Licensing data for model training means getting a written grant from whoever holds rights in the data. The grant decides what a team may do with the data, with the models trained on it and with their outputs. For teams building superintelligence (SI) systems, the parts that matter are the licensed material, the permitted uses, rights in models and outputs, scope, term, warranties, the terms of the underlying sources and the records that show each of them. This explainer takes them clause by clause, from published licenses, standards, filings and position papers. It is research, never legal advice.
Why the written grant matters
In the EU, recital 105 of the AI Act says any use of copyright protected content requires the rightsholder's authorisation unless a copyright exception applies [5]. Article 4 of the 2019 Digital Single Market Directive allows text and data mining of lawfully accessible works unless rightholders have reserved that use, by machine-readable means for content made publicly available online [3]. Article 7(1) makes contract terms contrary to the exceptions in Articles 3, 5 and 6 unenforceable, research mining included, and Article 4 is not on that list [3].
Where rights have been reserved, the same recital says providers of general-purpose models need authorisation to mine the works [5]. The Act's duties for those providers, including a copyright policy and a public summary of training content, have applied since 2 August 2025 [5]. The Commission's template for that summary asks separately about datasets licensed from rightsholders [6].
In the United States, the Copyright Office released its report on training on 9 May 2025, and it was still a pre-publication version on 7 November 2026 [2]. The report describes voluntary licenses negotiated directly or collectively and recommends letting the licensing market develop without government intervention [1].
Licensing data for model training: the grant clause
A grant of "use" leaves the hard questions open. The authors of the Montreal Data License, a 2019 proposal, write that licenses tend to grant a right to use data without saying what use means, and they call a bare ban on commercial use "problematically ambiguous" [8]. Their taxonomy splits rights on the data, such as access, labelling and distribution, from rights with models: benchmarking, research, publishing, internal use, output commercialization and model commercialization [8].
Machine-readable vocabularies split the uses further. Really Simple Licensing (RSL) 1.0, a steering committee recommendation of 10 December 2025, has ai-train for training or fine-tuning, ai-input for input such as retrieval-augmented generation and grounding, and ai-index for retrieval indexes [9]. Where a term both permits and prohibits a use, the prohibition takes precedence [9]. An IETF working document of 14 September 2026, which states that it does not reflect working group consensus, separates training a generative model from using an asset as its input. It says legal agreements that permit a use likely apply over stated preferences [10].
Fine-tuning and training can be separate grants too. A March 2026 buyer-side paper from the FISD Alternative Data Council describes fine-tuning as adding information to a pre-trained model and model training as building a new one, "a more intensive use of source data" [12]. Retrieval needs its own words as well. The Copyright Office notes that retrieval-augmented generation "also involves the reproduction of copyrighted works" [1].
Models, outputs and derived data
Rights in the data and rights in what is built from it are separate questions. The Linux Foundation's Community Data License Agreement, CDLA-Permissive-2.0, places no restriction on "Results", which it defines to include machine learning models [13]. The Montreal design can permit commercial use of outputs while the trained model stays away from third parties, or let the model itself reach them [8]. The FISD paper proposes testing outputs by whether the licensed data can be reverse-engineered from them and whether they could substitute for it [12].
The sources disagree on whether a model is a derivative work of its training data. The Montreal authors call it "highly uncertain" [8]. Creative Commons' May 2025 guidance says that in many cases neither a model nor its outputs would be, then takes a conservative approach to ShareAlike, NonCommercial and NoDerivatives content [16]. The Copyright Office notes that model weights may contain copies of works in the training data [1].
Scope: exclusivity, territory, field of use and onward sharing
Shutterstock's 2025 annual report says its library content is typically licensed non-exclusively and custom content exclusively [17]. Creative Commons Attribution 4.0 grants a worldwide, non-sublicensable, non-exclusive license [15], and RSL can restrict use to listed country codes [9]. The field-of-use limits the Montreal paper lists include exclusions for health-related fields and military use [8]. A restriction on sublicensing, the same paper notes, "would also prevent third party contractors from working with the licensed data" [8]. Common Crawl's terms grant a license that is non-assignable, non-transferable and non-sublicensable [22].
Term and the end of the term
Reddit's 2024 registration statement describes data licensing arrangements with terms of two to three years, a right of access over the contractual period, a fixed or usage-based fee, and delivery through its data API and quarterly transfers [18]. Shutterstock offers the content on its websites under perpetual licenses, apart from some custom, editorial and music content with special terms, and licensed content survives a contributor's removal of it, apart from periodic removals over third-party rights concerns [17]. Creative Commons Attribution 4.0 lasts for the term of the licensed rights and ends automatically on a breach, with reinstatement if the breach is cured within 30 days of its discovery [15].
The FISD paper reports that vendors have traditionally required deletion of raw data at expiry and that some now seek obligations covering "model artifacts" [12]. As one buyer group's position, it proposes deleting raw datasets and artifacts that could substitute for them, such as high-fidelity retrieval embeddings, and calls retraining models to excise past use "commercially impracticable" [12]. The FTC has ordered models deleted in settlements over data it said was collected or used deceptively or illegally [19, 20]. Its 2021 order against a photo app developer covered "the models and algorithms it developed" from users' photos and videos [19], and a 2022 settlement required destruction of algorithms derived from children's data collected illegally [20]. Neither order concerns the expiry of a license.
Warranties, source terms and personal data
CDLA-Permissive-2.0 and Creative Commons Attribution 4.0 give the data as is, with no warranty of title or non-infringement [13, 15]. Shutterstock's annual report describes a warranty against infringement for its content licenses, with indemnification generally capped at $10,000 to $250,000, and does not say its training data licenses carry the same terms [17].
A dataset license sits on top of the terms of its sources. FineWeb is released under the Open Data Commons Attribution License and is "also subject to CommonCrawl's Terms of Use" [23]. That license governs only the database and leaves its contents to other terms [24]. A study of nearly 4,000 public datasets found that less than 33% were restrictively licensed, while over 80% of the source content in widely used text, speech and video datasets carried non-commercial restrictions [27]. The Common Pile team calls redistribution under an incorrect license "license laundering" [28]. The provenance standard for datasets that state their limits cites the audits that found licenses missing or wrong on dataset hosting sites.
A license from a rights holder does not settle personal data questions. Creative Commons Attribution 4.0 does not license privacy or publicity rights [15]. The GDPR requires personal data to be collected for specified purposes and not further processed in a way incompatible with them [29], and the European Data Protection Board said in December 2024 that models trained with personal data "cannot, in all cases, be considered anonymous" [30].
What to keep on record
Datasheets for Datasets asks whether a dataset is distributed under a license or terms of use, for a link to the terms and any fees, and whether third parties have imposed restrictions on the data [31]. MLCommons' Croissant 1.0 gives a dataset a license field and recommends the URL of a known license from the SPDX list [32]. Kept on each record, the license answers the same questions item by item. Licenses on public resources can change after a team downloads them, as the comparison of financial sentiment methods records for one public model checkpoint. Public records can carry licensed fields of their own, as the account of where federal contract data comes from shows.
Rillor's training data acquisition and licensing finds the publishers and data holders that hold the data a team needs, agrees the use, the term and onward sharing in writing before delivery, and delivers every record with the license it came under.
Read the full report
The full report sets out ten findings, from copyright exceptions to the record each grant needs, with a table of the uses three published vocabularies name and all 33 sources.