When your work is used to train AI

You see a dataset listing that seems to include your work, or an AI output that feels uncomfortably familiar. You want to know what happened and whether you had any say in it. Start with what you can show, then decide what you want.

The U.S. Copyright Office's main analysis of this question is Part 3 of its report on copyright and AI. The Office released it as a pre-publication version on May 9, 2025, and says the final version will not change the analysis or conclusions. 3 The report is the Office's view, not a court ruling on your work. 1

Know which acts can involve your copyright

Training is several acts, and each one raises its own question. 1

• Collecting the data. Downloading, converting and filtering works to build a dataset makes copies, and the Office says this clearly involves the reproduction right, even if the copies are later deleted. 1
• Training the model. Training makes further copies. The trained model itself copies your work only if it has memorized substantial protected expression from it, for example when it can reproduce the work nearly verbatim. 1
• Outputs. An output that replicates or closely resembles your work likely infringes, separately from how the model was trained. 1

Keep these apart when you describe your concern. Evidence about an output does not prove what was in a dataset, and a dataset listing does not prove that an output copies you.

Separate what you saw from what you suspect

1. Save the page, dataset entry, output or message that raised your concern, with the date, the address and how you found it. If you ran a prompt, save the prompt and the output together. Saving evidence walks through how.
2. Keep your original files and the records that show you own the rights.
3. Write a short account that separates observation from inference. "This index names my file" is an observation. "I believe this model was trained on my file" is an inference. Platforms, lawyers and collaborators can act on the first; the second needs more evidence.

A similar style does not prove your work was used or copied. Copyright protects your expression, not ideas, methods or concepts, and the Office notes that style as such is not protected; a close reproduction of your actual work is a different matter. 21 When two things look alike explains how similarity is judged.

Know what decides whether a training use is lawful

In the United States the question is fair use, and a court weighs four factors: the purpose and character of the use, the nature of your work, how much was used, and the effect on the market for your work. 2 The Office expects the first and fourth factors to carry the most weight. 1

The Office's view, from its report: 1

• Some training uses will be fair and some will not. Noncommercial research or analysis that does not let the model reproduce the works is likely fair.
• Copying expressive works from pirate sources to generate content that competes with them, when a license was reasonably available, is unlikely to be fair.
• Knowingly training on a dataset of pirated or illegally accessed works weighs against fair use.
• Market harm counts, including lost licensing income and outputs in your style that compete with your work.

So the facts that matter most are where the developer got your work, what the model is for, and whether its outputs compete with you. Look for those facts first. How fair use works covers the factors in more depth.

Protect your position before you act

1. Register the work. For a U.S. work, you cannot sue for infringement until the Copyright Office has registered your claim or refused it. Statutory damages and attorney's fees are available only for infringement that began after registration, unless a published work was registered within three months of first publication. 4 Your first U.S. copyright application shows you how to file.
2. Note the deadline. A civil copyright claim must be filed within three years after it accrued. 5
3. Get legal advice before a demand or a lawsuit. Before you send a demand letter or file suit, have a copyright litigator review your registration, your evidence and that three-year deadline.

Block future crawling where you control the site

If you run your own website, add rules to its robots.txt file for AI-training crawlers:

• Google-Extended controls whether Google may use your pages to train future Gemini models. Blocking it does not affect your inclusion or ranking in Google Search. 6
• GPTBot signals that OpenAI should not use your content to train its foundation models. OpenAI's search crawler is a separate setting, so you can stay in its search results while blocking training. 7

robots.txt is a request that crawlers are asked to honor, not access control. 8 It only works on sites you control, and it only affects future crawling. Commenters told the Copyright Office that no current technology removes entire works at scale from a model after it has been trained. 1

Some services also train on what users post or type unless the user opts out. 1 Open the data or privacy settings of each service where you post work and choose the setting you want.

Ask for one specific thing

Decide what you want before you contact a developer, dataset host or platform, and say it in one sentence:

• Remove a hosted copy. Ask the host to take your file out of the dataset. This does not undo past training. How platform reports and takedowns work covers the process.
• Stop future use. Ask the developer to exclude your work from future training, and ask how they will confirm it.
• License the use. Licensing for AI training is growing in some sectors. 1 If someone offers you a training license, read which works, systems, purposes, downstream uses, payment, reporting and future rights it covers before you sign. Do not accept promises of "deletion" or "unlearning" that the agreement does not define.
• Claim compensation for past use. That is a legal claim, so take the three steps under "Protect your position before you act" first.

You do not have to decide at the start whether to sue, license or let it go. Save the evidence, register, block what you can, and ask for one thing at a time.