Facts aren't copyrightable, so who owns AI training data a vendor builds? How IP assignment, licensing, and architecture decide ownership, and what to verify.
Who owns AI training data a vendor builds for you is a question most buyers answer wrong, because they assume paying for collection means owning the result. It does not, automatically. In February 2025, a federal court ruled in Thomson Reuters v. Ross Intelligence that training an AI on someone else's data was not fair use, sharpening every ownership clause since. The way enterprise data sourcing is structured decides who actually holds the asset. Below: contract, copyright, and architecture.
You own it only if the contract assigns the intellectual property to you and the vendor cannot retain a copy. Paying for a collection project buys the service, not automatic ownership; ownership is created by an IP assignment clause, an exclusivity clause, and a delivery model that leaves nothing behind with the vendor.
This is the layer AIxBlock is built around. We source, validate, and deliver real-world training data across 100+ languages while the data stays inside the client's environment, so proprietary datasets are never copied, reused, or resold. Ownership here is not a single document. It is the combination of assigned rights over the copyrightable parts, a license scope that grants you exclusivity, and a no-copy architecture that makes the vendor structurally incapable of keeping the set. Where buyers lose ownership is trusting one of those three without the other two.
Raw data is not. The U.S. Supreme Court held in Feist Publications v. Rural Telephone Service Co., 499 U.S. 340 (1991), that facts are not original and therefore not copyrightable; only the original selection and arrangement of a compilation earns protection. Feist copied 4,000 directory listings and won, because names and numbers are facts.
The consequence for training data is direct. A raw corpus of transcribed calls or collected utterances is largely uncopyrightable at the fact level, so copyright alone will not secure your ownership. What is protectable is the original expression layered on top: the annotation schema, the curated selection, the structured arrangement. As the Court put it, "facts are not copyrightable," while original compilations are. This is exactly why ownership of a dataset rests on contract and delivery architecture, not on a copyright notice.
A license grants you permission to use the data; ownership gives you the exclusive right to control, resell, and exclude others. Many labeling contracts deliver a license while the vendor keeps ownership and reuse rights, which is the opposite of what an enterprise buyer usually assumes it is getting.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
The cleaner approach is to require a full assignment with exclusivity, not a perpetual license. Our comparison of licensing terms compared across custom and ready-made datasets shows how the license scope, not the invoice, determines what you walk away owning.
Work made for hire is a US copyright doctrine that assigns authorship of certain commissioned works to the hiring party, but it only reaches copyrightable expression, not raw facts. For datasets, it can cover annotations and structured compilations while leaving the underlying data untouched, because facts were never copyrightable to begin with.
This is where generic templates fail. A work-for-hire clause borrowed from a software agreement assumes the deliverable is fully copyrightable, which a fact-heavy dataset is not. The U.S. Copyright Office is explicit that work-for-hire applies to defined categories of authorship, so the durable move is to pair a work-for-hire provision with an express assignment of all rights plus an exclusivity covenant . This is general information, not legal advice; run the exact clause language past your own counsel.
Whoever the contract says, and by default it is often the vendor. Annotations, labels, transcriptions, and derived features are original expression a vendor can claim unless the agreement assigns them to you. The 2025 Thomson Reuters ruling turned on exactly this: derivative work built from another party's material.
In that case, Ross Intelligence hired a third party, LegalEase, to produce bulk memos derived from Westlaw headnotes, and the court found 2,243 of 2,830 headnotes had been copied and infringed. The lesson for buyers is that derivative datasets carry their own ownership questions. If your vendor subcontracts annotation, the assignment has to flow through every layer. Our guide to the end-to-end data lifecycle covers how sourcing, labeling, and RLHF outputs should all land under a single owner.
It should assign all right, title, and interest in the deliverable to the client, grant exclusivity, and bar the vendor from retaining, reusing, or reselling any copy. Vague phrasing like "customer may use the data" is a license, not an assignment, and it leaves ownership with the vendor.
Three elements make the clause enforceable in practice: an explicit assignment of all rights in both the raw compilation and the derived annotations, an exclusivity covenant naming your organization as sole owner, and a non-retention term backed by the delivery method. A contract that promises exclusivity while the vendor still holds a copy is only as strong as the next audit. The stronger design pairs the clause with ownership by architecture, so the promise and the technical reality match.
Because a clause allocates blame after a breach, while architecture prevents the copy that creates the risk. If the vendor never holds your data, there is no retained set to reuse, resell, or lose. Ownership stops depending on the vendor's honesty and starts depending on system design.
AIxBlock delivers data into client-owned storage from the first record, runs annotation inside the client's environment through scoped accounts, and ships every deliverable with a dataset card and audit log bound to the client's training run. Contributors are verified through KYC and biometric enrollment, and training, GPU compute, and inference stay entirely with the client's MLOps team. Worth noting: an assignment clause and a no-copy architecture are complementary, not redundant. The clause proves intent; the architecture removes the failure mode. Reviewing a vendor's contract red flags alongside its delivery model is how you confirm both hold.
It says the source and licensing of training data now carry real legal weight. Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence, decided February 11, 2025, was the first US ruling that using copyrighted material to train an AI is not fair use, because the use was commercial, non-transformative, and competed with the original.
The ruling does not make raw facts copyrightable; Feist still governs that. It does mean derivative datasets built on protected expression carry infringement exposure that flows to the buyer who trains on them. For an enterprise, the practical takeaway is that provenance and clean assignment are now due-diligence items, not paperwork. A dataset with murky origins is a liability on the balance sheet of any model trained on it, which is one more reason to commission collection you fully own rather than inherit someone else's licensing dispute.
Read for three things: an assignment of all rights, an exclusivity term, and a non-retention guarantee tied to the delivery method. If any of the three is missing, you have a license, not ownership, regardless of what the sales deck says.
Ask where the data lives during labeling, whether the vendor or any subcontractor keeps a copy, and whether the annotation IP is assigned to you. A vendor practicing no-copy delivery answers all three cleanly, because the data never left your environment. In practice, the vendors that route ownership through architecture close these questions in one call, while the ones relying on contract language alone ask to circle back with legal.
Ownership of a vendor-built dataset is never automatic. Facts are not copyrightable under Feist, derivative annotations are, and case law since February 2025 has made provenance a due-diligence line item. Real ownership comes from three things working together: an all-rights assignment, an exclusivity term, and a no-copy architecture that makes the vendor incapable of keeping what it built for you.
Commissioning a dataset you need to own outright? Book a dataset ownership review with the AIxBlock data team to structure collection that lands in your environment with rights fully assigned to you.
Not automatically. Payment buys the collection service, while ownership requires an IP assignment and exclusivity clause plus non-retention. AIxBlock structures delivery so custom-collected data lands in your storage with rights assigned to you, rather than leaving the vendor holding the set under a mere license.
Only if your contract and delivery model allow it, which many do by default. With AIxBlock's self-hosted, no-copy delivery, the operator never retains the data, so reuse or resale is impossible. Verizon's 2025 breach data shows why an unretained copy is the safer design.
No. The U.S. Supreme Court held in Feist Publications v. Rural (1991) that facts are not copyrightable; only original selection and arrangement are. That is why dataset ownership depends on contract assignment and architecture, not a copyright notice, especially for fact-heavy speech and dialogue corpora.
An assignment of all right, title, and interest, paired with an exclusivity covenant and a non-retention term. A perpetual license is not ownership. AIxBlock backs the clause with a no-copy architecture, so the contractual promise and the technical reality match rather than diverge under audit.
