Is training AI on copyrighted data fair use? Where US courts stand in 2026

Is AI training fair use?#

AI training is neither automatically fair use nor automatically infringement under US law; courts decide each case on its facts using the four fair use factors in the Copyright Act. Results so far have differed with the type of work, the way it was obtained, what the trained system does and the evidence offered about market harm.

That uncertainty is the main point for business leaders. A developer relying on fair use accepts litigation risk for each body of material it uses, while a developer using licensed material has a contract that states what it may do.

How the four fair use factors apply to AI training#

The four factors are weighed together rather than scored one by one, and in AI cases the first and fourth have done most of the work. The table shows what courts ask under each factor and how the argument has generally run when the use is AI training.

The phrase market for licensing training data matters for anyone who owns records. The more often material is licensed for AI use, the easier it becomes for a rights holder to argue that unlicensed copying takes away a real market rather than creating a new use.

Where US courts stand in 2026#

US courts in 2026 have split on outcomes but agree that facts decide: the cases that have produced rulings involve published books and editorial content, not private business records. The table summarizes the most-cited decisions; read the opinions and current counsel commentary before relying on any one of them.

District-court rulings bind only the parties and can be appealed, so headlines about a win or loss often overstate how far a ruling reaches. The US Copyright Office added a non-binding view in its May 9, 2025 pre-publication report on generative AI training: it said commercial use of vast troves of works to produce competing content, especially through illegal access, goes beyond established fair use boundaries, and it recommended letting voluntary licensing markets keep developing. Many other cases over news, music, images and code remain pending.

Acquisition and use are now separate questions#

Courts have begun to treat how a developer obtained material as a separate question from how it used the material. A use that might be defended as fair can sit alongside a claim over copies downloaded from unauthorized sources, and reported settlements in the book cases centered on exactly those copies.

For buyers, the split shifts attention to sourcing: where each dataset came from and what permission covered it. For suppliers, it means documentation of rights is part of what makes a dataset usable, alongside the content itself.

Why fair use rarely decides private business records#

Fair use rarely decides questions about private business records because those records were never published. A developer cannot lawfully collect a company's support tickets, dispatch notes or code reviews from the open web, so the practical question is access, and only the company can grant it.

Other laws also protect operational records. Trade secret law, confidentiality agreements with customers and employees, privacy laws such as CCPA where they apply, and vendor terms all shape what can be shared, and none of them is resolved by a fair use ruling about books or news articles.

Much of what makes operational records useful is factual: part numbers, timestamps, defect codes and delivery windows. Copyright protects expression, not facts, which is another reason contracts rather than copyright usually define these deals.

What founders and general counsel can do now#

Founders and general counsel can act now without waiting for the law to settle, because the most useful steps protect the company whatever courts decide. Each step below either preserves a protection or documents a right.

• 1Treat public website content, help centers and documentation as already exposed to crawling, and decide whether to state AI terms for it.
• 2Keep private records private through access controls, confidentiality terms and clear internal policies.
• 3License records only under a written agreement that defines permitted use, term, deletion and audit.
• 4Document provenance and rights for anything you license, before the buyer asks.
• 5Revisit your position when appellate courts rule, since their decisions can change the risk picture for developers.

Illustrative: an architecture firm separates public from private#

Illustrative: the managing principal of a fictional architecture firm reads that courts have sided with some AI developers on fair use and asks two questions. Were the firm's public project pages and design essays used to train models, and does that mean its private project records are fair game too?

Counsel separates the answers. The public pages may well have been crawled, and the firm can add AI terms and crawler controls going forward, with limited leverage over past copying. The private records, years of RFIs, submittal reviews and internal design review notes in Procore and Bluebeam, were never public, and nobody can use them without the firm's permission.

The firm updates its website terms, leaves the fair use question with counsel, and runs a metadata-only fit check on its RFI and submittal history. Client-owned drawings and deliverables are flagged for exclusion from the start.

How SourceX approaches fair use questions#

SourceX does not rely on fair use for anything it handles. Every package is a licensed transaction under a written agreement, records are licensed rather than sold outright, and the SourceX Evidence Packet documents provenance, licensing rights, permitted use, the privacy record and release authorization.

Within the SourceX Enterprise Data Value Framework, reproducibility reduces value and rights increase it. Records anyone could scrape are reproducible; private operational records with documented rights are not, whatever courts decide about published works.