First, acquisition: where did the file come from, and did obtaining it involve unauthorized reproduction? Second, retention: why is the developer keeping a general library, and which copies remain? Third, training: what copying and computational analysis occur to develop a model? Fourth, output and deployment: what does a user receive, and does the product reproduce protected expression or substitute for the source? A favorable argument at one stage does not automatically resolve the others.

The headline question—who owns the knowledge a model absorbs?—needs legal precision. Copyright protects original expression, not facts, ideas, or knowledge as such. The reporter does not own the underlying fact that a company polluted a river. The reporter may own the original wording and expressive organization of the investigation. Copying an entire article to acquire facts can still require analysis of the copying; the unprotected status of facts does not erase the protected expression in the source file. 17 U.S.C. §§ 102, 106 ↗

Worked example: an answer stating a publicly established date raises a different copyright question from an answer reproducing several distinctive paragraphs. Neither answer tells us whether a training file was purchased or pirated. Keep the acquisition record, training process, and output evidence separate before drawing a conclusion.

Transformative use adds a different purpose or character. It is not a synonym for changing a file format, running software, or producing something new. Ask what the challenged use actually does with the work. Both sides can agree that a model is technologically novel while disagreeing over whether a particular use replaces what the copyright owner sells.

Section 107 asks courts to consider four factors together. Purpose and character includes the asserted transformation and the commercial setting. Nature of the work considers characteristics such as creative expression and publication status. Amount and substantiality examines both quantity and the importance of what was taken, in relation to the purpose. Market effect considers harm to the work's market or value, including legally cognizable potential markets. There is no fixed percentage or automatic three-out-of-four score. 17 U.S.C. § 107 ↗

Apply the factors to our hypothetical investigation. A developer says full-text analysis was needed to learn language patterns: that explains a purpose and an amount, but it does not end the inquiry. The creator points to expressive reporting, full copying, and a paid product. We still need evidence of what the model does and the relevant market. Using an entire work may sometimes be justified by a transformative purpose; a smaller extract may still take the expressive heart.

Market substitution asks whether the use competes with or replaces the original in a legally relevant market. Compare a model explaining what investigative journalism is with a service supplying the key expressive passages of a particular paywalled investigation. The latter raises a stronger substitution concern. A lost licensing fee cannot simply be assumed to settle the fair-use question; parties dispute which licensing markets count and how harm should be established. These examples are analytical comparisons, not predicted verdicts.

The dispute can be taught in three phases, provided we do not mistake them for cleanly separated eras. Web scraping came first as an extensive and often opaque collection practice: automated systems fetched material at a scale creators could struggle to observe. Public accessibility never meant public-domain status. Collection also occurred through purchases, licensed datasets, and other sources; scraping was not the only route.

Source → copy → use → evidence

Copyright litigation then forced specific acts into court. Authors and publishers challenged copying, while developers raised defenses. A complaint states allegations; discovery gathers evidence; summary judgment asks whether a claim can be decided without trial under the applicable legal standard. A settlement resolves agreed claims, usually without deciding every disputed issue. These procedural distinctions prevent a headline about a filing from becoming a false headline about a ruling.

Dataset provenance review asks where individual files originated, what permissions accompanied them, and how copies moved into a training collection. It runs alongside collection and litigation today. Our reporter's book might appear in a vendor's dataset under a reassuring dataset name, but that name alone cannot tell us whether the source was a publisher license, a purchased book, or a shadow library. A shadow library is an unofficial repository offering copies often distributed without copyright permission. Identify the actual source and acts rather than assuming every item in a category has identical legal status.

In Bartz v. Anthropic, Judge William Alsup's 23 June 2025 order separated training copies, digitization of purchased books, and pirated copies assembled into a central library. The court treated the training use before it as highly transformative fair use. It also approved one-for-one digitization of purchased print books on the facts presented, including replacement of the physical copies. The pirated general-purpose library was not excused by the asserted training purpose. The ruling did not adjudicate infringing model outputs. Order on Fair Use ↗

Source → copy → use → evidence

This is more precise than saying the court approved all training on legally obtained books and prohibited all training on illegal ones. The court analyzed distinct copies and uses. Its training holding did not cleanse the separate creation of a pirated library; its acquisition analysis did not create a universal authorization for every downstream product.

For our hypothetical book, build separate evidence folders for the purchase or download, library retention, training process, and published answers. The same title appearing in multiple folders does not merge the legal questions. This is why provenance evidence can matter even when a developer has a substantial transformation argument.

The remaining dispute over pirated library copies created major exposure. The preliminary settlement order explains that the prior fair-use ruling left those issues for trial rather than granting an across-the-board victory. A settlement avoided litigating that remaining risk; it was not a damages verdict that all AI training infringes. Preliminary approval ↗

Source → copy → use → evidence

On 20 July 2026, Judge Araceli Martínez-Olguín granted final approval of the $1.5 billion class settlement. The order calls it the largest copyright class-action settlement in history. Distinguish Alsup's 2025 merits ruling from Martínez-Olguín's 2026 settlement approval. Approval evaluates the settlement for the class; it does not turn every unresolved allegation into a judicial finding. Final approval order ↗

The size matters to a media professional because provenance can create substantial economic exposure independently of the debate over learning. But a large payment is neither a nationwide precedent that all training is lawful nor proof that all training is unlawful. Identify which claims were resolved and which uses the court actually analyzed.

Example A: a developer buys print copies, replaces each with a searchable scan, and uses the copies for computational training without distributing book-like outputs. This resembles the purchased-book and training circumstances that received favorable treatment in Bartz. The resemblance supports an argument, not an automatic result in every court. Changed copying practices or product behavior can change the analysis.

Example B: a developer downloads unauthorized full texts from a shadow library into a permanent collection, then trains a useful general model. Separate the initial library reproduction from the later training purpose. A transformative use downstream does not retrospectively authorize the library acquisition. The developer should not cite the training holding as a blanket answer to the acquisition claim.

Example C: a publisher grants a license expressly covering specified training copies and uses. Start with the license's scope, term, permitted users, and restrictions. Conduct within a valid permission can be authorized without relying on fair use. A license for reading or internal reference should not be silently described as a training license.

Example D: a company obtains sources lawfully but offers answers that reproduce extensive expressive passages from an investigation and plausibly replace paid access. Lawful acquisition answers only the acquisition question. Output infringement and market effects need separate attention. None of these hypotheticals is legal advice or a ruling on a real product.

The New York Times sued OpenAI and Microsoft in December 2023. The publishers' position is that unauthorized copying of expensive journalism helped build competing services. The developers contend that training serves a transformative purpose and qualifies as fair use. Their disagreement includes what the systems do, what the technical record establishes, and whether products substitute for journalism. Treat each side's account as advocacy until the court resolves the relevant question.

Source → copy → use → evidence

The news plaintiffs' public September 2026 memorandum seeks summary judgment over several asserted uses. OpenAI's source hub links its own 4 September 2026 memoranda. Competing motions ask a judge to resolve issues; they are not competing judicial decisions. News Plaintiffs' filing ↗, OpenAI's filings ↗

Technical evidence could include copying logs, dataset contents, memorization tests, prompt design, retrieved source passages, and output comparisons. If a test produces a long matching passage, ask which system version was tested, whether it retrieved a source, how the prompt was constructed, and what the test demonstrates. A dramatic example can matter without establishing ordinary behavior across all users. A developer's claim that an example is unusual does not automatically dispose of it either.

For substitution, compare a general question about a reporting topic with a request that delivers a source's expressive account in place of paid access. Examine the evidence linking the challenged use to a relevant market. Do not confuse factual overlap with copying protected expression, or business competition with an automatic infringement finding.

Status note, checked 13 November 2026: the supplied brief says cross-motions were set for October 21, but that date was not corroborated in the sources reviewed. The verified record includes September 2026 summary-judgment filings. This lesson therefore does not present October 21 as an upcoming or established deadline and does not claim that a final Times merits outcome has been verified.

Source → copy → use → evidence

The DOJ filed a Statement of Interest on 1 September 2026 supporting a broad fair-use position on AI training. Publishers oppose that position. A government submission asks the court to adopt an interpretation; it does not amend copyright law, adjudicate the facts, or bind the judge. The 30 October congressional proposed amicus filing discusses the intervention and illustrates continuing disagreement. Proposed amicus filing ↗

When publishing a case update, label the date and source for each assertion. Write that a party argues, a government filing urges, or a judge rules according to what actually occurred. A schedule can change and a motion can be denied, granted in part, or deferred. Accurate procedural language protects the audience from mistaking momentum for a final answer.

Bartz provides a useful analytical template: separate how a work was obtained from how it was used for training. The slogan lawful acquisition plus training versus piracy captures the warning about acquisition, but it is not a complete fair-use test. One district-court decision is not a universal binding rule for every model, medium, jurisdiction, or factual record. The Copyright Office's May 2025 pre-publication report also treats training as a context-dependent copyright question; it is policy analysis, not a judicial holding. Copyright Office report ↗

Source → copy → use → evidence

The same provenance questions arise for image, music, audio, and video generators. Ask who supplied the files, what rights they held, what copying occurred, and what the deployed product produces. Do not transfer a books ruling into an assurance that every image or music model is cleared. Different works, rights, contracts, and outputs can require different analysis.

A platform or festival may require disclosure of AI use or assurances about rights. Those are policy and contractual layers, not a judicial determination that training was lawful. Requirements differ by venue and change over time; check the actual current rules. Honest disclosure helps a submission satisfy a rule, but it cannot supply missing permission for a copyrighted source.

Return to our reporter's investigation. A useful provenance record identifies the work and file version, original source, acquisition date and method, license or other asserted basis, supplier authority, retained copies, training use, and relevant restrictions. Link entries to evidence such as receipts, agreements, logs, and documented review. A record can also state unknowns; concealing them under a broad assurance weakens its value.

Suppose a vendor says its dataset is commercially safe but provides only a spreadsheet of titles. The list establishes possible contents, not permission. Ask for the source route and rights basis for the included works, the uses that basis covers, and how exceptions are handled. If the vendor cannot answer, record the uncertainty and seek appropriate rights or legal review before relying on its assurance. This is an evidence evaluation exercise, not a universal legal checklist.

Now suppose the vendor supplies receipts and an express training license, but your publication draft reproduces a source's distinctive paragraphs. Strong input documentation does not settle publication rights. Review the output, the intended use, the source comparison, and the applicable venue's disclosure terms. Provenance is practical evidence; it is not immunity.

Identify the act before evaluating the defense: acquisition, retention, training, and output can raise different questions. Apply all four fair-use factors, including transformation and market effects. Bartz favored the training use before it while refusing to excuse a pirated library; its $1.5 billion settlement received final approval on 20 July 2026. The Times litigation features contested technical and market evidence, and a DOJ submission is advocacy rather than a ruling. Provenance and venue disclosure are useful practical layers, each with limits. The key acquisition line is lawful sources versus piracy; lawful acquisition alone does not settle all later uses. The law remains contested and evolving. Educational information, not legal advice.

Source → copy → use → evidence