Can You Train AI on Copyrighted Data?

The training is usually defensible, while the data behind it may get companies sued.

by

Teddy Ellison

AI & Data Agreements

Summary

You can usually train AI on copyrighted material legally in the US and defend the training as fair use. A company that pulled its training data from pirate sites can still face infringement claims, even when the training itself is fair use. If you’re building a product on someone else’s model, be sure you can show where that model’s training data came from.


Courts have mostly treated training an AI model on copyrighted work as fair use, as long as the model transforms what it learns. In 2025, a court ruled that Anthropic's training was fair use. That said, Anthropic still had to pay $1.5 billion — the largest copyright settlement on record. The real risk in training AI is whether a company can show it had the right to use the data, and Anthropic didn’t. 

This isn't just Anthropic's problem. Any company building on AI has to account for its training data, whether it trained the model itself or licensed one built on sources it can't see. Where the data came from decides what the company can promise about its own product. A startup that can't trace its training set has little to stand on when a customer or an investor asks.

Is the risk in the training, or in how the data was collected?

Courts have mostly let the training stand on fair use. They weigh the four factors in Section 107 of the Copyright Act, and a model that transforms what it ingests, rather than reproducing it, has the stronger claim. In the Anthropic case, the district court held in 2025 that training on books was transformative, and no one appealed that ruling.

The $1.5 billion came from somewhere else. Anthropic had pulled millions of books from shadow libraries like LibGen, and the Bartz v. Anthropic settlement, approved in July 2026, resolved those acquisition claims. It covered only Anthropic’s past acquisition and copying of those books, not claims over what its models produce.

The training defense still has a gap. In September 2026, the Third Circuit handed down the first appeals-court ruling on AI training, in Thomson Reuters v. ROSS. It held that copying Westlaw headnotes to build a research tool was not fair use, and found the copying only minimally transformative. But ROSS only retrieved existing text, so the court said nothing about models that generate new work. No appeals court has blessed training a generative model, the kind most startups build on.

What if you didn't train the model yourself?

A company that licenses or fine-tunes someone else's model does not inherit the trainer's piracy liability. That stays with whoever did the copying. But it can still infringe through the model's output. A model can reproduce a copyrighted work in its output, and the authors and publishers suing OpenAI argue that infringes. That claim cleared a motion to dismiss and reached summary judgment in 2026, with no ruling yet on whether such output infringes.

Companies facing an output claim usually turn first to their vendor’s copyright indemnity. OpenAI, Anthropic, Google, and Microsoft all offer one on business tiers, but each typically covers only the model's unaltered output. Normal use generally keeps the coverage, though some vendors, such as Microsoft, require their content filters to stay on. But a claim that traces back to the customer’s prompts, fine-tuning data, edited output, or added tools usually falls outside it. 

For who owns what a model produces once the copyright question is settled, see our guide to who owns AI-generated content.

Settle where your training data came from before you have to prove it

The rules on training data are still being written, but it's safer to start documenting now. When dealing with regulations and rules that are still making their way through the courts and legislative system, it’s often better for founders to take a more conservative, “better safe than sorry” approach to cover future bases.

When it comes to managing your own training data, businesses can do most of the groundwork themselves without involving a lawyer. At its core, this requires establishing a record of where every piece of data came from, which should be detailed enough to hand to a skeptical reviewer. Be sure to:

  • Understand what you’ve shipped: The company that trained or fine-tuned the model answers for how the data was collected. A company that only licenses and deploys a model mostly inherits the output risk.

  • Trace the training data: Data you own, licensed, or took from the public domain is clean. Scraped or unknown-source data carries the exposure, so document what you can and flag what you can't.

  • Pin down the vendor’s coverage: Get the copyright indemnity in writing and read its conditions against how you use the model, including any fine-tuning that could drop the coverage.

  • Check which disclosure rules reach you: If you build or substantially modify a generative AI system offered to Californians, California’s AB 2013 requires you to post a high-level summary of its training data. The rule took effect January 1, 2026, and reaches systems released since 2022. The EU and China have their own training-data rules, covered in the FAQ below. 

  • Document everything before diligence: An investor or an acquirer will ask you to warrant your IP, and a provenance record built in advance lets you answer without renegotiating the deal.

With this record in hand, the company can confidently warrant its IP in a financing or a sale. The harder part is turning those answers into contract terms that survive a close read. 

To that end, legal counsel can write the training-data provenance into the vendor’s representations and warranties and find where its indemnity stops covering the company. And it can secure clean, licensed data early, so the product’s training set never becomes the reason a financing or a sale stalls or gets repriced. Serotonin Legal has extensive experience working with founders building on AI, and can give you a clear read on where your product stands.


Reach out to set up a free consultation.


Serotonin Legal advises technology founders on corporate, regulatory, and transactional matters at the intersection of AI, blockchain, and fintech. This guide is for general informational purposes and does not constitute legal advice. No attorney-client relationship is formed by reading this material.

FAQs

Can you train AI on your own licensed data?

Yes, and it’s the safest route. Data you own or have properly licensed carries none of the acquisition exposure that drove the Anthropic settlement. The license still has to permit AI training and model redistribution, which many don't cover by default. Confirm those rights before you build.

Who is liable if an AI model was trained on copyrighted data, the vendor or the customer?

Both, in different ways. The company that trained the model answers for how it collected the data. In the Kadrey v. Meta case, the court cleared Meta on training but kept alive a distribution claim over the books it torrented. A company that only deploys a model faces the separate risk of infringing output. Anthropic’s settlement covered only its own past copying, so a deployer facing an output claim would look to the vendor’s copyright indemnity instead.

What does an AI vendor's copyright indemnity not cover?

It excludes your own inputs and any data you fine-tune on, output you modified or combined with non-vendor tools, and anything on a free or beta tier. Even output you never edited can fall outside it if the claim traces back to your fine-tuning data. The indemnity applies to the base model’s unaltered output and little more, so read the exclusions against how your product is built.

Is training AI on copyrighted material legal outside the US?

It depends on the market. Because copyright is territorial, these rules mostly govern training done inside each country. For instance, Japan lets companies train on it even for commercial purposes, and a company falls outside that rule if it trains to reproduce the works’ creative expression or unreasonably harms the copyright owner’s interests. On the other hand, the UK allows it only for non-commercial research, so a company training a commercial model there generally needs a license. 

Some governments go further and regulate the training data behind AI products offered in their countries. The EU allows training on lawfully accessed works unless the owner opts out. Its AI Act then requires providers of general-purpose AI models offered in the EU to keep a policy for honoring those opt-outs, wherever the model was trained. China requires generative AI services offered to its public to use lawfully sourced training data, and Brazil is weighing training-data rules of its own. So a US startup that trains at home can still pick up training-data duties in the markets it sells into.

What does the US Copyright Office say about training AI on copyrighted works?

In its 2025 report, the Copyright Office warned that a model's weights can raise a copyright problem of their own, once the model has memorized its training data. Companies using it downstream could inherit the problem. It is guidance the Office hasn't finalized, and no court is bound by it.

Curious to learn more about Serotonin Legal? —

Get in Touch

Can You Train AI on Copyrighted Data?

The training is usually defensible, while the data behind it may get companies sued.

by

Teddy Ellison

AI & Data Agreements

Summary

You can usually train AI on copyrighted material legally in the US and defend the training as fair use. A company that pulled its training data from pirate sites can still face infringement claims, even when the training itself is fair use. If you’re building a product on someone else’s model, be sure you can show where that model’s training data came from.


Courts have mostly treated training an AI model on copyrighted work as fair use, as long as the model transforms what it learns. In 2025, a court ruled that Anthropic's training was fair use. That said, Anthropic still had to pay $1.5 billion — the largest copyright settlement on record. The real risk in training AI is whether a company can show it had the right to use the data, and Anthropic didn’t. 

This isn't just Anthropic's problem. Any company building on AI has to account for its training data, whether it trained the model itself or licensed one built on sources it can't see. Where the data came from decides what the company can promise about its own product. A startup that can't trace its training set has little to stand on when a customer or an investor asks.

Is the risk in the training, or in how the data was collected?

Courts have mostly let the training stand on fair use. They weigh the four factors in Section 107 of the Copyright Act, and a model that transforms what it ingests, rather than reproducing it, has the stronger claim. In the Anthropic case, the district court held in 2025 that training on books was transformative, and no one appealed that ruling.

The $1.5 billion came from somewhere else. Anthropic had pulled millions of books from shadow libraries like LibGen, and the Bartz v. Anthropic settlement, approved in July 2026, resolved those acquisition claims. It covered only Anthropic’s past acquisition and copying of those books, not claims over what its models produce.

The training defense still has a gap. In September 2026, the Third Circuit handed down the first appeals-court ruling on AI training, in Thomson Reuters v. ROSS. It held that copying Westlaw headnotes to build a research tool was not fair use, and found the copying only minimally transformative. But ROSS only retrieved existing text, so the court said nothing about models that generate new work. No appeals court has blessed training a generative model, the kind most startups build on.

What if you didn't train the model yourself?

A company that licenses or fine-tunes someone else's model does not inherit the trainer's piracy liability. That stays with whoever did the copying. But it can still infringe through the model's output. A model can reproduce a copyrighted work in its output, and the authors and publishers suing OpenAI argue that infringes. That claim cleared a motion to dismiss and reached summary judgment in 2026, with no ruling yet on whether such output infringes.

Companies facing an output claim usually turn first to their vendor’s copyright indemnity. OpenAI, Anthropic, Google, and Microsoft all offer one on business tiers, but each typically covers only the model's unaltered output. Normal use generally keeps the coverage, though some vendors, such as Microsoft, require their content filters to stay on. But a claim that traces back to the customer’s prompts, fine-tuning data, edited output, or added tools usually falls outside it. 

For who owns what a model produces once the copyright question is settled, see our guide to who owns AI-generated content.

Settle where your training data came from before you have to prove it

The rules on training data are still being written, but it's safer to start documenting now. When dealing with regulations and rules that are still making their way through the courts and legislative system, it’s often better for founders to take a more conservative, “better safe than sorry” approach to cover future bases.

When it comes to managing your own training data, businesses can do most of the groundwork themselves without involving a lawyer. At its core, this requires establishing a record of where every piece of data came from, which should be detailed enough to hand to a skeptical reviewer. Be sure to:

  • Understand what you’ve shipped: The company that trained or fine-tuned the model answers for how the data was collected. A company that only licenses and deploys a model mostly inherits the output risk.

  • Trace the training data: Data you own, licensed, or took from the public domain is clean. Scraped or unknown-source data carries the exposure, so document what you can and flag what you can't.

  • Pin down the vendor’s coverage: Get the copyright indemnity in writing and read its conditions against how you use the model, including any fine-tuning that could drop the coverage.

  • Check which disclosure rules reach you: If you build or substantially modify a generative AI system offered to Californians, California’s AB 2013 requires you to post a high-level summary of its training data. The rule took effect January 1, 2026, and reaches systems released since 2022. The EU and China have their own training-data rules, covered in the FAQ below. 

  • Document everything before diligence: An investor or an acquirer will ask you to warrant your IP, and a provenance record built in advance lets you answer without renegotiating the deal.

With this record in hand, the company can confidently warrant its IP in a financing or a sale. The harder part is turning those answers into contract terms that survive a close read. 

To that end, legal counsel can write the training-data provenance into the vendor’s representations and warranties and find where its indemnity stops covering the company. And it can secure clean, licensed data early, so the product’s training set never becomes the reason a financing or a sale stalls or gets repriced. Serotonin Legal has extensive experience working with founders building on AI, and can give you a clear read on where your product stands.


Reach out to set up a free consultation.


Serotonin Legal advises technology founders on corporate, regulatory, and transactional matters at the intersection of AI, blockchain, and fintech. This guide is for general informational purposes and does not constitute legal advice. No attorney-client relationship is formed by reading this material.

FAQs

Can you train AI on your own licensed data?

Yes, and it’s the safest route. Data you own or have properly licensed carries none of the acquisition exposure that drove the Anthropic settlement. The license still has to permit AI training and model redistribution, which many don't cover by default. Confirm those rights before you build.

Who is liable if an AI model was trained on copyrighted data, the vendor or the customer?

Both, in different ways. The company that trained the model answers for how it collected the data. In the Kadrey v. Meta case, the court cleared Meta on training but kept alive a distribution claim over the books it torrented. A company that only deploys a model faces the separate risk of infringing output. Anthropic’s settlement covered only its own past copying, so a deployer facing an output claim would look to the vendor’s copyright indemnity instead.

What does an AI vendor's copyright indemnity not cover?

It excludes your own inputs and any data you fine-tune on, output you modified or combined with non-vendor tools, and anything on a free or beta tier. Even output you never edited can fall outside it if the claim traces back to your fine-tuning data. The indemnity applies to the base model’s unaltered output and little more, so read the exclusions against how your product is built.

Is training AI on copyrighted material legal outside the US?

It depends on the market. Because copyright is territorial, these rules mostly govern training done inside each country. For instance, Japan lets companies train on it even for commercial purposes, and a company falls outside that rule if it trains to reproduce the works’ creative expression or unreasonably harms the copyright owner’s interests. On the other hand, the UK allows it only for non-commercial research, so a company training a commercial model there generally needs a license. 

Some governments go further and regulate the training data behind AI products offered in their countries. The EU allows training on lawfully accessed works unless the owner opts out. Its AI Act then requires providers of general-purpose AI models offered in the EU to keep a policy for honoring those opt-outs, wherever the model was trained. China requires generative AI services offered to its public to use lawfully sourced training data, and Brazil is weighing training-data rules of its own. So a US startup that trains at home can still pick up training-data duties in the markets it sells into.

What does the US Copyright Office say about training AI on copyrighted works?

In its 2025 report, the Copyright Office warned that a model's weights can raise a copyright problem of their own, once the model has memorized its training data. Companies using it downstream could inherit the problem. It is guidance the Office hasn't finalized, and no court is bound by it.

Curious to learn more about Serotonin Legal?

Get in Touch