AI’s Fair-Use House of Cards Is Starting to Fall

I made this bet already.

I said the artificial-intelligence industry's fair-use defense would eventually run into a wall because these companies stopped acting like researchers and started acting like owners.

Now the wall is coming into view.

The U.S. Court of Appeals for the Third Circuit has upheld a ruling against ROSS Intelligence, an AI legal-research company that used copyrighted Westlaw headnotes to help train a competing legal-search product.

ROSS called it fair use.

The court called it copyright infringement.

And somewhere inside every major AI company, a lawyer just sat up a little straighter.

You Cannot Call Everything Research Forever

Fair use is one of the most important protections in American copyright law.

It allows people to quote, criticize, report, teach, parody, study, and research copyrighted works without asking permission every single time. Without fair use, powerful people could use copyright to suppress criticism and control public discussion.

But fair use is not a magic phrase that turns taking into sharing.

Courts weigh four factors:

Why and how the copyrighted material was used
What kind of work was copied
How much was taken
Whether the new use harms the market for the original

Commercial use does not automatically lose. Research does not automatically win. Copying an entire work is not automatically illegal, and changing something is not automatically transformative.

The question is what the defendant actually did.

That is where the AI industry's story starts falling apart.

OpenAI Cannot Have It Both Ways

The defense goes something like this:

We are not stealing anyone's work. We are conducting research. The model learns from books, journalism, art, music, photography, code, and websites the way a human learns from reading.

That argument was easier to sell when these systems looked like laboratory experiments.

Then the experiments became products.

The products became subscriptions.

The subscriptions became enterprise contracts, platform deals, API businesses, advertising opportunities, and some of the largest corporate valuations in history.

Now the same companies want us to believe they are merely reading the internet for educational purposes.

No.

You do not get to crawl the work of humanity under the banner of research, lock the resulting system behind a commercial interface, charge the public for access, compete with the people whose work trained it, and then return to the courtroom wearing a lab coat.

At some point, research became product development.

At some point, the public experiment became private infrastructure.

At some point, “learning” became a business model.

That point matters.

The ROSS Decision Is the Crack

ROSS wanted to build a legal-research platform that could compete with Westlaw.

To train it, contractors created thousands of legal memoranda using Westlaw headnotes—short editorial summaries written by Thomson Reuters editors. ROSS did not merely study the history of legal headnotes. It used those headnotes to improve a product intended to serve the same legal-research market.

The Third Circuit did not buy the claim that inserting an AI-training step automatically made the use transformative.

The court found that ROSS and Westlaw used the material for substantially the same ultimate purpose: helping customers find responsive legal information. ROSS copied entire headnotes, used them to develop a competing platform, and interfered with both Westlaw's market and a developing market for licensing headnotes as AI-training data.

That last part should terrify the AI industry.

The court recognized that a market for licensing copyrighted material as training data can matter—even when the copyright owner has not already licensed that material widely.

Read that again.

AI companies have spent years arguing that creators cannot claim market harm from a licensing market the companies themselves refused to recognize.

The Third Circuit just upheld the opposite logic in this case: taking the material without authorization can usurp the owner's opportunity to license it for AI training.

That does not decide every generative-AI lawsuit in America.

But it puts a live round in the chamber.

“But ROSS Wasn't Generative AI”

Correct.

ROSS did not generate novels, images, songs, jokes, or articles. It retrieved passages from judicial opinions. The Third Circuit specifically distinguished ROSS from generative models capable of producing new expression.

The case is also an interlocutory appeal, meaning other parts of the litigation remain unresolved. It binds courts in the Third Circuit, not every court in the country. It does not automatically decide the lawsuits against OpenAI, Anthropic, Meta, or every other model developer.

Congratulations. The house of cards did not collapse in one afternoon.

It only lost a support beam.

The importance of the ruling is not that ROSS and ChatGPT are technically identical. They are not.

The importance is that a federal appellate court refused to treat “AI training” as a special legal category that floats above ordinary copyright law.

The court looked past the technology and asked old-fashioned questions:

What did you copy?

Why did you copy it?

How much did you take?

Did you build a competitor?

Did you damage the owner's market?

That framework is exactly where the biggest AI companies do not want this conversation to go.

Transformative Is Not a Get-Out-of-Licensing Card

AI companies love the word transformative.

They say models do not store and resell books. They extract patterns, relationships, styles, facts, and language. They produce something new.

Sometimes that argument may win.

But transformation cannot mean that the biggest copier receives the broadest immunity.

If I take one article and resell it, everyone understands the problem. If a corporation takes millions of articles, mixes them into a model, and sells access to an answer engine that reduces the need to visit the original publishers, suddenly we are told the scale of the copying makes it innovation.

That is backwards.

Scale should not erase consent.

Complexity should not erase ownership.

A sufficiently expensive machine should not acquire rights that an ordinary person does not possess.

And a company should not be allowed to destroy a licensing market, then point to the absence of that market as proof that no market was harmed.

The Market Harm Is Not Hypothetical

Generative AI does not operate in a separate universe from the work used to build it.

AI answers can replace visits to news sites.

AI images can replace commissioned illustrations.

AI summaries can replace research subscriptions.

AI-written marketing can replace freelance writers.

AI music can compete with musicians.

AI code can compete with the developers whose repositories helped train the system.

The companies call their models general-purpose tools. Rights holders see products entering the same markets that financed the training material in the first place.

That is not merely learning.

That is vertical integration without permission.

First, take the raw material.

Then build the machine.

Then sell the substitute.

Then tell the supplier that compensation would stop progress.

OpenAI Knows Fair Use Is Load-Bearing

OpenAI has publicly argued that training on publicly available internet material is fair use. It has warned policymakers that restricting access to copyrighted training material could damage American competitiveness and the AI race.

Of course it says that.

Fair use is not a minor defense for this industry. It is the foundation under the foundation.

If courts require licenses for large portions of training data, the cost of building frontier models changes. The competitive advantage of companies that copied first and asked questions later changes. The value of the data changes. The bargaining power of writers, artists, publishers, musicians, programmers, and other creators changes.

The entire industry was financed under the assumption that the largest input could be acquired at close to zero cost.

That input was human work.

This is why I called it a house of cards.

The models may be technically impressive. The products may be useful. The companies may continue operating and negotiating licenses.

But usefulness does not retroactively create permission.

“Publicly Available” Does Not Mean Public Domain

This trick deserves its own section.

A work being visible on the internet does not mean nobody owns it.

A newspaper article can be publicly accessible and copyrighted.

A photograph can appear in search results and remain copyrighted.

A song lyric can be posted on a website and remain copyrighted.

A book can be downloadable from an unauthorized source and remain copyrighted.

Publicly available describes access.

Public domain describes legal status.

Those are not the same thing, no matter how badly a data-hungry company needs them to be.

The Lawsuits Are Converging

ROSS is not the perfect generative-AI test case. Cleaner cases are coming.

Authors have sued over books.

Newspapers have sued over journalism.

Music publishers have sued over lyrics.

Artists have sued over images.

Rights holders are challenging both the copying that occurs during training and outputs that allegedly reproduce or substitute for protected work.

Anthropic is fighting music publishers over the use and reproduction of copyrighted song lyrics. OpenAI and Microsoft continue fighting publishers and authors. Different judges have already reached conclusions favorable to AI companies in some cases and favorable to copyright owners in others.

That split is not evidence that nothing is happening.

It is what the beginning of a legal reckoning looks like.

The law is testing where research ends, where transformation begins, when a product becomes a substitute, and whether an AI-training license is a market copyright owners have the right to control.

ROSS gives rights holders an appellate ruling they can cite.

The AI companies can spin it as narrow.

Narrow cracks still spread.

Pay the People Who Built the Intelligence

I am not arguing that artificial intelligence should be banned.

I am arguing that the people whose work made it possible should not be treated like an inconvenient cost center.

Build licensing systems.

Create collective rights organizations.

Offer meaningful opt-in agreements.

Pay publishers, artists, musicians, filmmakers, photographers, comedians, researchers, and independent websites.

Create public-interest research exceptions that are genuinely noncommercial.

Open training records to independent audits.

Stop hiding behind the word research after the research becomes a product.

The industry wants credit for building the future while refusing to pay the people who supplied the past.

That arrangement was never stable.

I Made the Bet. Now Watch the Cards Move.

I said OpenAI and the rest of the AI industry would lose the moral force of the research defense once they turned mass copying into a commercial empire.

To be legally exact, commercialization alone does not destroy fair use.

But let us be honest about what happened.

They scraped first.

They scaled second.

They sold third.

They offered deals to selected rights holders only after lawsuits and public pressure made universal permission look expensive.

Now courts are beginning to ask whether the word transformative can carry all that weight.

ROSS does not end the debate.

It proves the debate is real.

The AI industry built one of the most valuable technological structures in history on an unresolved legal assumption: that it could ingest copyrighted culture without negotiating with the people who created it.

That assumption is finally reaching appellate courts.

I made my bet.

The house is still standing.

But the cards are moving.