Retraining and Model Retention Rights After a License Ends

When a license of company data ends, most of the published AI data terms compared here require the buyer to delete the raw data but let it keep the models already trained on it. micro1's Data License Master Terms, updated 2026-11-08, state no duty to delete or retrain Licensee Materials (s.4.4), and Labelbox states that models already trained on the data "cannot be unlearned" (Prism FAQ, checked 2026-11-09).

Which rights outlast the license depends on the clause, and the answer belongs in any decision about AI training data licensing. A seller whose retention terms are unacceptable can decline the license, which is a valid outcome.

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

What Retention and Retraining Rights Decide When the License Stops

Retention rights decide what the buyer may keep after the license ends, and retraining rights decide whether it may train again or must rebuild a model without the data.

Published grants are broad. micro1's grant is nonexclusive, irrevocable and perpetual (s.1.2), so the license ends mainly through termination, on 30 days' notice for convenience or after a 10-day cure for breach (s.4.1). The licensee owns models and weights (s.2.2) and has no duty to delete or retrain Licensee Materials (s.4.4), so retention is the default unless a negotiated clause says otherwise.

Who owns those items is covered in ownership of derivatives and trained models; this page covers what the buyer may still do with them.

Four Rights That Can Outlast the License

Four separate rights can survive a license, and a contract can grant or refuse each one on its own.

The table compares the four rights as of 2026-11-10, with a published example and the source for each.

Keeping and Using Models Built Before the End Date

Keeping a trained model and continuing to sell products built on it are two rights, and published terms grant the first more often than they condition the second.

Defined.ai, as licensor of its own datasets, permits commercial exploitation of models trained on its data only if the data is not made available and the outputs contain no recognisable contributor content (1(c), agreement effective 2022-11-30). Under the Kopinor agreement of 2025-12-19, termination leaves models already trained and shared unaffected.

Training New Models After the End Date

Retraining after the end date means training a new or updated model on the licensed data, and only a clause that bars further training stops it.

• Labelbox states that after revocation there is no further training, transfer or sublicensing (checked 2026-11-09).
• micro1's master terms set no limit on new training and leave derivatives, weights and synthetic data with the licensee (s.2.2). Under Kopinor s.16, a rightsholder opt-out takes effect only for future training.

Copies, Backups and Downstream Users That Stay Behind

Copies can outlast the license in backups and in the hands of the buyer's own licensees, and some terms say so in numbers.

• micro1 excludes backups from deletion and confirms deletion in writing on request (s.4.4).
• Protege's licensees keep access for 12 months after termination and academic licensees for 3 years (s.7.2, checked 2026-11-09).
• DataVendor, a marketplace, states that a buyer's entitlement survives the vendor unpublishing; Pangea.ai, a code licensing platform, states that completed licenses survive (both checked 2026-11-09).

Retention Terms Buyers and Intermediaries Publish

Five published contract texts state what survives termination, and each says it differently.

The table compares them as of 2026-11-10 in alphabetical order, with roles labeled and a source for each.

The other clauses of these contracts are compared in the license clause matrix of buyer terms.

Where Published Terms Say Nothing About Retention

Most buyer program pages do not publish a retention or retraining term at all, so the signed agreement is the only place to find it.

Scale AI, Handshake AI, Mercor, Mode, Nyne and Claru publish no retention or retraining term on their program pages (checked 2026-11-09). Endless, an aggregator that licenses data on to labs, states deletion at the end of the agreed term. The Project Lazarus page of Turing, a selling agent, does not state what happens to licenses already sold once a seller opts out.

Why Trained Models Usually Survive the Deletion of Raw Data

A trained model no longer holds the data as files, so deleting the dataset leaves the model in place, and buyers do not promise to remove a dataset's influence afterwards.

micro1 deletes raw data at the earlier of 120 days after delivery completes or termination (s.4.4), and Labelbox states that trained models cannot be unlearned. None of the published terms cited here offers an unlearning commitment.

Those clauses have their own page on deletion, return and machine unlearning clauses.

When Regulators Have Ordered a Model Deleted

A regulator, not a contract, has ordered a model deleted: the FTC's final Everalbum settlement of May 7, 2021 required deletion of models built on users' photos and videos. The FTC release states that the order requires deletion of "the models and algorithms it developed by using the photos and videos uploaded by its users" (File No. 192-3172; the order text itself was not checked). Orders of this kind are the subject of FTC rules on AI training data.

For a seller, a model built on unlawfully used personal data can carry the problem past the end of the license.

In the EU, EDPB Opinion 28/2024, adopted 2024-12-17 under Art. 64(2) GDPR, states that models trained on personal data cannot in all cases be considered anonymous. Anonymity requires that extraction and query-based retrieval of personal data are insignificant; the opinion is soft law. The rules on GDPR and AI training data decide whether a retained model still processes personal data.

Clause Options That Limit What Survives the License

A seller who does not accept open-ended retention can ask for one of six clause types, and each has a published or filed example.

1. No training after the end date, mirroring Labelbox's wording on revocation.
2. Conditions on commercial use of retained models, mirroring Defined.ai 1(c).
3. Deletion with written certification, mirroring Defined.ai 8(c) and micro1's written confirmation on request (s.4.4).
4. A defined tail period for downstream users, mirroring Protege's 12 months and 3 years (s.7.2).
5. An opt-out with effect for future training only, mirroring Kopinor s.16.
6. A license back to the model: under the Pathos AI license from Tempus AI, the data serves only to train a foundation model, and Tempus receives a license to that model (Tempus 10-K for fiscal 2025).

No public figure exists for the price of any of these concessions. Labelbox states that buyers pay for stricter exclusivity, and no comparable statement exists for retention limits.

Accepting Model Retention: Benefits, Risks, Downsides and Alternatives

Accepting that a buyer keeps its trained models is common in the published terms compared here, and whether it is acceptable depends on what the data reveals and on what the seller can still license elsewhere. Retention is one input to the question of should you sell or license your data at all.

Benefits

• Raw data still deleted on a deadline (micro1: 120 days, s.4.4).
• Non-exclusive grants leave the data licensable to others (micro1 s.1.2).

Risks

• Personal or confidential content may be inferable from a retained model.
• Retention passes with the contract when the buyer is merged or acquired (micro1 s.9.1).
• Labelbox's FAQ does not state a refund term for revocation.

Downsides

• The license cannot be undone by deleting the data.
• Downstream users keep access for a tail period, as in Protege's terms.

Alternatives

• Negotiate one of the clause options above.
• License only de-identified or less sensitive subsets.
• Choose a short term.
• Do not license at all.

Other routes are set out in alternatives to selling data to AI labs.

How Retention Fits With the Other Clauses in the Contract

Retention interacts with the grant, the term, deletion and derivative ownership, so it is read together with those clauses rather than alone. The parent page explains every clause of an AI data licensing agreement.

• Exclusive vs non-exclusive data licenses
• License term and perpetual licenses
• Field of use and scope limits
• Audit rights in AI data licenses
• Sublicensing and onward transfer
• Ownership of derivatives and trained models
• Negotiating an AI data licensing deal
• Evaluating an offer for your data
• Rights audit

Retention When Data Is Sold Rather Than Licensed

A sale has no end date, so the retention question does not arise: the Spirit Airlines estate offered de-identified business data under a bill of sale, not a license. Its successful bid from Google LLC, USD 10,000,000, is not approved as of 2026-11-09, with a hearing set for 2026-11-17 (Dkt. 1463 and Dkt. 1779), and the form of bill of sale says the seller's representations do not survive consummation. See US bankruptcy data sales and Spirit Airlines' data archive sale.

Trade Secret Protection for Limits on Further Training

In the EU, keeping a dataset secret and limiting its use by contract matters, because the Trade Secrets Directive (Directive (EU) 2016/943, in force since 2016-07-05) treats use in breach of a contractual limitation as unlawful use. See EU Trade Secrets Directive and licensed datasets.

Information, Not Legal Advice

This page explains published terms and public rules and is not legal advice for a specific agreement.

Terms, Programs and Deals Cited on This Page

The terms above come from these programs' and parties' own documents, checked on the dates shown.

• Claru
• DataVendor (marketplace)
• Defined.ai (licensor of its own datasets)
• Endless (aggregator that licenses on to labs)
• Handshake AI
• Kopinor and the National Library of Norway (agreement of 2025-12-19)
• Labelbox Prism
• Mercor
• micro1
• Mode
• Nyne
• Pangea.ai (code licensing platform)
• Protege
• Scale AI
• Tempus AI clinical data licensing deals
• Turing Project Lazarus (selling agent)

The wider field is listed under who buys AI training data.

Frequently Asked Questions

Does the buyer keep a model trained on my data after the license ends?

Usually yes: the published terms compared here delete the raw data but leave trained models with the buyer. micro1 states no duty to delete or retrain, and Labelbox states that trained models cannot be unlearned.

Can I require a buyer to retrain its model without my data?

Not under the published terms cited here; micro1 states no duty to retrain, and a retraining duty would have to be negotiated. The model deletion example above came from a regulator, not a contract.

Can a buyer train new models on my data after termination?

Only if the contract allows it or says nothing; Labelbox states no further training after revocation. The Kopinor agreement limits the opt-out to future training.

Does accepting model retention change what my data is worth?

No public figure prices retention limits, so their effect shows only in written offers compared on the same terms.

The valuation page explains how much company data is worth to AI companies.