Public domain material is material no copyright protects, because the term has run out or the law never granted one, and openly licensed material is still protected but released under terms that let anyone copy it, so both can enter a training set without paying the author.
This is the free end of AI training data licensing, and it holds only within each license's conditions, with privacy, contract and database rights untouched.
The page gives the term rules for the US, EU and UK with worked dates, Creative Commons' own statements on what each condition means for training, and the corpora built from this material, each with its section, source and check date.
The statutes were re-read on 2026-11-17, and the sources and dating rules are in how we source, date and check every figure.
What counts as public domain or openly licensed material for training?
Four kinds of material differ by who still holds rights: public domain works have none, openly licensed works keep a holder who has granted conditions in advance, protected works copied under an exception keep a holder who never agreed, and licensed works come under a contract.
"Publicly available" is not "public domain". Text reachable on the open web stays protected unless its term ended or a license covers it. Creative Commons states that its licenses apply only when copyright permission is required, so an open license matters only where copying needs permission.
An exception is a different route. EU training on protected online works rests on Article 4 of Directive (EU) 2019/790, which the holder can reserve by machine-readable means, and UK section 29A covers non-commercial research only. Public domain material needs neither, and the exceptions are compared in copyright and AI training across jurisdictions.
| Material | Who holds rights | What a developer needs | Example on record | Source |
|---|---|---|---|---|
| Public domain (term expired or never protected) | no one | nothing under copyright | works of authors dead 70+ years (EU, UK); US federal works | 17 U.S.C. 105; Directive 2006/116/EC; CDPA 1988 s.12 |
| Openly licensed | the author, conditions granted in advance | to follow the license conditions | CC-licensed works | creativecommons.org, checked 2026-11-17 |
| Protected, used under an exception | the author, no agreement | the exception's conditions (EU: no machine-readable reservation) | DSM Directive Art. 4 | eur-lex.europa.eu |
| Licensed | the author, by contract | the contract | paid licenses (see the border below) | n/a |
Table as of 2026-11-17.
When copyright ends: US, EU and UK terms with worked dates
Copyright in the EU and UK ends 70 years after the end of the year the author died, and in the US it ends 70 years after death for works created from 1978 and 95 years from first securing for older renewed works, each term running to 31 December.
The EU rule is Directive 2006/116/EC Article 1(1), life plus 70 years, with Article 8 counting terms from 1 January of the year after the event, so the count never starts on the date of death. The UK rule is CDPA 1988 section 12(2): 70 years from the end of the calendar year of death.
In the US, 17 U.S.C. 302(a) sets life plus 70 years for works created on or after 1978-01-01, section 304(b) gives 95 years from the date of securing to works in their renewal term on 1998-10-27, and section 305 runs every such term to the end of the calendar year.
Three worked dates follow, written as calculations from these provisions and not as any court's or office's statement.
An author who died in 1955 is protected in the EU and UK through 2025-12-31 and in the public domain from 2026-01-01. A US work secured in 1930 and renewed had its 95 years end on 2025-12-31, so it is free from 2026-01-01. A work secured in 1931 is free from 2027-01-01.
The table does not settle pre-1923 rules, unrenewed works, notice formalities or unpublished works, which this page does not cover.
Terms differ by country: an author who died in 1960 reaches the EU term's end on 2030-12-31 by the Directive's arithmetic, even if a US work of that author has already left US protection. Other jurisdictions are not covered here.
| Jurisdiction and work | Term | Runs to | Provision | Source |
|---|---|---|---|---|
| US, created on or after 1978-01-01 | life of the author plus 70 years | end of the calendar year (s.305) | 17 U.S.C. 302(a) | govinfo.gov |
| US, anonymous, pseudonymous or made for hire | 95 years from first publication or 120 from creation, whichever expires first | end of the calendar year | 17 U.S.C. 302(c) | govinfo.gov |
| US, pre-1978 work in its renewal term on 1998-10-27 | 95 years from the date copyright was originally secured | end of the calendar year | 17 U.S.C. 304(b), Pub. L. 105-298 | govinfo.gov |
| US federal government works | no copyright | n/a | 17 U.S.C. 105(a) | govinfo.gov |
| EU | life of the author plus 70 years | counted from 1 January after the death | Directive 2006/116/EC Arts. 1(1), 8 | eur-lex.europa.eu |
| UK, literary, dramatic, musical, artistic | 70 years from the end of the calendar year of death | 31 December | CDPA 1988 s.12(2) | legislation.gov.uk |
| EU, reproductions of public domain visual art | no copyright or related rights unless original | n/a | Directive (EU) 2019/790 Art. 14 | eur-lex.europa.eu |
Table as of 2026-11-17.
US federal works: free of copyright, with three limits
Works of the US Government have no US copyright: 17 U.S.C. 105(a) says "Copyright protection under this title is not available for any work of the United States Government".
Three limits apply:
- The rule covers federal works only, and it is never extended to state government works.
- Works transferred to the Government by assignment or bequest can still be protected.
- The rule is US law, and this page does not record how the same work is treated abroad.
Public bodies elsewhere release data under licenses rather than by statute. The EU high-value datasets rule, Implementing Regulation (EU) 2023/138, in force 2023-02-09 and applying from 2024-06-09, requires CC0, CC BY 4.0 or less restrictive terms, and its Recital 6 exempts libraries, museums and archives from the free-of-charge requirement.
The open-data statutes are covered in governments and open data as AI training data sources.
Digitized copies of out-of-copyright works and the access terms around them
Scanning a public domain work does not create a new copyright in the EU for visual art: under Directive (EU) 2019/790 Article 14, a reproduction of a work of visual art whose term has expired is not protected unless the reproduction is itself original.
Member States had to transpose it by 2021-06-07 (Article 29(1)), and the text says nothing about scanned books.
Holders of digitized collections control access by terms of use, whatever each volume's copyright status. Institutional Books 1.0 releases 983,004 volumes of 1,075,899 scanned (arXiv 2506.08300, 2025-06-10), and its Hugging Face card offers noncommercial early-access terms, with commercial use including AI model training requiring contact with the Institutional Data Initiative (checked 2026-11-09).
What Creative Commons says each license condition means for training
Creative Commons states that its licenses apply only when copyright permission is required, and where they do apply, each condition carries a duty in training: a link for attribution, the same license for shared models under ShareAlike, no commercial gain at any stage under NonCommercial, and no training under NoDerivatives.
These are statements in Creative Commons' practical guidance, which it says is likely to lead to overcompliance, and each is quoted as its statement, checked 2026-11-17. The ShareAlike line is conditional: the same license is required only if models or outputs based on the content will be shared publicly.
The guidance speaks about works. A license placed on a compilation, such as LAION-5B's metadata under CC BY 4.0 while the images keep their own copyright, is a different question, handled under dataset licenses for AI training.
One buyer's criteria show how the conditions bite commercially: Mercor states that content under non-commercial or no-derivatives licenses is not a fit for its video listing (checked 2026-11-09).
| Condition | What Creative Commons states for AI training | Source |
|---|---|---|
| All licenses | apply "only when copyright permission is required" | creativecommons.org, checked 2026-11-09 |
| BY (Attribution) | "attribution could be a simple link to the source of the dataset used to train the model" | creativecommons.org, checked 2026-11-17 |
| NC (NonCommercial) | "all stages, from copying the data during training to sharing the trained model, must not be for commercial gain" | creativecommons.org, checked 2026-11-17 |
| ND (NoDerivatives) | ND-licensed content is "not be used as training data" | creativecommons.org, checked 2026-11-17 |
| SA (ShareAlike) | if models or outputs based on SA content "will be shared publicly", the developer uses "the same CC license as the original works" | creativecommons.org, checked 2026-11-17 |
Table as of 2026-11-17, ordered as the license element letters; no court ruling on these conditions in training is recorded.
Training corpora built from public domain and openly licensed text
Three text corpora on record are built for training around public domain, openly licensed or scanned library text: Pleias's Common Corpus, the Common Pile v0.1 and Institutional Books 1.0, and only the last records a contact step for commercial training.
The Common Pile v0.1, released 2025-06-05 by EleutherAI and collaborators, is "an eight terabyte collection of openly licensed text" from 30 sources, with two 7B Comma v0.1 models trained on 1T and 2T tokens. Its license and access terms are not recorded here.
Common Corpus is named in a frontier developer's EU training summary. Pleias SAS states 2,267,302,720,836 tokens, first released November 2024, while its paper says about two trillion, and SpaceXAI's EU training summary for Grok 4.7, placed on the EU market 2026-09-21, names Common Corpus with OpenAlex and Crossref metadata. It is hosted by Hugging Face but built by Pleias.
The license on each document governs, and Common Corpus has no single dataset license.
| Corpus | Built by | Size as stated | Rights basis as stated | Terms for training | Source |
|---|---|---|---|---|---|
| Common Corpus | Pleias SAS | 2,267,302,720,836 tokens (Pleias); paper: about two trillion | public domain and openly licensed text, license per document | named in SpaceXAI's Grok 4.7 EU summary | huggingface.co; media.x.ai |
| Common Pile v0.1 | EleutherAI and collaborators | 8 TB, 30 sources | "public domain and openly licensed text" (paper title) | not recorded | arXiv 2506.05209 |
| Institutional Books 1.0 | Institutional Data Initiative (Harvard Law School Library collection) | 983,004 volumes, 242B tokens | scanned library volumes (rights basis not recorded) | noncommercial early access; commercial AI use by contacting IDI | arXiv 2506.08300; huggingface.co |
Table as of 2026-11-17, alphabetical; each value is as its builder states it.
How protected text ends up in sets labeled open
Protected text enters "open" sets in three ways: public domain sources mixed with unlicensed ones, a license label that differs between a publisher's own pages, and expiry criteria set by the dataset builder rather than a court.
LLaMA 1's Books share of 4.5 percent (85 GB) combined Project Gutenberg with Books3 (arXiv 2302.13971, 2023-02), so a public domain source sat beside a set of protected books.
Common Voice is CC0 in its community playbook, while a later Mozilla post says "a permissive licence", and both wordings stay visible. Rights labels in a corpus are its builder's own assessment, not a court's or a registry's.
The remedy is a record per work: source, term basis, license text and version, as set out under training data provenance and the chain of rights.
Releasing material openly, licensing it or holding it back: benefits, risks, downsides and alternatives
A holder of out-of-copyright collections or of its own works can release them openly, license access for payment, attach conditions or keep them back, and not releasing or licensing at all is a valid outcome.
Benefits of open release
- Reach and attribution: CC licenses require attribution, which Creative Commons states could be a link to the source.
- Public bodies meet the EU high-value datasets rule with CC0 or CC BY 4.0.
Risks
- A release cannot be recalled once copied.
- A CC license does not apply where no permission is needed.
- Scans of public domain visual art carry no new EU right to fall back on.
Downsides
- Mercor states that content already open access is not a fit for its publisher listing (checked 2026-11-09).
- No payment arrives for the material itself.
Alternatives
- Access terms with a commercial contact route, as Institutional Books uses.
- A paid delivery service on open content: Wikimedia, LLC offers Enterprise APIs and named Amazon, Meta, Microsoft, Mistral AI and Perplexity as new partners on 2026-01-15, with no deal values or partner pricing disclosed.
- NonCommercial or ShareAlike terms as a condition on commercial training.
- Not releasing or licensing at all, covered in alternatives to selling data, including not selling at all.
Beyond free material: paid licenses, unlicensed copies and the law behind each term
This page covers material that needs no payment; the pages below cover the paid route, unlicensed copies and the law each term rule sits in.
Licensed data comes under a contract that sets scope, term and payment, and how it compares with scraping is covered in licensed vs scraped training data. This page is information, not legal advice.
- How AI companies get training data: the parent of this page.
- Pirated books and shadow libraries in AI training: unlicensed copies.
- Major AI training datasets, The Pile and Common Pile: the corpora.
- Institutional Books contents and access terms: the Harvard-collection set.
- Libraries, archives and museums licensing collections for AI and books as AI training data: the holders and the market.
- Wikimedia Enterprise AI deals: the paid delivery of open content.
- The EU text and data mining exception and the opt-out and the EU database right and licensing datasets: the EU rules.
- UK copyright and AI training and fair use and AI training in US law: the UK and US rules.
Frequently asked questions
Can AI be trained on public domain works without permission?
Yes as far as copyright goes, because none subsists in a public domain work. Privacy, contract and database rights still apply.
Is content that is publicly available online in the public domain?
No. Online content stays protected until its term ends, and open licenses apply only where copyright permission is required.
Which US works are in the public domain in 2026?
By the statutes' arithmetic, renewed works secured in 1930 or earlier, plus federal works. Works created from 1978 run for life plus 70 years under 17 U.S.C. 302(a).
Does a ShareAlike license reach a trained model?
Creative Commons states that if models or outputs based on ShareAlike content are shared publicly, the same CC license applies. No court ruling on it is recorded.