社会ガイド
Fair Use and AI Training Data
Fair use is the US copyright doctrine that courts are using to decide whether training AI models on copyrighted works without permission is lawful.
このページでは4 分で読めます
概要
Judges weigh four factors: the purpose of the use, the nature of the work, how much was copied, and the effect on the market for the original. The first 2025 rulings split. Training was found highly transformative in some generative AI cases, while pirated data and non-generative copying drew liability. The answer affects every AI developer and every creator whose work is in training data.
ディープダイブ
Section 107 of the Copyright Act lists four factors. Factor one asks whether the use is transformative and commercial. The Supreme Court's 2023 Warhol v. Goldsmith decision emphasized that a different purpose matters more than adding new meaning. Factor two considers whether the work is creative or factual. Factor three looks at how much was taken. Factor four asks whether the use harms the market for the original, and courts often treat it as the most important. The first rulings came in 2025. In February, Judge Stephanos Bibas held in Thomson Reuters v. Ross that Ross's copying of Westlaw headnotes to train a legal search tool was not fair use. Ross was building a direct competitor, and the AI was not generative. The case went to the Third Circuit on interlocutory appeal. In June, Judge William Alsup ruled in Bartz v. Anthropic that training on books was 'exceedingly transformative' and fair use, and that scanning purchased print books was also fair. He held that building a central library from pirate sites was a separate use that fair use did not cover. After class certification, Anthropic agreed to a settlement reported at $1.5 billion, about $3,000 per covered work. Days later, Judge Vince Chhabria ruled for Meta in Kadrey v. Meta, but only because the authors did not develop evidence of market harm. He stressed that 'market dilution' from floods of AI-generated competing works could weigh heavily against fair use in a better-argued case. A common misconception is that courts have declared AI training fair use across the board. These are district court rulings on specific facts. Appeals, other cases and the method of acquiring data all matter.
戦略的影響
リスクと安全性
AI による壊滅的な被害も日常的な被害も、誰がリスクを理解し、誰が行動できるかにかかっています。
より明確な判決
国民と専門家のリテラシーは、強力な安全政策が政治的に可能かどうかを左右します。
誇大広告を打ち破る
明確な説明は、誇大広告、研究室の PR、曖昧な倫理劇場に囚われることを減らします。
The Future of Fair Use and AI Training Data
Appellate decisions, including the Third Circuit's review of Thomson Reuters v. Ross, are likely to shape the doctrine more than any single trial ruling. Expect courts to keep separating how data was acquired from how it was used, and to look closely at market harm evidence. Licensing deals will probably keep growing, since they reduce legal risk regardless of outcome. Congress could legislate, and the Copyright Office has published its own analysis of generative AI training. Until then the law remains unsettled and depends on the facts of each case.
現実世界の実装
In Bartz v. Anthropic, a judge found training Claude on lawfully purchased and scanned books was fair use, but downloading millions of pirated books into a library was not excused.
In Thomson Reuters v. Ross Intelligence, a court rejected fair use for a legal research startup that used Westlaw headnotes to build a competing, non-generative search tool.
In Kadrey v. Meta, authors lost at summary judgment because they did not prove market harm, even though the judge suggested such harm could exist in other cases.
A news publisher signs a paid licensing deal with an AI company, which both earns revenue and supports arguments that a training license market exists.
リスクとガードレール
能力が複雑になる一方で、実存的なリスクを SF として扱います。
高度な自律性の下での調整による表面製品の安全性を混乱させる。
英語以外や専門家ではない聴衆には、低品質の情報源しか提供されません。
実装ロードマップ
製品の危害、誤使用、制御不能/調整不良のリスクを分離します。
どのような証拠がタイムラインと重大度についてのあなたの見方を変えるかを尋ねてください。
マーケティング上の主張よりも、一次情報源と具体的な評価を優先します。
意識だけでなく、キャリア、政策、資金、スキルなど、行動経路を 1 つ特定します。
探検を続けましょう
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Fair Use and AI Training Data quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
よくある質問
What is Fair Use and AI Training Data?
Fair use is the US copyright doctrine that courts are using to decide whether training AI models on copyrighted works without permission is lawful. Judges weigh four factors: the purpose of the use, the nature of the work, how much was copied, and the effect on the market for the original. The first 2025 rulings split. Training was found highly transformative in some generative AI cases, while pirated data and non-generative copying drew liability. The answer affects every AI developer and every creator whose work is in training data.
Which fair use factor asks about harm to the market for the original work?
Factor four considers the effect on the potential market for or value of the original, and courts often treat it as the most important.
Why did the court reject fair use in Thomson Reuters v. Ross Intelligence?
Judge Bibas found Ross's use was not transformative enough and served as a direct market substitute for Westlaw.
In Bartz v. Anthropic, what did Judge Alsup find was NOT covered by fair use?
Alsup held training and scanning lawfully bought books were fair use, but acquiring and keeping pirated books in a library was a separate, unexcused use.
What settlement did Anthropic reach in Bartz, as reported?
After class certification, the reported settlement was about $1.5 billion, approximately $3,000 per covered work.
Why did Meta win in Kadrey v. Meta?
Judge Chhabria ruled narrowly because the plaintiffs failed to show market harm, while suggesting a better-argued case could succeed.
学び続ける
関連ガイド
このトピックのために選ばれたその他のガイド