Assignment and change of control in AI data licenses

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

How deal structure decides whether a data license "moves"

Whether a license transfers depends first on how the deal is structured, then on the exact contract words. In an asset purchase the license must be assigned to the buyer, so any anti-assignment clause bites directly. In a forward merger the target disappears into the buyer, which many courts and drafters treat as a transfer "by operation of law." In a stock purchase or reverse triangular merger the licensee entity survives, so nothing is technically assigned unless the contract says a change of control counts.

The leading Delaware example is Meso Scale Diagnostics v. Roche. In 2013 the Court of Chancery held that a reverse triangular merger was not an assignment by operation of law under the clause at issue, and it declined to follow the 1991 California federal decision in SQL Solutions v. Oracle, which reached the opposite result for a copyright license. An earlier ruling in the same case had refused to dismiss the claim, a reminder that outcomes turn on drafting and governing law; have counsel read the opinions themselves rather than summaries.

For data licenses this matters because the licensed asset is often copyrighted compilations, trade secrets or contract-only rights. US federal courts commonly treat nonexclusive patent and copyright licenses as personal to the licensee, so default rules can lean against free transfer even when the contract is silent; contract-only and trade secret rights turn more on state law and the contract's wording. Counsel should read the clause under the license's own governing-law provision, not the acquirer's home state.

Anti-assignment and change-of-control clauses compared

The clause types below behave very differently at closing, so classify each license before anyone negotiates. An anti-assignment clause restricts transfer of the agreement; a change-of-control clause restricts or reprices the agreement when ownership of the licensee changes, even without any transfer.

Change-of-control definitions usually reference more than 50% of voting securities, the power to elect a board majority, or a sale of all or substantially all assets. Watch for low thresholds that a priced equity round could trip, such as 30% to a single investor, because a financing is not an exit but can still fire the clause.

Clauses a startup should negotiate before the term sheet

The cheapest time to fix transferability is at signing, when the licensor wants the deal. Prioritize these positions in roughly this order:

• Permitted assignment without consent to affiliates and to a successor acquiring all or substantially all of the business or the product line that uses the data. Sample data license clauses already extend use to affiliates and contractors within internal-use limits [1].
• Consent not unreasonably withheld, conditioned or delayed where the licensor refuses a free successor right, with a deemed-consent period after written notice.
• Notice instead of consent on change of control, or a termination right that applies only if the acquirer is a defined competitor of the licensor.
• A narrow competitor definition, listed by name or by revenue segment, not "any entity offering AI products."
• Trained-model survival so that models, weights and checkpoints trained before termination remain usable, even if no new training is allowed. Contracts often address data return but not the model [2].
• Financing carve-out stating that equity issuances to investors, IPOs and internal reorganizations are not a change of control.

These terms sit next to the scope terms covered in per-model vs enterprise-wide training licenses and derivative and successor model rights. A license limited to "Licensee's Model v2" can survive the deal yet still be useless to an acquirer that wants to fold the data into its own foundation model.

Protecting trained models if the license ends at closing

If a change-of-control termination fires, the key question becomes whether the models trained on the data must be retired. A well-drafted training license allocates rights in outputs and trained models expressly rather than relying on a one-sentence grant [3], and termination terms should say what happens to data, derived artifacts and models separately [2].

Draft survival language around concrete artifacts: the training corpus copies, cached embeddings and vector indexes, fine-tuned adapters such as LoRA weights, full model checkpoints, and evaluation sets. A common compromise is that copies of the raw data are deleted with a certificate, embeddings and indexes are purged, and models trained before the termination date may continue in production but not be retrained on the data. Acquirers will price a deal very differently if a core model has to be rebuilt.

Bankruptcy adds a separate layer. Section 365 of the US Bankruptcy Code governs whether a licensee in bankruptcy can assume or assign an IP license without licensor consent, and whether a licensor's rejection affects the licensee's rights. Buyers of distressed AI companies should have bankruptcy counsel analyze both directions before relying on any data license.

What acquirers and lead investors check in data-license diligence

Acquirers check whether each dataset that feeds a production model is licensed for that use, transferable to them, and documented well enough to meet their own regulatory duties. Expect the diligence request list to cover the following.

• A schedule of every data license, click-through, API terms and commissioned-data contract, mapped to the models and products that rely on it.
• Assignment, change-of-control, termination, exclusivity and most-favored-licensee clauses, flagged by consequence at closing.
• Field-of-use and prohibited-use clauses that conflict with the acquirer's business, including competitor bans.
• Usage reporting and audit rights, and whether past reports match actual use.
• Data warranties and indemnities, and whether they run to successors.
• Provenance and consent evidence, de-identification records and deletion certificates, following the AI training data due diligence checklist.

EU exposure raises the bar. As of October 2026, providers of general-purpose AI models must keep a copyright compliance policy and publish a training-content summary under Article 53 of the EU AI Act, obligations that have applied since 2 August 2025 [4]. The Commission's template for that public summary was published on 24 July 2025 [5]. An acquirer placing a GPAI model on the EU market will want license records that let it complete that summary accurately.

A change-of-control clause map for each license

Build one row per agreement and refresh it before each financing round. Data room reviewers read this faster than the contracts themselves.

Illustrative example: invented to show structure; it does not describe an available dataset.

license_id: DL-014
counterparty_type: "US insurance claims processor (licensor)"
data: "Claims adjuster notes, 2019-2024, de-identified"
governing_law: "Delaware"
models_dependent: ["claims-triage-v3", "summarizer-ft-2"]
assignment_clause: "No assignment by operation of law or otherwise without consent"
change_of_control:
  definition: ">50% voting power or sale of substantially all assets"
  effect: "Licensor may terminate on 60 days' notice"
  competitor_carve_out: "Applies only if acquirer is a listed claims processor"
financing_carve_out: false
trained_model_survival: "Models trained before termination may remain in production; no retraining"
data_on_termination: "Delete raw copies and embeddings; deletion certificate in 30 days"
warranties_run_to_successors: true
risk_rating: "Medium: needs financing carve-out before Series B"
action: "Request amendment: notice-only for equity financings"

Pair the map with the cluster guidance in the AI training data licensing hub, and for refreshed feeds check renewal and refresh mechanics in subscription data licenses, since a subscription that ends at closing stops future deliveries even when past data survives.

Sourcing new data with transferability in mind

New licenses are the easiest place to get successor and model-survival terms right. SourceX sources operational datasets from US companies on request and manages the commercial process, including licensing agreements and ongoing purchases; each dataset is rights-reviewed for ownership and consents and delivered under a license that defines records, uses, term and delivery. Nothing is contracted until the supplying company agrees, and every release is approved by that supplier. Buyers describing a need can start at the SourceX buyer page; have counsel raise successor and model-survival terms before any license is signed. More cluster pages live under AI data guides.

Request licensed data with transfer terms on the table

If your team needs support histories, engineering records, documents or finance and legal workflows from US companies, describe the data rather than the businesses that might hold it. SourceX looks for suppliers, assesses data and licensing permissions, and agrees pricing and allowed uses in a license per deal; a request does not guarantee a match. Describe the licensed data your team needs.