License Clause Matrix: How Buyer Contracts Differ on Exclusivity, Deletion and Indemnity

This matrix sets the contract clauses that AI data programs publish side by side, so a company can see how one buyer's terms on exclusivity, deletion and indemnity differ from another's before it signs. Of 45 programs, 24 publish enough to compare and 21 publish nothing or leave every clause to the signed agreement, as of 2026-11-09. Rows are alphabetical, never ranked, and no program pays DataPacts.

Contract terms are one part of AI training data licensing, and the matrix reports only what a program's own documents say and the date they were read. It does not show what any company would be offered.

DataPacts Editorial Team read each document in the matrix as of November 9, 2026; the page on how we source, date and check every figure sets out the rules.

Every Published AI Data Contract, Clause by Clause

The matrix holds 24 programs whose own contract, terms page or FAQ states at least one of the nine clauses, each quoted with the date it was checked. Filters narrow it by kind, show one clause column at a time, or hide every cell that reads "not published".

Kinds: CD company data program; IN seller-side intermediary (broker or adviser that handles data and does not buy it); MK marketplace (seller lists, platform takes a fee); CL content licensing platform; CC custom capture; RL RL environments and data; SV service terms. Section numbers are the program's own. As of 2026-11-09.

Programs That Publish No Contract Terms

Twenty-one programs, among them Handshake AI, Mercor and Scale AI, publish no contract clauses, or say only that the signed agreement will set them. The table gives what each program's pages say about terms, as of 2026-11-09.

Kinds as above, plus DV dataset vendor.

An empty row is a question to ask before signing, not a sign that the terms are worse. How each of these works as a program is described in company data partnership programs.

How to Read the Clause Matrix

Read each cell as a dated quotation of one program document, not as DataPacts' reading of what the buyer would sign with a given company. The order form can override master terms, as at micro1 (introduction to the Master Terms and s.4.2), so a master-terms cell is a starting point.

Contract Text, FAQ or Marketing Page: Why the Source Column Matters

A clause in a published master agreement binds more than the same words on a FAQ or landing page, so every row names the document it comes from. The matrix has three source classes:

• Contract text: micro1's Data License Master Terms, the Defined.ai Supplier Program terms, Protege's Provider Terms and Grepped's Terms (2026-11-08).
• FAQ only: Labelbox Prism, whose values are taken from its FAQ.
• Marketing page: Turing, Endless, SuperAnnotate and most other rows.

What "Not Published", "Order Form Only" and "Agreement Defines It" Mean

Four labels fill the empty cells, and each one points to a different question for the buyer:

• Not published: the program's pages are silent on the clause.
• Order form only: the master terms defer the clause to the order form, as micro1 does for audit and exclusivity (s.4.2).
• Agreement defines it: the clause is named, but the value is left to the signed agreement.
• Not recorded: DataPacts holds no value for the cell.

Platform Kinds in the Matrix

Not every row is a buyer: marketplaces, seller-side intermediaries and content licensing platforms appear beside companies that license data for their own use, and the kind column says which is which. Aba Compute, Corpus Bridge and Endless handle data and do not buy it; Grepped, HUMXN, Kled AI, Pangea.ai and Protege are marketplaces; Newsflare and TollBit are content licensing platforms. Fees and eligibility for the marketplaces are compared on AI training data marketplaces.

When the Contract and the Program Page Disagree

Three programs give a different answer on the same clause in two of their own documents, four contradictions in all, and the matrix shows both answers without choosing. The table is as of 2026-11-09, with Defined.ai's two documents re-read on 2026-11-10.

Contradictions about payouts and payment timing are collected in payouts and program terms compared, dated.

How the Published Contracts Differ, Column by Column

The widest gaps between buyers sit in five places: exclusivity, how long the license runs, what is deleted, who carries third-party claims, and the rules for audit, exit and disputes.

Exclusivity Column

Published grants run from plainly non-exclusive to exclusive by default, and the label alone does not show whether the data can be licensed again.

• Non-exclusive: micro1 (s.1.2), the Defined.ai Supplier Program (s.3.5), Protege (s.1.1), Claru, Endless and Pangea.ai, among others.
• Time-limited sole seller: Turing is the only seller for the first 90 days, then the owner can renew or opt out.
• Exclusive by default or required: Labelbox Prism, Kled AI and Newsflare, and Polyshares on its track-record page.
• Seller chooses: Aba Compute, Grepped, HUMXN and SuperAnnotate.

Exclusivity also works as a screen. Mercor's listings name data "already licensed exclusively through a broker" and studies "under an exclusive contract with another buyer" as not a fit, and micro1 asks the licensor to warrant that it has not granted a third party rights to the same data on the same or substantially similar terms (s.5.1). Troveo states that 95 percent of its licensors signed exclusively (page dated 2026-07-25). What exclusivity costs a seller is covered under exclusive vs non-exclusive data licenses.

Duration Column

Four published duration patterns appear: perpetual and irrevocable, a fixed or agreed period, until one side terminates, and pages that say two things.

• Perpetual and irrevocable: micro1 (s.1.2), Kled AI, and Grepped for licenses to buyers.
• Fixed or agreed period: Corpus Bridge's illustrative draft states 3 years, SuperAnnotate "a set period", Sureform whatever the signed agreement defines.
• Until terminated: the Defined.ai Data License Agreement (s.8(a)).
• Two answers: Defined.ai (the FAQ says perpetual) and Polyshares.

Deletion Column: Raw Data Versus Trained Models

Every published deletion clause covers the raw data, and none of the 24 programs promises to delete or retrain a model already trained on it. Four deletion patterns appear:

• Fixed deadline: micro1 deletes at the earlier of 120 days after delivery completes or termination (s.4.4), Ooak Data within 30 days of delivery, AIChamp within 30 days after project end.
• On demand or kill clause: Labelbox Prism within 30 days, SuperAnnotate and Newsflare on request.
• End of term or no deal: Endless, Aba Compute and Sureform at the agreed date; Origin Lab and Patronus AI if no deal is made.
• No duty: micro1 has no duty to delete or retrain Licensee Materials (s.4.4), and Kled AI may but need not delete.

Models are the gap. The licensee owns models and weights under micro1 (s.2.2), Labelbox states that models already trained "cannot be unlearned", and Aba Compute states that deletion does not reverse completed training. An FTC staff post of 2024-01-09 describes model deletion as a remedy in enforcement actions, not a contract term. What happens to a trained model after a license ends is covered under retraining and model retention rights.

Indemnity and Liability Cap Columns

Where an indemnity is published, the seller or provider defends third-party claims arising from the data, and the published caps are multiples of fees or business volume.

• Who indemnifies: the micro1 licensor for IP, privacy, publicity and data protection claims (s.8.1), the Protege provider (s.4.1), the Grepped seller, and the TollBit publisher.
• Caps: micro1 3x fees under the Order Form (s.7.2), Protege 2x fees in the prior 12 months (s.6.2), and the Defined.ai Supplier Program 2x total business conducted.

No published survey of typical caps exists.

Audit, Termination and Governing Law Columns

Audit rights are rare in the published terms, termination notice runs from 30 to 60 days, and four US state laws govern the contracts that name one.

• Audit: the Defined.ai Supplier Program allows one audit a year on 15 business days' notice; Endless states audit rights.
• Termination: 30 days' notice for convenience at micro1 and 60 days at Defined.ai; Pangea.ai asks 30 days after a withdrawal notice.
• Governing law: Delaware (AIChamp, Grepped, Protege), California (micro1, with JAMS arbitration), Washington (Defined.ai) and New York (TollBit, with JAMS arbitration).

What Platforms Grant Their Own Customers After You License

When a platform resells or sublicenses a seller's data, its own contract with the AI developer decides how long that developer keeps it, and several platforms publish that side too. The table is as of 2026-11-09, Defined.ai FAQ 2026-11-10.

Ending a contract with the platform does not by itself end its customers' licenses: Protege's licensees keep access for 12 months after termination and Pangea.ai states that completed licenses survive.

Using the Clause Matrix Before You Sign: Benefits, Risks, Downsides and Alternatives

Published clauses show where a negotiation starts, but the order form a company is sent can change every one of them, and none of these documents says whether licensing suits a given company.

Benefits

• The matrix shows which programs publish terms at all: 24 of 45.
• Section numbers can be compared before a first call.
• Clauses that outlast the deal are visible in advance.

Risks

• A perpetual, irrevocable grant cannot be taken back, as in micro1's master terms.
• Trained models stay with the licensee under micro1 and Labelbox Prism.
• Some contracts make the seller indemnify the platform, and published caps differ in what they cover.

Downsides

• An order form can override the master terms.
• FAQ wording can differ from the contract, as the micro1 and Defined.ai rows show.
• 21 programs give nothing to compare.

Alternatives

• Asking for the order form before sharing any data.
• A time- or purpose-limited grant, as SuperAnnotate and Corpus Bridge describe.
• A platform that publishes deletion and confirmation terms, such as Ooak Data.
• Not selling or licensing at all.

Whether to license at all is weighed in should you sell or license your data to AI companies.

Where the Clause Matrix Sits Among the Seller Tools and Contract Guides

The clause matrix is one of the seller tools: it shows what programs publish, while what each clause does and when to push back is explained in the contract guides. The tools for AI data sellers page lists the others, and AI data licensing agreements, every clause explained covers each clause.

• AI data buyer match.
• Who buys AI training data.
• Clause guides: license term and perpetual licenses, indemnities and warranties, liability caps, audit rights, deletion, return and machine unlearning clauses, ownership of derivatives, acceptance and deemed acceptance, sublicensing and onward transfer and governing law and arbitration.
• Buyer estimators side by side.

How the Matrix Is Checked and Updated

Each program document is re-read before a change is published and at least monthly, and a row's check date changes only when the document was read again. Dated term changes are logged in the buyer program changelog. No program in the matrix pays DataPacts, the page carries no referral links, and DataPacts takes no fee from any program.

Information, Not Legal Advice

This matrix reports what programs publish and is not legal advice; a lawyer in the relevant jurisdiction should review any agreement before it is signed. Regulation (EU) 2023/2854 (Data Act), Art. 13, makes unfair data-use, liability or remedy terms that one enterprise unilaterally imposes on another non-binding, for contracts concluded after 2025-09-12 (Art. 50; EUR-Lex).

Frequently Asked Questions

Which AI data programs publish their full contract terms?

Few do: micro1's Data License Master Terms, Defined.ai's Supplier Program terms, Protege's Provider Terms and Grepped's Terms are the published contract texts in the matrix. The other 20 rows come from FAQs and landing pages, as of 2026-11-09.

Which published contracts make the seller indemnify the platform?

micro1, Protege, Grepped and TollBit publish terms under which the seller or content provider covers claims arising from the data. micro1 also makes the licensee indemnify use outside the license (s.8.2), as of 2026-11-09.

Do any published contracts promise to delete trained models?

No program in the matrix promises to delete or retrain a model already trained on licensed data; deletion clauses cover the raw data. Labelbox Prism states that trained models cannot be unlearned, as of 2026-11-09.