01
A creator’s field lesson
Training Data
on Trial.
One dataset. Four different questions.
Priya is building a story-research assistant for documentary creators. Her team can license a local-news archive, buy books, or accept a cheap bulk download whose origins are unclear. Each route supplies text. Each leaves different questions about copying, permission, and the product's effect on creators.
Follow one batch of text from its source to a user-facing answer. You will separate the uses, weigh fair-use arguments, read two disputes carefully, and evaluate Priya's sourcing plan. This lesson concerns US copyright and is educational, not legal advice. Its examples cannot establish whether a particular product is lawful.
What you’ll practice
Priya and her product are hypothetical. US focus · Sources checked November 15, 2026.
01 / Separate the questions
A download is only the beginning.
The model’s learning
Many large language models learn statistical patterns from enormous mixtures of text, which can include journalism and literature. Model types and datasets vary: not every modern model uses these materials. Training often involves making copies; learning patterns does not mean that no copying occurred, and it does not mean every model stores a searchable library of originals.
The copies behind it
Priya needs two questions at minimum: How did we get these copies? What are we doing with them? A training purpose may support a fair-use argument without excusing a separate unauthorized source library. Output that reproduces protected expression raises further questions. Copyright protects expression, not ownership of facts or ideas themselves.
First, web scraping makes large collections technically accessible. Then copyright litigation challenges particular copying and uses. Dataset provenance review asks for the history of the data: its source, permissions, transfers, and handling. These activities overlap; this sequence describes a shift in scrutiny, not a rule that scraping ended when lawsuits began.
For Priya, a download URL is the start of an investigation, not proof of permission. Publicly reachable text is not automatically public domain or licensed for model training.
02 / The legal framework
What exactly is being weighed?
Creators’ position
Creators argue that mass ingestion exploits protected work without authorization and can power products that replace purchases, subscriptions, or visits.
Developers’ position
Developers argue that training learns patterns for a different purpose and can qualify as transformative fair use.
These are competing positions; neither slogan resolves Priya's particular facts.
Four factors · considered together
Fair use permits some uses of copyrighted works without permission. Courts weigh four factors together: purpose and character of the use, including commerciality and transformation; nature of the copyrighted work; amount and substantiality taken; and effect on the potential market for or value of the work. There is no automatic percentage, checklist score, or universal AI exemption.
01
Purpose & characterWhat is the use doing?
02
Nature of the workWhat kind of expression is copied?
03
Amount & substanceHow much, and why that much?
04
Market effectWhat demand might be displaced?
For Priya, factual reporting and creative narrative may raise different nature-of-the-work considerations. Copying an entire archive raises amount questions even if a technical purpose requires substantial material. A commercial product is relevant but does not automatically lose; a new technical process does not automatically win.
03 / Drive the fair-use balance
Change the facts. Watch the argument shift.
Transformative use involves a further purpose or different character, assessed in context. Priya's tool identifying themes across reports presents a different argument from a tool delivering the substance of a paid feature so users can skip it. Market substitution means replacing demand for the original or a legally relevant market for it.
Ask what the product actually returns and what users do instead of buying or visiting. Evidence matters: output examples, overlap with protected expression, usage patterns, and the asserted market. Not all competition is legally cognizable harm, and claiming a lost licensing fee alone does not settle fair use. A favorable purpose argument and a serious substitution concern can coexist.
Any simplified balance in this lesson illustrates the direction of arguments only. It cannot calculate fair use. All four factors and the particular facts still matter.
04 / Case file: Bartz v. Anthropic
Bartz: separate uses, different outcomes.
Training use at issueFair use
Purchased-book digitizationFair use on these facts
Pirated central libraryNot fair use
In Bartz v. Anthropic, Judge William Alsup's June 23, 2025 order treated the training use at issue as transformative fair use. It separately approved digitizing purchased print books on the facts presented, including replacement of the physical copies without redistribution. It rejected fair use for downloading and retaining pirated books in a permanent central library. Later training, or later purchases, did not erase that separate library-copying problem.
The important distinction is between uses, not a slogan that all training is lawful if a book was bought. No infringing outputs were alleged in that case. Different outputs, markets, acquisition conduct, or evidence can change another analysis. A district-court ruling is not nationwide blanket permission.
$1.5 billion
Settlement · July 20, 2026The remaining pirated-copy claims led to a $1.5 billion settlement, granted final approval on July 20, 2026 by Judge Araceli Martínez-Olguín. It was reported as the largest known copyright settlement. A settlement resolves covered claims on agreed terms; it is not a ruling that every disputed practice was lawful or unlawful.
Sources: June 23, 2025 order · Court-authorized settlement site · Largest-known recovery described in AP’s report, attributing plaintiffs’ counsel.
Priya says: “If training is fair use, keeping our pirated source library is covered too.” What is the strongest response?
05 / Drive the provenance pipeline
Follow Priya's provenance pipeline.
Dataset provenance connects these stages with a documented history. Records help identify and investigate risk; a complete spreadsheet does not make an unauthorized act lawful.
06 / Case file: NYT v. OpenAI & Microsoft
A claim is not a court decision.
The New York Times alleges that OpenAI and Microsoft copied its journalism without authorization and built products that can substitute for its work. The defendants contest the claims and argue fair use. Priya should describe these as allegations and defenses, rather than proven findings about every model output.
Summary-judgment motions ask a court to decide issues without trial where the governing legal standard is met. September 4, 2026 filings are documented; filing a motion is not winning it. The supplied October 21 cross-motion date was not verified and is not taught as fact.
ComplaintAlleges
BriefArgues
OrderRules within scope
DOJ · September 1, 2026 · Advocacy
The DOJ's September 1, 2026 statement of interest supported the fair-use position. Government advocacy can influence a debate but is not the judge's decision. A court may also issue procedural rulings without deciding the ultimate training fair-use question. No final merits outcome on that issue is established by this lesson's reviewed sources as of November 15, 2026.
Priya’s reporting rule
Priya's reporting rule: identify who said it, what document contains it, and exactly what the court decided, if anything. A complaint alleges; a brief argues; an order rules within its stated scope.
A headline says: “DOJ supports OpenAI, so the judge has ruled training lawful.” What needs correcting?
07 / Bring it back to Priya
Priya chooses a reviewable route.
01 / Verify the source and scope
Priya pauses the unknown-origin bulk download while investigating. She evaluates the licensed archive's actual terms: which works are covered, whether the licensor controls the relevant rights, and whether training, retention, and intended outputs fall within scope. She records sources and permissions, tracks the training mix, and tests for reproductions and substitutes before launch.
02 / Evaluate the available routes
Licensing is one emerging route to negotiated access and compensation. It is not the only possible lawful route, nor does a license from one provider necessarily cover third-party material or every downstream act. Public-domain materials and sufficiently authorized creator contributions may also be useful; verify their status and scope.
03 / Keep the conclusion bounded
Bartz makes source-copy history hard to ignore. Treat applying its acquisition/training distinction to other scraped datasets as a useful question to investigate, not a settled legal template for all data. Priya's decision is to document and review each use rather than announce that a case has cleared her product.
Before you decide
Source → copies → training use → user-facing output. Keep each question visible as you move through the product.
Assessment
Assessment Starting
You are about to begin the assessment. Select the best answer for each question.
Priya's training use has a strong fair-use argument. Which issue still needs its own analysis?
Priya's assistant returns protected passages that let users skip paid features. Which analysis is strongest?
Which summary of Bartz is best supported?
The Times' complaint says defendants built substitutes for its journalism. How should Priya label that statement?
Which next step best addresses Priya's uncertain-origin dataset?
Your review / Training Data on Trial.
Results
—Final score
Complete the assessment
Answer all five questions to see your review.
02
Priya's assistant returns protected passages that let users skip paid features. Which analysis is strongest?
03
Which summary of Bartz is best supported?
04
The Times' complaint says defendants built substitutes for its journalism. How should Priya label that statement?
05
Which next step best addresses Priya's uncertain-origin dataset?
Priya's recap: separate acquisition, retention, training, and output. Weigh purpose alongside market effects and the other factors. Read holdings narrowly and label allegations and advocacy. Check provenance and permission scope. No single case, document, or safeguard guarantees legality.
Certificate of completion
Training Data on Trial.
Your name
Completed an educational lesson on US AI training-data copyright issues. This certificate records lesson completion, not legal expertise or legal clearance.