Exhibit 01 · The core question

Who owns what a model learned?

Every modern generative model is trained on vast collections of text, journalism and literature. The fight isn't about the model's answers — it's about how it was trained and who owns what it absorbed.

One distinction runs through every case: courts treat the training of a model differently from how the data was acquired. Learning from a work can be lawful even when the way it was obtained is not.

Hold this line

Lawful acquisition + training on one side. Piracy on the other. Almost every hard case argues which side a fact pattern lands on.

Educational, not legal advice. U.S. copyright focus, as of 13 November 2026. Allegations, rulings and settlements are kept distinct.

The brief

What you'll be able to do

1
Separate acquisition, training and output when spotting the copyright question.
2
Apply the four fair-use factors and the idea of market substitution.
3
Explain Bartz v. Anthropic and its settlement without over-reading them.
4
Tell allegations and advocacy apart from rulings in the NYT case.
5
See why licensing and dataset provenance are the emerging path.

Exhibit 02 · The parties

Creators vs. developers

Creators say…

Feeding our work into a model at scale is unauthorized exploitation. No one asked — and the model then competes with us using what it took.

Developers say…

Training is transformative learning. The model studies patterns; it doesn't republish the book. That's classic fair use.

Two ideas decide most fights. Fair use is strongest when a use is transformative (new purpose, not re-serving the original). Market substitution — output that replaces the original — cuts the other way.

Exhibit 02b · The four-factor test

What a court actually weighs

U.S. fair use isn't a single rule — it's a four-factor balancing test. Two factors do most of the heavy lifting in AI cases (marked in red).

1 · Purpose & characterIs the use transformative — a new purpose — or commercial re-use? This is where “training is learning” lives.

2 · Nature of the workFactual works get thinner protection than highly creative ones. Usually a minor factor here.

3 · Amount usedHow much was taken? Training often ingests whole works — but courts ask whether that was reasonable for the purpose.

4 · Market effectDoes the output substitute for the original in its market? Often the decisive factor.

The takeaway: factors 1 and 4 — transformation vs. substitution — are the tug-of-war you'll weigh on the next page.

Do it · weigh the factors

The fair-use balance

Toggle each factor and watch the scale tip. No single switch decides it — find a mix that leans fair use, then one that leans infringement.

BalancedAdd factors to tip the scale

Use has a new, different purpose+FU

Learns patterns, not verbatim copies+FU

Output is highly transformative+FU

Output competes in the same market−FU

Output substitutes for the work−FU

Whole works copied, no new purpose−FU

Knowledge check

Which factor most weakens a fair-use defense?

Exhibit 03 · Procedural history

Three phases of the dispute

The conflict moved through three stages. Reconstruct the record — click them from earliest to latest.

?

Copyright litigation — creators take developers to court.

?

Dataset provenance review — scrutiny of where data came from and how it was obtained.

?

Web scraping — an early period of largely unmonitored collection.

Click the earliest phase first.

Exhibit 04 · Case one

Bartz v. Anthropic

The landmark outcome so far — and it drew the key line in plain view. Judge Alsup split the question in two:

Training = fair use. Training on lawfully acquired books was fair use — the analysis is transformative.

Piracy ≠ protected. Downloading pirated texts from shadow libraries was a separate, unlawful act.

Read the caveat. Bartz does not mean all AI training is legal — the outcome turned on how the data was acquired, licensed and used.

Exhibit 04b · The remedy

A $1.5 billion signal

The unlawful acquisition created enormous exposure — and it settled for a historic $1.5 billion, the largest copyright class settlement in U.S. history, finalized in July. Why that number matters:

It was a class action

One suit stood in for many authors at once. That aggregation is what turns per-work statutory damages into a company-threatening figure.

It priced the piracy, not the training

The settlement answered the acquisition problem. The transformative-training holding stood — the money was for how the books were obtained.

Signal to the industry: buy or license your corpus. The cheapest-looking shortcut — a pirated dump — became the most expensive line item.

Do it · trace provenance

Where does liability enter?

Follow one dataset through the pipeline. Click each stage — the question at every node is the same: was this step lawful?

Stage 1

Acquisition

→

Stage 2

Retention

→

Stage 3

Training

→

Stage 4

Output

Pick a stage. Bartz tells us the first stage is where the piracy problem lived — not the training itself.

Knowledge check

In Bartz v. Anthropic, what did Judge Alsup rule was not protected?

Exhibit 05 · Case two

The New York Times v. OpenAI & Microsoft

The most-watched case still live. Publishers say their investigative journalism was ingested without permission; the firms say it's fair use.

Cross-motions for summary judgment were set for October 21. The DOJ filed a letter backing the fair-use defense, which publishers are fighting. Courts weigh technical evidence and market substitution.

Reading skill: keep three apart — an allegation (a complaint's claim), advocacy (a brief or the DOJ letter), and a ruling (what the judge decides). Only the last sets the law.

Exhibit 06 · The emerging path

Licensing moves to center stage

If piracy is the exposure and lawful acquisition is the safe harbour, the market's answer is becoming obvious: license the data. A fast-growing layer of deals now sits between creators and model-builders.

Content licensing deals

Publishers and archives sign paid agreements letting a developer train on their catalogue — turning a lawsuit risk into a revenue line.

Provenance & opt-outs

Dataset cards, robots signals and opt-out registries aim to record consent before training, not litigate it after.

Why it follows from Bartz: if the decisive question is how data was acquired, the cleanest way to win it is to acquire lawfully — by licence.

Exhibit 07 · Why it matters

The provenance template

Bartz's split — lawful acquisition + training = fair use, piracy = infringement — is becoming the template for how courts treat scraped data, and it reaches every generator: text, images, audio, code.

With many suits active and interpretations competing, the law here is still contested and evolving.

The one line to keep

The decisive question is almost always: lawful acquisition + training — or piracy?

Verdict

0/2

Knowledge-check score

Acquisition ≠ training. Courts separate how data was obtained from whether learning is fair use.

Four factors, two that matter. Transformative purpose vs. market substitution decide most AI cases.

Bartz drew the line. Lawful books = fair use; pirated downloads = infringement → $1.5B class settlement.

NYT is still live. Separate allegations and advocacy (incl. the DOJ letter) from actual rulings.

Licensing is the path. Lawful acquisition — by licence — is the clean way to win the decisive question.

Glossary

Fair use — a U.S. defense allowing some unlicensed uses, especially transformative ones. Transformative use — adds new purpose/character. Shadow libraries — pirate repositories of copyrighted works. Market substitution — whether output competes with/replaces the original. Dataset provenance — the record of where training data came from and how it was obtained.