AI Training Data Licensing Is a Real Market Now | Elacity
The first U.S. appeals ruling on AI training treats licensing data to AI as a real market. Enforcing it took Thomson Reuters six years. Owners need a gate instead.
AI Training Data Licensing Is a Real Market Now. Selling Into It Takes a Gate.
Everything useful you have ever written, labelled, or summarised is worth something to a machine, and until this week no American appeals court had said whether that worth was yours to sell. On September 29, one did. AI training data licensing is now a market a federal appeals court recognises, and the open question is whether you have any way to sell into it.
The Third Circuit upheld Thomson Reuters' win against ROSS Intelligence, rejecting ROSS's claim that training its legal search engine on Westlaw headnotes was fair use. Reuters called it the first ruling of its kind from a U.S. appeals court in the wave of AI training cases.
What the Court Actually Decided
The detail that matters most for owners sits in the fourth fair-use factor, market harm. MLex reports the panel found that ROSS's copying devalued Thomson Reuters' data, including in the potential market for licensing that content as AI training material.
Read that slowly. A court treated "licensing your work so a machine can learn from it" as a market in its own right, one a taker can damage. That is the legal shape of a thesis we keep returning to: when AI drives the cost of labour toward zero, value moves to whoever owns the inputs the machine consumes.
The ruling is also narrower than the headlines. ROSS built a direct competitor, and the court put competition at the centre of its reasoning. Commentators note the panel distinguished generative AI cases, where a district judge in 2025 found that training on lawfully purchased books was fair use. The law is splitting by facts, not settling.
The Failure of Enforcement by Lawsuit
Look at how this right was actually defended. ROSS asked to license Westlaw content, was refused, and obtained training material derived from the headnotes through a third-party contractor instead.
Thomson Reuters sued in 2020. The district court rejected fair use in February 2025. The appeal landed in September 2026, against a startup that is already defunct.
Six years, a global information company's legal budget, and a defendant that no longer exists. The market for your data is real on paper. Defending it the way Thomson Reuters did is a privilege available to very few owners, and the copying finished long before the verdict.
For an independent researcher, a small publisher, or a photographer, a recognised market with no way to control access is an IOU written in case law. The court can say a taking hurt you. It cannot stop the next one.
AI Training Data Licensing Needs a Gate, Not Only a Verdict
The fix is to move the license from the courtroom to the moment of access. That is precisely what Elacity dDRM does, and it is why we build it: Elacity exists to turn your data into capital you control, and ElastOS is the open-source runtime that enforces it on a machine you own. Our decentralized DRM explainer covers the architecture; here is how it applies to a training-data license.
1. The dataset becomes a Wealth Capsule with its terms inside
You package a corpus, a set of annotations, or a body of notes into an encrypted, programmable on-chain good. Who may use it, for what, and the royalty it carries are written into the asset, not into a PDF you hope someone reads. Our guide to turning your data into capital walks through the packaging step.
2. A quorum checks the license before any key exists
The key that unlocks the content is split across an owned 2-of-3 quorum of independent machines. No single operator holds it, Elacity included, and each machine re-checks the buyer's on-chain rights before releasing its share. No license, no key. The ROSS route, getting the material from someone else who already had it, runs into a lock instead of a lawsuit.
3. Use happens in a sealed moment, never as a handover
Content stays encrypted everywhere except inside a sealed sandbox at the moment of use. The key exists in the clear only for that split second, welded to that one transaction, then wiped. Access is a narrow, expiring capability: revoke it and it stops mid-action, and the system fails closed. The sealing is post-quantum-hybrid today, so data licensed now is not waiting to be decrypted later.
What a Gate Cannot Do
Honesty matters more than a slogan here. A gate controls access; it cannot make a model forget what a paying licensee legitimately trained on. Once you sell a use, that use happened. What changes is that taking without paying now requires breaking a split key rather than downloading a file.
Two more edges, stated plainly. The quorum is trust-minimised, not trustless: colluding operators could in principle reconstruct a key, which is why the nodes are independent. And AI agents buying licenses on their own, with wallets and an autonomous approval loop around keys they use but never see, is what we are building toward, not a finished product.
- Old model: your work is a data source, and enforcement is a lawsuit years later.
- Elacity model: your work is an asset, and enforcement is a key that is never released without the license.
- Old model: whoever holds a copy decides the terms.
- Elacity model: you set the terms, and the rights check runs every time.
As our founder Sasha Mitchell puts it: "The people who create the value should own it. That is the entire reason Elacity exists." The Third Circuit just agreed that value can be harmed. For more on how this market is forming, follow our Market Intelligence coverage.
The court found the market. Taking part in it starts with controlling the key. Get ElastOS.