Sell side · Licensing mechanics

Data Licensing Guide

How AI training-data licenses actually work — the terms that decide what your data earns and what risk you keep. This guide is general information drawn from public sources; it is not legal advice. For a live deal, the professionals who do this work are data-licensing counsel (law firms such as Frankfurt Kurnit Klein & Selz publish in this area) and de-identification specialists (such as Tonic.ai).

License vs sale

A license grants permission for defined uses. An assignment transfers the asset itself. In AI data deals the money is almost always for a license: Troveo states it plainly to video owners — “you license your videos, you don’t sell them” — and HUD’s DataVendor licenses codebases non-exclusively by default. If a counterparty’s contract says assignment while its pitch says license, that mismatch is the negotiation.

The terms that matter

• Grant & field of use. Training and evaluation of AI models — name it, and exclude everything else (no resale of raw data, no public release, no unrelated product use) unless separately priced.
• Exclusivity. Exclusive rights command a premium because they end your other deals. Sell them narrow (one buyer, one field, a fixed window) or not at all.
• Sublicensing. Vendors and marketplaces resell to labs. Know which labs, whether you are identified to them, and whether the sublicense survives your agreement’s end.
• Term & derived models. Licenses end; trained models do not un-train. Negotiate what the buyer may keep (trained models vs. your raw data), and remedies that reflect that reality.
• Delivery & acceptance. Define the dataset spec, delivery format, acceptance criteria and the payment trigger. “Paid on delivery and acceptance” (Nyne’s published model) is verifiable; open-ended royalty promises are harder to audit — though per-licence royalties (Troveo’s model) can out-earn a flat fee for media with repeat demand.
• Fees & splits. Intermediaries charge visibly or invisibly: DataVendor publishes a 20% withdrawal fee; FileYield’s public fee statements conflict (3% vs 14%) — get the number, and who pays it, in writing.
• Warranties & indemnity. You will be asked to warrant that you own the data and hold the rights to license it. That warranty is why the rights audit comes first: do not warrant what you have not checked.
• Confidentiality & identification. Standard, and worth insisting on: the seller is not identified to end buyers (Nyne’s stated practice), and the data is never published.

Rights and consent hygiene (before you sign anything)

• Inventory every record class and its origin: customer data, employee data, user content, third-party licensed material. Each has a different rights answer.
• Check the instruments: customer contracts and privacy notices, employee agreements, platform terms for anything collected through third parties, and sector rules (HIPAA; IRC §7216 for tax data; state privacy law such as CCPA/CPRA; GDPR where EU data is involved).
• Original consent rarely covers AI-training resale. Where it does not, your options are re-consent, narrowing the dataset to classes you do control, or not transacting that class at all.
• Document the audit. In diligence, a written rights memo is worth money; in a dispute, it is worth much more.

De-identification, properly understood

Removing names is not de-identification. Free-text fields leak identities through rare details, small populations re-identify through attribute combinations, and voice and face are biometric identifiers. A defensible process uses a documented standard, automated PII detection plus human review for sensitive corpora, and a contractual ban on re-identification by the buyer. Marketplaces and vendors that process data themselves (Troveo’s clearance pipeline; Nyne’s de-identification step) are selling partly this: provenance and risk reduction a lab can document.

Pricing factors

• Uniqueness and scarcity — the single biggest driver. Data that exists in exactly one company is priced like it.
• Depth of history — years of records beat months; long-horizon workflow data is what buyers name first.
• Human authorship — for code and documents, human-written material prices above generated or templated content.
• Rights cleanliness — documented consent and ownership convert directly into price; messy rights convert into discounts or dead deals.
• Preparation state — cleaned, formatted, de-identified data transacts faster and higher.
• Exclusivity and field breadth — wider grants cost the buyer more.

Grounded reference points (all cited on our sell-side hub): published buyer tiers for company records start at $100K floors (Micro1, Nyne); reported footage rates run about $0.75–$3 per minute; landmark content deals (Reddit, News Corp) are platform-scale exceptions, not comparables for an operating company’s records.

Negotiation checklist

• Two written offers, minimum — use published price-matches and tiers as leverage.
• Fee, split and who-pays confirmed in writing before exclusivity is discussed.
• Payment trigger you can verify (delivery + acceptance, or auditable royalty reporting).
• Use restriction, no-publication and non-identification clauses present.
• De-identification standard, owner and verifier named.
• Term, post-term data handling, and remedies that survive the un-training problem.
• Your own counsel reviews before signature — the buyer’s paper is the buyer’s paper.