Back to Learning Area

Copyright & Data Theft

Copyright and AI: two concepts that haven't figured out how to deal with each other yet.

  • Ethics
  • 8 min
  • Beginner

Plain Language

This preview is in Plain Language. The whole portal is also available in Plain Language. You can change the language at the top. Or use the button below. Our Plain Language texts are AI-assisted and editorially reviewed. They follow the core principles of easy-to-read language, but they are not certified Easy Read: a review panel of people with learning difficulties does not test them.

Preview: AI and copyright

Can AI use my work?

AI programs learn from very many texts and images. Many of them are protected by copyright. Copyright means: the author decides about their work. This is not yet settled. Courts are debating it.

Can I protect myself?

Authors can say no to AI training. This is called opt-out. Opt-out means: I say no. There are ways to do this. One possibility is the robots.txt file. This file blocks AI crawlers. Crawlers collect data from the internet.

AI images belong to no one

Do you make an image with AI? The image does not automatically belong to you. In the USA, the rule is: AI images are not protected by copyright. This means: no one owns the image. The law only protects your own part.

AI models learn from billions of images, texts, and songs — most of them copyrighted. Artists discover their styles replicated in AI generators without ever giving consent. Companies generate millions in content but often hold no rights to it. Three questions define this conflict: Is AI training on copyrighted data legal? Can creators opt out? And who owns what an AI produces? The answers are far less clear than either side claims.

Caution

Legal status as of November 2026. The legal landscape for AI copyright is evolving rapidly — court rulings may change the positions described here.

Fair Use — When May AI Use Your Data?

Definition

Fair Use

Analogy

Imagine reading every cookbook ever published, absorbing techniques and flavor combinations, then writing your own recipes from memory. Did you steal or did you learn? That is the core dispute in AI training — and courts have not yet given a definitive answer. Breaking point: A human genuinely transforms and reinterprets. An AI model can sometimes reproduce near-exact copies (the Getty lawsuit rests on this), which considerably weakens the "transformation" argument.

US Fair Use vs. EU TDM Compared

US: Fair Use Doctrine

Case-by-case assessment using four factors (purpose, nature, amount, market impact). AI companies argue: training is "transformative" — the model extracts patterns, not copies. No statutory opt-out right. Pending cases: NYT vs. OpenAI, Getty vs. Stability AI.

EU: TDM rules of the DSM Directive (2019)

Statutory regulation of Text and Data Mining (TDM), i.e. the automated analysis of large volumes of text and data. Art. 3: TDM for scientific research permitted. Art. 4: TDM also commercially permitted — but with an explicit opt-out right for rights holders. Legally binding within the EU.

Example: Getty Images vs. Stability AI

Stability AI trained Stable Diffusion on parts of the LAION-5B dataset — a list of over five billion image links with captions from the internet. Generated images repeatedly contained distorted versions of Getty watermarks. For the plaintiffs, this is evidence that the model memorized specific copyrighted images rather than learning only abstract patterns.

Caution

Misconception: "AI training is clearly illegal"

Wrong. The legal status is genuinely unsettled. Both sides present valid legal arguments. In 2025, two US federal judges ruled that training on books was fair use (Anthropic, Meta) — explicitly only for the specific case at hand. In Anthropic's case, that did not cover collecting pirated copies; Anthropic therefore agreed to a $1.5 billion settlement. In 2026, a US appeals court rejected fair use for training an AI, albeit a non-generative legal search engine (Thomson Reuters v. Ross). In the US, a Supreme Court ruling is still pending. In the EU, the DSM Directive addresses at least part of the question. Its exception for text and data mining generally covers AI training, the Advocate General at the EU Court of Justice argued in November 2026; the ruling is still pending. It will likely take years before established precedents exist.

Opt-Out — The Right to Say No

Opt-out is the right of creators to prohibit the use of their works for AI training. Since the default on the internet is that anything accessible gets crawled, creators must take active steps. Four mechanisms are available — but all share a fundamental limitation.

robots.txt Website operators can deny access to AI crawlers. Technically just a request to crawlers; in the EU, though, such an entry can count as a machine-readable objection under Art. 4.3.
EU TDM Opt-Out Art. 4.3 of the DSM Directive: rights holders can object in machine-readable form. Legally binding within the EU; since August 2025, the AI Act has explicitly required providers of general-purpose AI models to respect such reservations.
Platform Tags DeviantArt, ArtStation, and others offer "No AI Training" tags. Activatable per click or already on by default, but only effective on the respective platform.
Glaze Tool from the University of Chicago: Alters pixel structure invisibly to the human eye, but deliberately disrupts the AI's learning process.

Example: The Photographer and His Stolen Style

A photographer discovers that his unique style is perfectly replicated through the prompt "in the style of [Name]." He adds a restrictive robots.txt to his website. This blocks future crawling — but the existing model has already encoded his style in billions of weights. Only training a completely new model from scratch without his data could remove his style — practically infeasible. On top of that, a style as such is not protected by copyright, only specific works. He can only act against an AI image once it recognizably takes over a specific photo of his.

Caution

Misconception: "Opt-out fully protects me"

No. Opt-out mechanisms only protect against future training runs. Many widely used AI models today build on datasets collected before such opt-outs were common. Retroactive opt-out is practically impossible because the extracted patterns are embedded in billions of model weights.

AI-Generated Content — Who Owns the Output?

The US Copyright Office established in 2023: Purely AI-generated works without significant human creative contribution are not copyrightable. The rationale: copyright protects human creativity — a machine is not an author. For mixed works (human arrangement + AI generation), the human portion can be protected.

LAION-5B Images

5  Bn

Over five billion image-text pairs from the internet: Stable Diffusion learned from parts of them

Copyright Protection (US)

0 %

Purely AI-generated images receive no copyright protection under US law — they effectively belong to nobody

Zarya Ruling

2023

US Copyright Office: text and layout protected, AI-generated images not

Note

Case Study: Zarya of the Dawn

Kris Kashtanova created a comic in 2022 using her own text and Midjourney-generated images. In February 2023, the US Copyright Office ruled: the text and page layout (human creation) are copyrighted. The individual AI-generated images are not — they lack the required human authorship.

Caution

Misconception: "AI images are free of copyright issues"

Dangerous half-truth. The generated image itself may not be copyrightable — but if it closely resembles an existing copyrighted work, the original rights holder can still sue for infringement. Regardless of whether a human or a machine created the copy.

Interactive: Is Your AI Output Original Enough?

Use this checklist to evaluate whether your AI-generated content is on solid legal and ethical ground. Each item carries a different weight — at the end you will see an overall assessment.

Originality Check for AI Content

Check your AI-generated content for copyright risks. Answer the questions honestly — you will receive an assessment at the end.

1.Have you verified that the output does not closely resemble any specific existing work?
Use reverse image search or text comparison. If a recognizable existing work is being reproduced, there is a high risk of infringement.
2.Have you added your own creative contributions (editing, selection, arrangement)?
According to the Zarya decision, only the human-created portions are eligible for copyright protection. The more original creativity, the better.
3.Is the output free of recognizable trademarks, logos, or watermarks?
Getty argues: if the AI model reproduces protected watermarks, this is strong evidence of memorization rather than transformation.
4.Are you familiar with the AI tool's terms of service regarding ownership rights?
Some tools reserve rights to the output or restrict commercial use. Check the terms before publishing.
5.Are you using the output transformatively — in a new context, not as a replacement for an original?
Fair use factor 1: Transformative use (new purpose, new meaning) is more likely to be considered permissible than simple copying.
6.Are you confident the output contains no verbatim copyrighted text passages?
AI models can reproduce training texts verbatim. For longer text outputs, spot-check for known passages.
7.Have you documented your creative process (prompts, editing steps)?
In legal disputes, documented human creativity can make the difference. Keep a record of prompts and post-processing steps.
0 / 7 answered

The distorted Getty watermarks in AI-generated images are legally significant: according to Getty, they demonstrate that the model memorized specific copyrighted images — not merely extracted abstract patterns. This memorization considerably weakens the central "transformative use" argument of AI companies. If a model stores a work precisely enough to reproduce watermarks, it becomes difficult to argue that it merely "transformed" the work. The English High Court saw it differently in November 2025: the model itself does not store copies of the works. Getty prevailed only in a few instances, and on trademark grounds because of the watermarks, not on copyright. Getty has been granted permission to appeal the copyright ruling. The Munich Regional Court took the opposite view, in November 2025 in GEMA v. OpenAI (song lyrics) and in July 2026 in GEMA v. Suno (music): if a model has memorized works, that is a reproduction, and the exception for text and data mining does not cover it. The Munich rulings are not final.

The legal landscape for AI copyright differs fundamentally from country to country. US: Purely AI-generated works receive no copyright protection (Copyright Office 2023; a US appeals court upheld this in 2025, and in 2026 the Supreme Court declined to review it). UK: Offers a limited protection framework for "computer-generated works" — even without human authorship; in March 2026 the government proposed abolishing this protection. EU: Largely unsettled, national courts issue partly contradictory rulings. China: Early court decisions suggest that AI works guided by precise human prompts may be protectable. There is no global consensus — jurisdiction matters enormously.

Note

Key Takeaways

  1. Legal status globally unsettled: Whether AI training on copyrighted data is legal remains disputed worldwide. Neither "clearly legal" nor "clearly illegal" is accurate.
  2. Opt-out has limits: Protection mechanisms exist (robots.txt, EU TDM, platform tags, Glaze) — but they cannot retroactively remove patterns already learned by existing models.
  3. Pure AI works are not protected: Under US law, purely AI-generated works receive no copyright protection. Only the human-created portions of mixed works are protected.

Check

Not completed

What does the US Fair Use Doctrine evaluate when determining whether the use of copyrighted material is permissible?

Select one answer

Answer Key: 1) B · 2) B · 3) B · 4) B · 5) B

Check

Learning Goals

Can you explain this in your own words? Tick what you've got.

0 of 3 done