Dataset licensing and usage rights are set by the license attached to a dataset, and often by the terms of the sources inside it. This explainer compares three families of terms: open licenses, noncommercial and research-only terms, and commercial agreements. For superintelligence (SI) developers it asks each text the same three questions: what purposes it allows, what it requires when a changed copy is shared, and whether it says anything about trained models. Each answer restates the license text, and the explainer is research, never legal advice.
Dataset licensing and usage rights, side by side
| Terms | Purpose limit | Sharing the data or a changed copy | Trained models or results | Source |
|---|---|---|---|---|
| CC0 1.0 | Any purpose, commercial included | No conditions, rights waived | Not named | [4] |
| PDDL 1.0 | None, recipients "may use this work commercially" | Changes may be shared or kept secret | Not named | [9] |
| CC BY 4.0 | None | Adapted material may be shared, with attribution notices | Not named, "all media and formats" | [1] |
| ODC-By 1.0 | None | Free to share and modify, subject to its attribution requirements | Not named | [8] |
| CC BY-SA 4.0 | None | Shared adaptations under the same or a compatible license | Not named | [2] |
| ODbL 1.0 | None | Public Derivative Databases stay under the ODbL or a compatible license, internal use excluded | Not named, Produced Works need a notice | [7] |
| CDLA-Permissive-2.0 | None stated | Shared with the agreement's text | Results, machine learning models included, unrestricted | [12] |
| CDLA-Sharing-1.0 | None stated | Published data and Enhanced Data under the same agreement | Results of computational use carry no obligations | [13] |
| CC BY-NC 4.0 | "NonCommercial purposes only" | Adaptations for NonCommercial purposes only | Not named | [3] |
| C-UDA 1.0 | "solely for Computational Use" | Each recipient bound to the C-UDA | Models holding no more than a de minimis portion of the data are Results, unrestricted | [14] |
| IMDb non-commercial datasets | "personal and non-commercial use" | Not to build "any kind of online/offline database" | Not named | [26] |
| Data Subscription Agreement for AWS Marketplace | Use and modify, no service bureau use | No access of any kind for third parties | Models are Derived Data, owned by the subscriber | [27] |
The first six rows are the Open Definition's recommended licenses, which must allow use "for any purpose" [10, 11]. The two CDLA rows are Community Data License Agreements [12, 13].
Open licenses: public domain, attribution and share-alike
Open Definition 2.1 requires an open license to allow use, redistribution, modification and compilation for any purpose, sale included [10]. Its six recommended licenses fall into three pairs: CC0 and the Public Domain Dedication and License (PDDL) with no conditions, CC BY and the Open Data Commons Attribution License (ODC-By) with attribution, and CC BY-SA and the Open Database License (ODbL) with share-alike [11]. No Creative Commons NonCommercial or NoDerivatives license is on its lists, and the Common Pile authors state that NC and ND licenses "are not considered open under this definition" [11, 20].
CC0 waives copyright and related rights, database rights included, "for any purpose whatsoever", and falls back to an unconditional license where a waiver is invalid [4]. It waives no trademark or patent rights and disclaims responsibility for clearing other people's rights [4].
In CC BY, CC BY-SA and the ODbL, conditions turn on sharing. Creative Commons states that attribution, ShareAlike and NoDerivatives are "triggered only upon public sharing" [6]. The ODbL turns share-alike on when a Derivative Database is used publicly, and says use inside an organisation is not public [7]. A Produced Work made from the database, such as an image or a text, needs a notice when used publicly and does not itself create a Derivative Database [7].
Government texts set similar terms. U.S. law defines an "open license" for a data asset as free to the public with no restrictions on copying, publishing, distributing, transmitting, citing or adapting it [34]. The UK's Open Government Licence 3.0 allows commercial use and exempts personal data and third-party rights it cannot license [35].
The Common Pile authors also note that an open license "does not necessarily imply that the rights holder has specifically contemplated use of their content to train LLMs" [20].
What noncommercial and research-only terms restrict
CC BY-NC 4.0 defines NonCommercial as "not primarily intended for or directed towards commercial advantage or monetary compensation" [3]. Creative Commons says the definition "does not turn on the type of user": a nonprofit's use can breach it, and a company's use does not necessarily do so [5]. Under 4.0, the NC term applies to any use of a licensed database restricted by copyright or database rights, "even if the database is not publicly shared" [5]. Creative Commons says it cannot advise on what counts as commercial [5].
Its 2025 guidance on training is deliberately conservative, by its own account [6]. Under it, training copies of an NC work and later use and distribution of the model would need to be noncommercial [6]. The same guidance says that in many scenarios compliance with the license conditions is not required for training [6]. The financial sentiment note records a public model checkpoint that later gained a CC BY-NC-SA 4.0 license.
"Research-only" names no standard license. ImageNet's terms of access say the researcher "shall use the Database only for non-commercial research and educational purposes", bind a for-profit employer to the same terms and allow access to be ended at any time [24]. IMDb offers data subsets "for personal and non-commercial use", may withdraw permission at any time and licenses commercial use separately [26]. The authors of the Montreal Data License called the commercial-use restriction "problematically ambiguous", and noted that research is the first step toward products [17].
In the EU, Article 3 of the 2019 copyright directive lets research organisations and cultural heritage institutions mine lawfully accessed works for scientific research, and Article 7(1) makes contract terms contrary to it unenforceable [32].
Commercial terms set by contract
The Data Subscription Agreement for AWS Marketplace, a template dated 14 July 2022, is one published example [27]. It grants "a nonexclusive, worldwide, nontransferable license" to receive, retain, use and modify the data and to create Derived Data [27]. It bars giving any third party access, acting as a service bureau and using de-identified data to infer anyone's identity [27]. Derived Data includes models, and the subscriber owns it apart from the provider's data inside it [27]. The marketplace explainer covers how a license like this comes with a listing.
Some texts sit between open and commercial. The Computational Use of Data Agreement (C-UDA) limits use to "Computational Use" and states that "Artificial intelligence models trained on Data (and which do not include more than a de minimis portion of Data) are Results", which it leaves unrestricted [14]. RAIL-D licenses put use restrictions on data and require downstream redistribution to carry them [16].
When the contents carry their own terms
Several dataset licenses cover the collection and leave its contents to other terms. COCO licenses its annotations under CC BY 4.0, says the consortium "does not own the copyright of the images" and points users to the Flickr Terms of Use [25]. The ODbL and ODC-By govern the database "and not the contents of the Database individually" [7, 8].
FineWeb is released under ODC-By, and its card adds that use "is also subject to CommonCrawl's Terms of Use" [29]. Those terms say crawled content may carry its owners' own terms, and the user indemnifies the foundation against third-party claims arising from use of crawled content "in connection with artificial intelligence, machine learning, or other similar technologies" [28]. Dolma's card says users are also bound by the terms of the original sources [30]. The contract review benchmarks report follows one CC BY 4.0 dataset into the later benchmarks built from its contracts.
What audits found about dataset license labels
The Data Provenance Initiative audited more than 1,800 text datasets in 2023 [18]. More than 70% of licenses for popular datasets on GitHub and Hugging Face were "Unspecified", and 66% of the Hugging Face licenses analyzed sat in a different use category, often more permissive than the author intended [18]. GitHub, Hugging Face and Papers with Code suggested licenses that were too permissive 29%, 27% and 16% of the time [18].
A 2025 follow-up across nearly 4,000 datasets found 25%, 33% and 32% of text, speech and video datasets licensed noncommercially, while over 80% of each modality's content carried undocumented restrictions in its sources [19]. A 2025 paper reports that its automated system judged only 605 of 2,852 datasets with commercially viable license terms, or 21%, permissible for commercialization once their dependencies were traced [21]. A July 2026 preprint traced 232,270 dataset-to-model-to-application chains and found that 62.3% passed through at least one artifact with no declared license [23].
What to record for each dataset
The SPDX License List gives each license a standard short identifier, such as CC-BY-NC-4.0, ODbL-1.0 or C-UDA-1.0 [15]. A record per dataset can hold that identifier, the license text, the date it was read and each source's terms, so a change stays visible. Dolma, released in August 2023, moved to ODC-By on 15 April 2024 [30]. CC BY 4.0 says a licensor who stops distributing the material or offers other terms does not end the license already granted [1].
Rillor's data marketplace sets use, term and onward sharing in writing before access starts. The license grant explainer reads one grant clause by clause.
Read the full report
The full report adds the Open Definition's license list, the clauses behind each table row, EU database rights and U.S. compilation copyright, a table of five license audits and all 35 sources.