fair use

advancedwidespread

A US copyright rule allowing the use of protected works without permission or payment in certain cases if the use is fair.

In plain terms

Copyright law generally states: if you use someone else's work, ask for permission and pay for it. Fair use is an exception to this in the United States. You can quote a book in your review, parody a hit song, or show a film clip in class, and this typically does not require the author's consent.

It is like a film review. If you play a short scene to give your opinion on it, that is fine, because you give the viewer something different from the film itself. However, if you screen the entire film for an admission fee, you are replacing the original and depriving the creators of their revenue. Fair use is not a precise boundary, but a weighing: how you used the work, how much you took from it, and whether you caused harm to the copyright holder. There is no rule that a certain number of lines or seconds is always free.

It became daily news because of AI. Large language and image generation models were trained on countless books, articles, and images, mostly without asking the authors. Developer companies say this is fair use because the model does not copy but learns from the works. Authors, publishers, and newspapers, however, consider it unauthorized commercial use.

It is important that this is a US legal doctrine. A few other countries, such as Israel and South Korea, have adopted a similar rule, but the European Union and Hungary have no such flexible, general exception. Here the law itemizes the permitted cases; this is called free use (statutory exceptions).

How it works

There is no prior authorization. Fair use can be cited afterward as a defense when a rights holder sues someone, and the court decides on a case-by-case basis.
The judge first examines the purpose and character of the use. It matters whether it is for commercial or nonprofit educational purposes, and especially whether it is transformative, meaning it serves a different purpose than the original work or merely substitutes it.
The second aspect is the nature of the work (factual text or fictional, creative work), and the third is the size and significance of the portion taken from it.
The fourth aspect is market impact: does the usage take away the original work's buyers, or does it diminish its value?
The court weighs the four criteria together; there is no formula. In AI cases, it is therefore examined separately how the training data was obtained, what the model does with it, and whether it returns the original verbatim.

An everyday example

You read in the news that writers or a daily newspaper sued an AI company because it trained its chatbot on their works. The company almost always defends itself by claiming the training was fair use. The case hinges on whether the court accepts this.

Who is affected

lawyer, citizenThis rule determines in the United States whether AI companies can train on others' works without permission, and it also affects European debates. → Law

developerIf you train a model or build an agent that collects web content, you must know that American fair use does not protect you in Europe. → Code

marketerIf you use an AI-generated image or text in a campaign, you are also affected by whether the output resembles a protected work too closely. → Marketing

entrepreneur, executiveWhen selecting an AI tool, a legal risk is the source and legal basis on which the provider trained the model.

A common misconception

MythIf something is freely available on the internet, or I cite the source, it already counts as fair use.

In realityNeither availability nor citation makes use lawful. Courts evaluate fair use based on four factors, and this is a rule of US law: it cannot be invoked in Hungary or the EU.

Limits and risks

In Hungary and the EU, you cannot cite it: here, the statutory cases of fair use and the text and data mining exception apply.
It is unpredictable because two judges may decide differently in similar cases, and only a final court ruling provides certainty.
The fact that training a model may be fair use does not mean the output is legal: if the AI returns a protected work almost word for word, that could be a separate infringement.
The method of acquisition also matters: in 2025, a US court ruled that storing books downloaded from pirate sites did not constitute fair use, while considering the training itself to be fair use. This is a first-instance decision; another judge could rule differently.

Similar, but not the same

fair use – Hungarian and EU law list permitted cases in a closed list, while fair use is an open, assessment-based rule.
public domain – A work in the public domain has expired protection and can be used by anyone for any purpose, whereas fair use is a narrow exception applying to protected works.
text and data mining exception – This is an EU rule specifically tailored for machine analysis. Research institutes' scientific analysis cannot be prohibited by the rights holder, but for all other uses (such as commercial), they can object with a machine-readable statement.

Related concepts

Broader

Narrower

transformative use

On the edge

fair use, text and data mining, training data, licensing agreement, public domain

A brief history

Fair use grew out of American judicial practice; the 1841 Folsom v. Marsh ruling is considered its starting point. It was codified in law in Section 107 of the 1976 US Copyright Act, together with the four balancing criteria.

Where things stand

Long established as a legal institution, its application to AI training remains an open question. In 2025, several US first-instance courts ruled with different emphases and partly contradictory outcomes. The US Supreme Court has not yet issued guidance in an AI training case. Meanwhile, companies are signing more and more licensing agreements with publishers, and several lawsuits have been closed with settlements.

Status: 2026 November

In the news

Appears in 14 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Publishers file amicus brief against OpenAI's fair use defense in copyright caseOct 29, 2026 · Jog
Thomson Reuters v. ROSS: Third Circuit recognizes licensing market for AI training dataOct 27, 2026 · Jog
Copyright and AI: India, Australia and Taiwan debate rules as international study appearsOct 26, 2026 · Jog
Anthropic asks court to rule lyrics training is fair use in dispute with music publishersOct 19, 2026 · Jog
Ninth Circuit weighs AI training and copying as another court rejects fair use in first AI copyright appealOct 12, 2026 · Jog
Third Circuit rejects fair use defense for AI training in Thomson Reuters v. RossOct 5, 2026 · Jog
Appeals court upholds Thomson Reuters win: AI training on its material is not fair useSep 29, 2026 · Jog
Justice Department backs OpenAI's fair-use defense in copyright case; news publishers objectSep 28, 2026 · Jog

Also known as

fair use · fair use · fair use · fair use · fair use doctrine · fair use doctrine · fair use defense · fair use defense

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.