Is AI training fair use, and what does it mean for licensed business data?

The short answer

It depends, and no single ruling settles it. Under US copyright law, fair use is a defense that a court weighs for a particular use of particular works, so whether training an AI model is fair use gets decided case by case, on the facts of each lawsuit. As of October 2026, AI developers, authors, publishers and other rights holders are still arguing those facts in court.

For business records licensed under a signed agreement, the question mostly falls away. A buyer that licenses a company's support tickets and project files for an agreed term relies on permission, not on a defense. What remains is a narrower question: does the company granting the license actually hold the rights it is granting?

What does fair use weigh?

Fair use allows some unlicensed uses of copyrighted works. Courts consider four factors together, and no single factor decides the outcome. The US Copyright Office's Copyright and Artificial Intelligence initiative is where the agency publishes its analysis of how those factors may apply to training.

The fourth factor is where licensing enters the legal debate. Whether rights holders and developers already license training material to each other is part of the market-effect question.

What has the US Copyright Office said?

The Office is publishing a multi-part report on copyright and AI. Part 3, which covers generative AI training, appeared in May 2025 as a pre-publication version. It examines which steps in building a model may involve copying that implicates copyright, how fair use may apply, and whether licensing approaches are practical. It also observes that a model's performance depends heavily on the quality of its data.

Two cautions before quoting it. The report is the agency's analysis, not law, and courts are not bound by it; judges decide fair use. And because it was released in pre-publication form, check the Office's site for the current version. This page deliberately does not summarize the report's conclusions; read them in the original or ask counsel.

Where do the US court cases stand?

This page does not report individual rulings, because we have not verified them against the court opinions and the picture keeps shifting as cases move through appeals, settlements and new filings. Ask counsel for the current status of any case, and read any ruling through five questions rather than relying on a headline.

The ruling reader:

1. Which court decided it? A federal trial court decision does not bind other courts. An appeals court decision binds trial courts in its circuit, and only the Supreme Court sets a nationwide rule.
2. At what stage? A ruling on a motion to dismiss, a summary judgment decision and a jury verdict carry different weight, and a settlement sets no precedent at all.
3. What was copied, and how was it obtained? Where the training copies came from can matter as much as what the training did with them.
4. What does the model produce? Outputs that reproduce or substitute for the original works raise different questions from outputs that do not.
5. Is there a licensing market for that material? Evidence of willing licensors and paying licensees feeds straight into the market-effect factor.

For the commercial side of that last question, AI data licensing trends for 2026 tracks how the market is developing.

Why do licensed business records sidestep the question?

Because the buyer has written permission for an agreed scope, and because the records were never public to begin with.

Three points explain the gap:

• Fair use does not open doors. Even a ruling broadly favorable to training would give nobody access to a company's email, Slack or Teams threads, CRM history or ticket queue. Those records sit behind company systems, and the practical route to them runs through the company.
• Copyright is only one right in play. Many operational records are largely factual, and copyright protects original expression, so confidentiality duties, contracts and privacy law often matter as much as copyright for this material.
• Buyers value a paper trail. A signed license documents what the buyer may do, for how long and under which de-identification rules, which matters to developers that need a clear record of where their training data came from.

Developers have signed licenses even while the fair-use question stays open. The Associated Press reported in July 2023 that it would license part of its text archive, dating back to 1985, to a major AI developer. Business records add something news archives lack: the exceptions, escalations and decisions behind real work, which is why AI models need rare examples. For the full map of sources, see where AI training data comes from.

Does the company actually hold the rights it licenses?

This is where the real work sits, because a license is only as good as the licensor's rights.

• Staff-created material. The Copyright Office's circular on works made for hire explains that the employer, not the employee, is treated as the author of material prepared as part of the job. Freelance material is treated that way only in listed categories with a signed written agreement, so contractor work needs checking.
• Partial rights. Copyright can be divided: 17 U.S.C. section 201 allows any exclusive right to be transferred and owned separately. That is why a company can grant AI-training rights for a set term while keeping ownership of its records.
• Third-party material. Purchased research, vendor manuals, licensed images and files supplied by clients arrive with their own terms and are normally carved out unless the owner agrees.

How does this play out in common partner situations?

Disclosure and consent good practice

Partners make introductions and share basic fit information. They do not give legal opinions or handle records.

• Do not offer a view on whether a client's records would be fair game without a license; point to the company's counsel and the rights review.
• Never forward records, screenshots or samples to show what a company holds.
• Tell the company you may earn a referral reward; licensed professionals should check their own rules on referral fees and disclosure.
• Expect the company to approve scope, de-identification and redaction before any work starts; nothing is delivered without an executed agreement and the company's authorization.

If a client raises the lawsuits, a short answer works:

Questions to ask counsel

The company's counsel, or yours if you advise the company:

1. Which of our records include third-party copyrighted material, and how should it be excluded?
2. Do our contractor, agency and freelancer agreements assign work product to us in writing?
3. Do customer contracts, employee notices or our privacy policy limit using records for AI training?
4. What rights and warranties would we give in a license, and how is liability limited?
5. Does any recent AI-copyright decision change our analysis, and which cases should we watch?
6. For partners who hold a professional license: does my professional body restrict referral fees, and what must I disclose?

This is general information, not legal, tax or financial advice. Confirm with your own counsel, tax adviser or professional body before acting.

Where does this leave a referral partner?

You do not need a view on fair use to make a useful introduction. The better question for a client is practical: did the company create its records, and does it control the systems they sit in? For the other legal layers, read whether it is legal to license business data to AI developers; for how the public deals were structured, see enterprise AI data licensing deals.

Partners earn 25% of the eligible platform fees SourceX actually collects from the referred company's licensing deals, up to $100,000 per referred company, payable only after the buyer pays and SourceX receives its fee. No reward is guaranteed, and the reward is never deducted from what the company receives.

Next step

If a client has years of records it created itself, run it through the company fit checker and read how it works. Then register as a partner to make the introduction, or have the company apply directly at sourcex.si/apply.