Trade Secrets in the Age of Vector Databases
Embeddings are derived data, but they are still derived from something confidential. What reasonable steps look like now.
Authorship, originality and the contractual fixes available while the case law settles in the UK, EU and US.
The question arrives from clients weekly, and the honest answer is that ownership of purely machine-generated output remains unsettled in most jurisdictions. What is settled is how to contract around the uncertainty.
Copyright protection generally requires human authorship or a human intellectual creation. The more autonomous the generation, the weaker the claim to protection — not because the output is unprotectable in principle, but because there may be no author to whom rights attach. Output with no protection is not owned by the user; it is owned by nobody, and may be freely copied.
United States. The Copyright Office has consistently required human authorship, registering works containing AI-generated material only in respect of the human-authored contributions, with disclosure of the AI-generated portions.
United Kingdom. Section 9(3) of the CDPA provides for computer-generated works with no human author, deeming the author to be the person who made the arrangements necessary. Its application to modern generative systems is contested and its future has been the subject of consultation.
European Union. The originality standard requires the author’s own intellectual creation reflecting free and creative choices, which points towards protection only where human creative input is meaningful.
Assume the output may be unprotected and address the consequences: an assignment of whatever rights do subsist, together with an obligation to assist in perfecting them; a licence back sufficient to operate if the assignment fails; representations about training data provenance and third-party rights; an indemnity for third-party IP claims, appropriately capped; and an express statement on whether the supplier may use your inputs and outputs for further training.
Document the human contribution. Selection, arrangement, editing, iterative direction and creative judgement are what supports a protection claim in most jurisdictions. A record of that process, made contemporaneously rather than reconstructed, is worth more than any clause — and it costs nothing to keep.
This article is general information about legal technology and practice, not legal advice, and it does not create a lawyer–client relationship. JuriPro is a technology company, not a law firm. Take advice from a qualified lawyer admitted in the relevant jurisdiction before acting on anything here.
Legal Research Lead, JuriPro
Former law librarian and knowledge manager; curates the primary-source corpora behind the Legal Research Assistant.
Embeddings are derived data, but they are still derived from something confidential. What reasonable steps look like now.
A plain-English account of how a language model reads an agreement, where its judgement is genuinely useful, and the four failure modes every reviewing lawyer should know about.
Uncapped indemnities, silent auto-renewals, unilateral change rights: the provisions that rarely make the negotiation summary but decide who pays when something goes wrong.
Start a 14-day trial, or book a 30-minute walkthrough with someone who has practised.