D.C., filed September 2025); Google's motion to dismiss and Penske's February 2026 opposition. https://www.cjr.org/analysis/penske-ai-lawsuit-google-tactics.php (Note: Live litigation)

10. NewsGuard AI Tracking Center, recording 3,749 AI content farm news and information sites across 16 languages as of 23 June 2026. https://www.newsguardtech.com/special-reports/ai-tracking-center/

11. Reported terms of the News Corp–OpenAI agreement (May 2024) and the New York Times–Amazon licensing deal. https://pressgazette.co.uk/platforms/news-publisher-ai-deals-lawsuits-openai-google/ (Note: Deal values are reported, not confirmed)

12. Getty Images (US) Inc & Others v Stability AI Ltd [2025] EWHC 2863 (Ch), 4 November 2025. Getty abandoned its primary training and output copyright claims during proceedings; the court rejected the secondary infringement claim and found limited historic trade mark infringement. Getty was granted permission to appeal the secondary infringement finding in December 2025. https://www.judiciary.uk/wp-content/uploads/2025/11/Getty-Images-v-Stability-AI.pdf (Note: Appeal pending).