Derivative and successor model rights in AI data licenses

Why copyright's "derivative work" test cannot set model scope

Copyright asks whether protected expression was copied or adapted; model scope asks which models were built from the licensed data, and the copyright answer for trained models is unsettled. As of October 2026, the U.S. Copyright Office's Part 3 report is still a May 2025 pre-publication version. It finds that compiling training datasets implicates the reproduction right and considers whether model weights can themselves be infringing copies where outputs closely resemble training inputs [1]. The builders of the GPT-NL corpus excluded NonCommercial and ShareAlike data to keep it usable for commercial training, and describe the law on whether a trained LLM is a derivative work as unclear [2].

US courts decide training disputes on their own facts. On 29 September 2026 the Third Circuit, in a precedential opinion, affirmed that ROSS's use of Westlaw headnotes to train a non-generative legal-research tool was not fair use [3]. In Kadrey v. Meta, a June 2025 district court ruling found fair use on its record for training on the plaintiffs' books; as of October 2026, other claims continue [4]. Neither outcome settles whether a distilled student derives from your corpus.

A license that uses "derivative works of the Data" as its boundary imports that uncertainty in both directions:

• Licensor reading. Every descendant model is a derivative work, so all fall under restrictions and deletion duties.
• Licensee reading. A student trained only on a teacher's generated outputs holds little or none of the licensed expression, so it escapes every restriction, though its capability came from the data.
• Output status. Lemley and Henderson argue model outputs lack the human authorship copyright requires, so limits on their use rest on contract [5].

Instead, define Model, Successor Model and Derived Outputs by lineage (weights inherited, data used, outputs used) and add "whether or not a derivative work under applicable law."

Three ways to define the licensed Model

Most model-scope clauses follow one of three patterns, and the right one depends on how long the licensed data must keep contributing to your roadmap.

Illustrative example: invented to show structure; it does not describe an available dataset. Not legal advice; adapt with counsel.

PATTERN A
"Licensed Model" means the model identified in Schedule 2 by name, base
checkpoint (SHA-256 of the weight files) and training run ID, and any
Permitted Derivative. "Permitted Derivative" means a copy of the Licensed
Model that is quantized, pruned, converted to another file format or adapted
with low-rank adapters, where no Licensed Data is used after the Schedule 2 run.

PATTERN B
"Successor Model" means a model that (i) is initialized in whole or in part
from parameters of a Covered Model, (ii) is trained on a data mixture that
includes Licensed Data, or (iii) is released to replace a Covered Model in
the same product or endpoint, in each case whose training run starts before
[date].

PATTERN C
"Model" means any machine learning model trained in whole or in part,
directly or indirectly, on Licensed Data or Derived Outputs, including by
distillation, whether or not it is a derivative work under applicable law.
"Derived Outputs" means data generated by a Model, including synthetic
examples, preference labels and logits.

"In whole or in part" is also how the FTC describes models and algorithms it has required companies to delete when developed with unlawfully obtained data [6]. If your grant reaches less far, a challenge to the data can reach models your license does not cover. Fuller grant language is in writing the AI training rights grant.

Which descendants each pattern reaches

List your stack's derivation techniques in a schedule and mark each in or out, because a model can inherit from licensed data through weights, retraining on the data or another model's outputs.

Merges and customer-built models cause the most surprises. A merge averaging a covered fine-tune with an outside model can ship under a new name, outside a name-based family. Models your customers build are not "trained by Licensee," so they need a flow-down or sublicensing clause for platform customers. Open-weight release puts every later derivative outside your control; see releasing open-weight models trained on licensed data.

Distillation and synthetic outputs need their own clause

Distillation transfers capability through outputs rather than weights, so a definition built only on weight inheritance misses it. In one published method, a differentially private teacher generates synthetic text that is then used to distill a student, which learns from those outputs rather than from the teacher's weights [7].

Model providers already draft for this. Anthropic's help center says its terms do not allow using outputs to train models that compete with its own, while non-competing uses such as sentiment analysis or content categorization tools are allowed [8]. Commentators question whether such terms bind parties who never accepted them [5].

Expect a data licensor to ask for comparable limits on outputs of models trained on its data. Settle three questions:

1. Generation. May a covered model produce synthetic data, preference labels or logits for training?
2. Downstream training. May those Derived Outputs train models outside the covered set, such as small on-device students?
3. Survival. Do Derived Outputs, and models trained on them, survive termination, or must synthetic sets be deleted with the data copies?

A workable buyer position treats students distilled from covered models as covered, with the same rights and survival, and asks the licensor to disclaim rights in ordinary outputs except verbatim reproduction of licensed records. See synthetic data generated from licensed data and who owns model outputs under a training data license.

Successor tests: weights, data, replacement and time

A successor clause holds up when it relies on facts your training logs can prove, not on product names that marketing can change.

• Weights test. The model is initialized in whole or in part from covered parameters. Checkpoint lineage proves it.
• Data test. Licensed manifest IDs appear in the run's data mixture. Data-loader manifests prove it.
• Replacement test. The model replaces a covered model in the same product or endpoint. It catches renamed models but is harder to prove; use it as a backstop.
• Time box. Tie coverage to the date a training run starts, not the release date, so a run that begins during the term and ships later is covered. Add a run-off period for runs in progress at expiry. For termination, see what happens to trained models when a data license ends.

Rights to past training do not extend forward on their own. In Bartz v. Anthropic, the class settlement received final approval in July 2026 [9]. A license that names one model generation works the same way. Successors built by an acquirer or affiliate need assignment and change-of-control terms.

How model scope changes the price

Broader model scope moves more of the data's long-term value to the buyer, so expect licensors to price it higher or attach a renewal trigger. Agree the structure before the number:

• a per-model fee with a pre-priced option for the next generation;
• a family grant capped at an agreed number of successors, then a renewal fee;
• a time-boxed grant covering every run that starts within the term;
• an enterprise-wide flat fee.

Price distilled and quantized variants together with their teacher, or per-model counts multiply quickly. Structures are compared in AI data license pricing structures and per-model vs enterprise-wide training licenses. SourceX does not publish prices; terms depend on scope, volume, history, rights and exclusivity, and are agreed per deal in writing.

The lineage register that makes model scope provable

Whatever pattern you sign, keep a record linking every released model to its parents, derivation method and licensed datasets, because disclosure rules and auditors will ask. California's AB 2013 requires developers of generative AI systems offered to Californians to post training-data documentation before each release or substantial modification, covering dataset sources or owners, whether datasets include licensed material and whether synthetic data was used [10]. AI Act Article 53(1)(d) requires a public training-content summary for each general-purpose AI model [11].

License terms get lost downstream: one audit found license omission above 70% and license errors above 50% on popular dataset hosting sites [12]. NIST's Generative AI Profile lists value chain and component integration among generative AI risks [13]. FTC orders have also reached derived models: the 2021 Everalbum order required deletion of models and algorithms developed using users' photos and videos [14], and the 2023 proposed Rite Aid order reached data, models or algorithms derived from collected images [15]. A register shows which models fall inside a lineage and which do not.

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

model_id: support-agent-v3-int4
parents: [support-agent-v3]
derivation: quantize        # fine_tune | adapter | distill_outputs | distill_logits | quantize | prune | merge | continued_pretrain | from_scratch
training_run_id: null       # no new training run
licensed_lineage:
  - manifest_id: LD-0142
    license_id: DLA-2026-017
    scope_pattern: B
    reached_via: weights    # weights | data | outputs
synthetic_inputs: []
release_channels: [api, on_device]
disclosures: { ca_ab2013_doc: "2026-09", eu_art53_summary: "v3" }
license_status: covered     # covered | option_needed | out_of_scope

See audit and usage-reporting rights for audit clauses and the AI data provenance guide for dataset-level records.

Drafting sequence and who owns each step

Start from the model roadmap rather than the licensor's template, so the definition describes models you will actually build.

1. Research lead. List descendants planned within the term: next base generation, distilled students, quantized on-device builds, merges, reward and embedding models.
2. ML platform. Confirm the derivation methods in use and what logs capture: checkpoint hashes, data-loader manifests, generator IDs for synthetic sets.
3. Counsel. Choose pattern A, B or C and draft Model, Successor Model and Derived Outputs.
4. Counsel and research lead. Attach the technique schedule, marked in or out.
5. Procurement. Price the options and record them in the AI data license term sheet.
6. Privacy reviewer. If the data contains personal information, confirm whether distilled or synthetic outputs inherit de-identification and no-re-identification duties.
7. ML platform. Turn on the lineage register before the first training run.

SourceX sources operational datasets from US companies and manages the commercial process, including licensing agreements and ongoing purchases. Every dataset goes through rights review and is delivered under a license that defines which records are included, what they can be used for, how long the license runs and how delivery happens. Datasets are sourced on request, so a request does not guarantee a match; state the scope pattern you need when you describe your data needs to SourceX.

Wording that turns a derivative clause into a dispute

Most model-scope disputes trace back to a few phrases; search the draft for them before signing.

Fallback positions are in the AI data license negotiation checklist. For general terms, see SourceX's guide to AI data license terms and its clause-by-clause agreement guide; the AI training data licensing hub maps other clauses.

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

Need data licensed for future model versions?

Describe the data you need and the models it must cover, from one fine-tuned model to successor generations and distilled students. SourceX looks for US businesses that hold that data, checks the data and each supplier's licensing permissions, and manages the license, delivery and future purchases. Specify your data and model-scope requirements.