DataPacts

Ownership of Derivatives, Annotations and Models Trained on Licensed Data

Checked

At a glance

In published AI data licenses the seller usually keeps ownership of the data itself, while the buyer owns what it builds from it: annotations, models and their weights, outputs and synthetic data. micro1 states the split in two sections of its Data License Master Terms, last updated November 12, 2026 (s.2.1 and s.2.2).

Independent reference. Every buyer is covered on the same terms, every figure carries its source and check date, and no buyer pays to be listed or placed higher.

Request an independent introduction

Tell us what data your company holds. We reply with the buyers that fit and what their terms leave out. Your details reach a buyer only after you agree to a named introduction.

IndependentYour details stay with us until you agreeNot selling is a valid outcome

A license-back or a narrow grant can shift the split. Conceding ownership, limiting it or not licensing are all valid outcomes in AI training data licensing.

Every term below carries its source and check date; how DataPacts sources, dates and checks every figure is set out in the methodology.

What Counts as a Derivative of Licensed Data

A derivative is anything the buyer creates from the licensed data, from labels added to records to a model whose weights were trained on them.

The table lists seven kinds of derivative as of 2026-11-17 and shows whether micro1's Licensee Materials definition (s.2.2) names each one.

Derivative What it is Named in micro1 s.2.2? Source
Annotations and labels tags, ratings or corrections added to records no; covered by "derivatives, modifications, and improvements" or "datasets" micro1.ai terms, read 2026-11-17
Cleaned or restructured datasets the data reorganized or combined "datasets" same
Embeddings numeric representations computed from records no; covered by "other materials ... using or based on the Data" same
Models, weights and parameters trained or fine-tuned models "models (including weights and parameters)" same
Outputs what the model generates "outputs" same
Synthetic data generated records imitating the data "synthetic data" same
Analyses and insights reports and findings drawn from the data "analyses, insights" same

The contract definition, not the technical category, decides what is owned, and micro1's list ends with "other materials created by or for Licensee ... using or based on the Data".

Ownership Terms in Published AI Data Contracts

Thirteen published terms say who owns the data or what is built from it, and only micro1's master terms list the derivatives one by one.

The table compares the thirteen terms, in alphabetical order, as of 2026-11-17, with each stating party's role and the date its source was checked.

Party (role) Document and date Data Derivatives and models Source and check date
AIChamp (factory study service) service terms, effective 2026-10-22 customer owns its data; limited license to deliver the project not stated aichamp.com/terms.html, 2026-11-09
Claru (buyer) partnerships page partner keeps ownership; non-exclusive training rights not stated claru.ai/partnerships, 2026-11-09
Collinear (buyer) data acquisition page seller keeps ownership; Collinear gets usage rights not stated collinear.ai/data-acquisition, 2026-11-09
Defined.ai (licensor of its own datasets) Data License Agreement, effective 2022-11-30 licensed, non-exclusive (1(a)) models may be exploited commercially if the data is not made available and outputs hold no recognisable contributor content (1(c)) defined.ai/data-license-agreement, 2026-11-09
Endless (aggregator that licenses on to labs) home page seller keeps ownership not stated endless-data.com, 2026-11-09
Kopinor and National Library of Norway (collective licensor, public licensee) agreement of 2025-12-19 non-exclusive license models may be shared under royalty-free, non-exclusive terms (s.7); unaffected by termination kopinor.no agreement PDF, 2026-11-16
Labelbox Prism (buyer) Prism FAQ seller retains full ownership and continued use trained models "cannot be unlearned" prism.labelbox.com/faq, 2026-11-09
micro1 (buyer) Data License Master Terms, last updated 2026-11-12 licensor retains title; "licensed and not sold" (s.2.1, s.2.2) licensee owns Licensee Materials incl. models, weights, outputs, synthetic data (s.2.2) micro1.ai/legal/data-license-master-terms, read 2026-11-17
Outlier (Scale AI; contributor platform) Terms of Service, effective 2026-05-11 contributor assigns Work Product (s.4(b)) platform holds all rights outlier.ai, 2026-11-09
Pangea.ai (code and company data platform) company data page company keeps ownership not stated pangea.ai/data-licensing/company-data, 2026-11-09
Proximal (builder of custom data and environments) MSA, last updated 2026-07-06 customer materials returned (s.6.4) Proximal owns Deliverables, may reuse them, keeps learnings and model training (s.8.1 to 8.3) proximal.ai/msa, 2026-11-09
Tempus AI and Pathos AI (data holder, model developer) order form 2025-04-17, Tempus filings licensed for the sole purpose of training the Foundation Model Tempus receives a license to the model, sublicensable to AstraZeneca Tempus 10-K FY2025; 10-Q to 2026-09-30, 2026-11-16
Turing Project Lazarus (selling agent) program page ownership does not change hands not stated lazarus.turing.com, 2026-11-09

Each term is the stating party's own wording; the signed order form can change it.

Abaka AI's Project Reborn page states no ownership term, and Scale AI, Handshake AI and Mercor publish no derivative term (checked 2026-11-09), so the signed order form decides. The clauses of these buyers can be set side by side in the license clause matrix.

The Usual Split: The Seller Keeps the Data, the Buyer Owns What It Builds

The clearest published term, micro1's, combines two promises: the data stays the seller's property, and everything derived from it belongs to the buyer.

micro1 s.2.1 states that the licensor "retains all right, title and interest (including all intellectual property rights) in and to the Data", and s.2.2 adds that the licensee "acquires no ownership of the Data itself, which is licensed and not sold".

Labelbox, Turing, Claru, Collinear, Endless and Pangea.ai state that the seller keeps ownership (checked November 9, 2026). Those statements concern the data, and none of these pages says who owns models built from it.

Each Derivative and Who Owns It

Published terms name some derivatives explicitly and catch the rest with general wording.

Annotations and Labels

No published AI data contract recorded here names annotations or labels, so they fall under general words such as micro1's "derivatives, modifications, and improvements of the Data" and "datasets"; the words "annotation" and "label" do not appear in its terms (read 2026-11-17).

When people paid by a platform write the labels, the platform takes them by assignment: Outlier's terms of service, effective 2026-05-11, have the contributor assign "all right title and interest throughout the world" in Work Product (s.4(b)).

Under 17 U.S.C. 101 and 201(b), an employer owns a work made for hire, while a commissioned work needs a signed agreement and one of nine categories, which is why platforms use assignments.

Embeddings and Other Representations

Embeddings, the numeric representations a model computes from records, are not named in any published term recorded here. micro1's catch-all "analyses, insights and other materials ... using or based on the Data" would cover them.

Fine-Tuned Models and Their Weights

Models and their weights are the derivative published terms address most directly, and they assign them to the buyer. micro1's grant covers use to "develop, train, fine-tune, evaluate and test" AI (s.1.2), and s.2.2 gives the licensee "models (including weights and parameters)".

Labelbox states that "Models already trained on the data cannot be unlearned" (checked November 9, 2026).

Synthetic Data Generated From Licensed Data

Synthetic data generated from licensed records is listed as the buyer's property in micro1's terms, and that matters because a synthetic set can travel further than the data it imitates.

Section 2.2 names "synthetic data", while the resale bar in s.1.4(e) covers the Data as received. No published term recorded here limits resale of a synthetic set; see synthetic data for AI training.

Restrictions That Protect the Seller Inside a Buyer-Owns Grant

A buyer can own its derivatives and still be barred from using the data, or what it learns from it, against the seller.

micro1's s.1.4 (read 2026-11-17) bars these uses:

  • building or improving a product that directly competes with the licensor's business (b);
  • knowingly identifying or re-identifying any natural person or customer in de-identified Data (c);
  • reverse-engineering the licensor's business processes, pricing methodologies, customer relationships or workflows (d);
  • reselling the same Data, as received, as a standalone data product (e).

Defined.ai, licensing its own datasets, permits commercial use of models trained on them only if the data is not made available and outputs contain no recognisable contributor content (1(c), agreement effective 2022-11-30). The Kopinor agreement bars training specialised models for media and news services (s.2).

Trade secret law excludes "reverse engineering, independent derivation, or any other lawful means" from improper means (18 U.S.C. 1839(6)(B), read 2026-11-17), so only a contract bar like s.1.4(d) excludes it; see the Defend Trade Secrets Act page.

License-Backs: When the Data Holder Gets Rights in the Model

A license-back gives the data holder rights in the model trained on its data, and two filed or published agreements contain one.

Tempus licensed a de-identified multimodal dataset to Pathos AI "for the sole purpose of assisting in the development and training of the Foundation Model" and receives a license to that model, sublicensable to AstraZeneca (Tempus 10-K for fiscal 2025 and 10-Q for the quarter ended 2026-09-30, checked November 16, 2026).

The Kopinor agreement of 2025-12-19 lets the National Library of Norway share trained models "under royalty-free, non-exclusive terms of use (license)" (s.7), and termination does not affect models already trained and shared. Both differ from the company-data programs above, as the filings on Tempus AI clinical data licensing deals show.

Commissioned Data: When the Vendor Owns What It Makes for You

When a company pays a vendor to build data or environments from its own materials, the vendor's terms may keep the result.

Proximal's MSA, last updated 2026-07-06, says Proximal owns the Deliverables (s.8.1), the customer gets a non-exclusive license for internal business use during the Term (s.8.2), and Proximal may reuse Deliverables for other customers absent customer confidential information (s.8.1) while keeping "learnings, improvements, optimizations, or model training" (s.8.3).

AIChamp's terms run the other way: the customer owns its data and AIChamp uses it only to deliver the project (effective 2026-10-22).

Labeling vendors' own terms are compared on the data labeling companies page.

The US Copyright Office's view is that a model's weights can contain copies of training works, but only where the model has memorized substantial protectable expression.

The Office states that training "may result in model weights that contain copies of works in the training data", and that copying those weights, "even by parties not involved in the training process", could be prima facie infringement (Part 3, p. 28).

Weights that "have memorized protectable expression from training data may also infringe the derivative work right" (p. 29).

Whether weights implicate those rights "turns on whether the model has retained or memorized substantial protectable expression", and copying them infringes "only where there is substantial similarity" (p. 30). The US Copyright Office report on generative AI training page holds the full analysis.

Part 3 exists only as the pre-publication version released 2025-05-09, and copyright.gov states on 2026-11-17 that "A final version of Part 3 will be published in the future" without substantive changes expected. The report is not law, and the Office notes that courts have reached varying conclusions (p. 29).

Owning the weights does not settle whether they carry the seller's works.

Owning a Derivative Versus Using It After the License Ends

Ownership answers who holds a derivative; a separate set of clauses decides whether the buyer may keep and keep using it once the license stops.

micro1 s.4.4 sets no duty to delete or retrain Licensee Materials built on Data the licensee accepted and paid for, and retraining and model retention rights cover the rest.

Clause Options for Derivative Ownership

A seller who does not accept a blanket buyer-owns clause can narrow it in six ways, each with a published or filed example except the last.

  1. Retained title to the data and its improvements, as micro1 s.2.1 does for the Data.
  2. A sole-purpose grant tying derivatives to one model, as in Tempus and Pathos.
  3. A license-back to the model, as in Tempus and Kopinor s.7.
  4. Bars on competing products, reverse engineering and re-identification, as in micro1 s.1.4.
  5. Conditions on commercial use of trained models, as in Defined.ai 1(c).
  6. A carve-out keeping annotations or synthetic sets out of the derivatives definition (no published example).

No public figure prices any of these concessions.

Conceding Derivative Ownership or Not: Benefits, Risks, Downsides and Alternatives

Conceding derivative ownership is what micro1's published terms provide, while the other buyer pages compared here are silent, and whether it is acceptable depends on what the derivatives could reveal or replace.

Benefits:

  • It matches published buyer terms (micro1 s.2.2), so negotiation can focus elsewhere.
  • The seller keeps title to the data (s.2.1).
  • Non-exclusive grants leave the data licensable again (s.1.2).

Risks:

  • Synthetic sets and models can substitute for the data in later deals.
  • Weights may carry memorized expression, in the Copyright Office's view (pp. 28 to 30).
  • Vendor terms may keep commissioned deliverables (Proximal s.8.1).

Downsides:

  • Derivatives survive deletion of the raw data (micro1 s.4.4).
  • No public price exists for a license-back or a carve-out.

Alternatives:

  • Narrow the definition or add a license-back.
  • License only less sensitive subsets.
  • Choose a buyer whose terms condition model use (Defined.ai 1(c)).
  • Do not license at all.

The wider question is covered under alternatives to selling data to AI labs.

How Derivative Ownership Works With the Other Clauses of the License

Derivative ownership is read together with the grant, the permitted purpose, onward transfer and deletion, because each changes what the buyer can do with what it owns. The parent page explains every clause of an AI data licensing agreement.

Related clause pages:

Process steps:

This page explains published ownership clauses and a government report for information and is not legal advice; a lawyer in the seller's jurisdiction should review any derivatives clause before signing. The Copyright Office report is a pre-publication agency report, not a statute or ruling.

Terms, Filings and Reports Cited on This Page

The examples above come from these documents, each checked on the date shown in the tables.

Frequently Asked Questions

Who owns an AI model trained on licensed data?

Under micro1's published master terms, the buyer, which owns models including weights and parameters; the other buyer pages compared here are silent, while a license-back such as Tempus's can give the data holder rights.

Does the seller keep ownership of data licensed for AI training?

Usually yes: micro1 states the data is "licensed and not sold", and Labelbox, Turing, Claru and Pangea.ai state the seller keeps ownership (checked November 9, 2026).

Who owns synthetic data generated from licensed data?

Under micro1's terms the licensee owns it as Licensee Materials (s.2.2); no published term recorded here limits its resale, so a seller has to negotiate any limit.

Is a model trained on copyrighted data a derivative work?

The US Copyright Office says weights implicate the derivative work right only where the model memorized substantial protectable expression (Part 3, pre-publication version, p. 30).