
LAION took the set down on December 19, 2023, after a Stanford report identified links in it to suspected child sexual abuse material, and released Re-LAION-5B on August 30, 2024 with 5,526,641,167 pairs.
This page reads both sets from the side of the people whose images are linked. The licenses are CC BY 4.0 for the LAION-5B metadata and Apache-2.0 for Re-LAION-5B, and the images keep their own copyright.
Within AI training data licensing, the set is the free link counterpart to licensed image data: the dataset brings no payment to the holders of the linked images. A German suit over its creation awaits a Federal Court of Justice pronouncement scheduled for December 17, 2026, so no ruling exists as of 2026-11-10.
LAION e.V.: The Hamburg Association That Publishes the Dataset
LAION-5B is published by LAION e.V., a registered association in Hamburg entered in the association register of the Hamburg Local Court under VR 25085. The table lists the identity facts as of 2026-11-10.
| Attribute | Value | Source |
|---|---|---|
| Legal name | LAION e.V. | laion.ai/impressum; bundesgerichtshof.de I ZR 281/25 |
| Register | Vereinsregister Amtsgericht Hamburg, VR 25085 | laion.ai/impressum |
| Status | registered association recognized as non-profit (tax status) | laion.ai/impressum |
| Seat | Hamburg, Germany | laion.ai/impressum |
| Website | laion.ai | laion.ai |
| Founded | not stated | n/a |
| Role in court | defendant in BGH I ZR 281/25 | bundesgerichtshof.de |
The non-profit recognition is a tax status and not part of the legal name, so the name is written LAION e.V. throughout.
The Federal Court of Justice notice for case I ZR 281/25 names the association as defendant, which makes the Hamburg register entry the identity that matters for any legal question about the set. No founding year, founder or board member is stated on this page, because the sources used here do not record them.
For a holder of a linked image, the practical consequence is that the party behind the dataset is a small registered association with a named seat, not a company that sells data, and it offers no license fee or contract for the images it links.
LAION-5B is one of the public corpora the parent page compares among the major AI training datasets.
What LAION-5B Holds: Links, Captions and Metadata Instead of Images
LAION-5B holds 5.85 billion pairs of an image URL and the text that accompanied the image, plus metadata, and the images themselves stay on the websites that host them.
LAION states in its post of March 31, 2022 that the set has 5.85 billion image-text pairs, of which 2.3 billion are English. The Federal Court of Justice notice describes the entries as links and descriptions, not images.
Each row therefore points outward: it names where a picture sits on the web and carries the words that were attached to it, and a user who wants the picture must fetch it from the host.
If the host removes the picture, the row can survive without it, and if the picture stays online, the row does not change its copyright position.
The stated size differs between sources. LAION's LAION-5B post and the Federal Court of Justice notice give 5.85 billion, the UK High Court judgment in [2025] EWHC 2863 (Ch) cites 5.85 billion URL-text pairs at paragraph 47 and adds LAION subsets of about 3 billion, and LAION's later Re-LAION-5B page gives 5.8 billion for the original set. The page keeps both figures and does not average them.
| Attribute | Value | Source |
|---|---|---|
| Published | 2022-03-31 | laion.ai/blog/laion-5b |
| Image-text pairs | 5.85 billion, as LAION states | laion.ai/blog/laion-5b |
| Same set, later LAION page | 5.8 billion | laion.ai/blog/relaion-5b |
| English pairs | 2.3 billion, as LAION states | laion.ai/blog/laion-5b |
| What an entry holds | image link, caption, metadata; no image file | bundesgerichtshof.de I ZR 281/25 |
| Size cited by the UK High Court | 5.85 billion URL-text pairs, [2025] EWHC 2863 (Ch) para 47 | caselaw.nationalarchives.gov.uk |
| Upstream source | Common Crawl web archive | commoncrawl.org/about |
| Status | taken down 2023-12-19 | laion.ai/blog/relaion-5b |
The set is a list compiled by LAION, not a store of works, which is also why LAION describes its license as covering the compilation.
The links were drawn from Common Crawl's archive, which is upstream of LAION-5B. For how that crawl works and who else builds on it, see the page on what Common Crawl contains.
The December 2023 Takedown and the Stanford Abuse-Material Report
LAION took LAION-5B down on December 19, 2023, after a Stanford report identified links in it to suspected child sexual abuse material. LAION's Re-LAION-5B post refers to that report as the trigger for the takedown and states that it identified 1,008 links in the set.
LAION then ran its own review. It matched the set against hash lists supplied by partners and counted 2,236 matches with suspected links, which it calls the upper bound for links leading to potential abuse material.
The 2,236 includes the 1,008 links from the Stanford work, and LAION removed the matched links from the set it later released.
For anyone holding a downloaded copy, the takedown does not recall copies. Links and captions copied before December 19, 2023 stay where they are, and LAION points users to the new release. The page makes no statement about models trained on the set, and it does not describe the material itself.
Two figures matter for reading the episode: the 1,008 links in the Stanford work and LAION's own 2,236 matches, which LAION frames as an upper bound because its hash lists cover suspected as well as confirmed items.
Both are small against a set of billions of links, and both are LAION's statements about its own data. The Stanford work is cited here only as LAION's Re-LAION-5B post refers to it, as the report that preceded the takedown.
Re-LAION-5B: What the August 2024 Release Removed
Re-LAION-5B, released on August 30, 2024, holds 5,526,641,167 text-link pairs, LAION's count after removing the matched links from the original set. It also carries a different license, Apache-2.0, named in the table below.
The table compares the two releases using LAION's figures, as of 2026-11-10.
| Attribute | LAION-5B | Re-LAION-5B | Source |
|---|---|---|---|
| Released | 2022-03-31 | 2024-08-30 | laion.ai blog posts |
| Pairs | 5.85 billion | 5,526,641,167 | laion.ai blog posts |
| License | CC BY 4.0, metadata | Apache-2.0 | laion.ai blog posts |
| Matched links removed | n/a | 2,236, an upper bound, including 1,008 from the Stanford report | laion.ai/blog/relaion-5b |
| Research variant | n/a | 1.121 % removed (65 million) | laion.ai/blog/relaion-5b |
| Research-safe variant | n/a | 3.044 % removed (176 million of 5.8 billion) | laion.ai/blog/relaion-5b |
| Status | taken down 2023-12-19 | released | laion.ai/blog/relaion-5b |
The removal shares are LAION's own and are measured against 5.8 billion, the size LAION's Re-LAION page gives for the original, not against the 5.85 billion of the 2022 post.
Re-LAION-5B is therefore a filtered successor and not a new crawl: it keeps the structure of the original, links with captions and metadata, and drops the matched entries.
The count 5,526,641,167 is written in full because LAION gives it to the digit. LAION publishes two variants of the new set. The research variant removes 1.121 % of the samples, 65 million, and the research-safe variant removes 3.044 %, 176 million of 5.8 billion, as LAION states.
LAION-5B and Re-LAION-5B Licenses: What They Cover and What They Leave Out
LAION released the LAION-5B metadata under CC BY 4.0 and Re-LAION-5B under Apache-2.0, and in both cases the license covers LAION's compilation of links and captions, not the linked images, which remain under their own copyright.
The two licenses are not interchangeable and the page never merges them. What CC and Apache terms mean in general is covered on the page about dataset licenses for AI training.
| Release | License | What it covers | What it leaves out | Source |
|---|---|---|---|---|
| LAION-5B | CC BY 4.0 | the metadata LAION compiled | the images, which remain under their own copyright | laion.ai/blog/laion-5b |
| Re-LAION-5B | Apache-2.0 | the released link set | not stated on record | laion.ai/blog/relaion-5b |
CC BY 4.0 is an attribution license, and in this case it attaches to the metadata LAION compiled. A holder of a linked image keeps the copyright in it, whatever license LAION applies to its list, and a license from LAION grants no rights in the picture because LAION does not own them.
The step from LAION-5B to Re-LAION-5B changed the license from CC BY 4.0 to Apache-2.0 and did not change who owns the images. The license also does not settle whether training on the linked images is lawful, which is a question for the courts below.
LAION-5B in Court: The Hamburg Photographer Suit and Image-Model Cases
Three court records on file name LAION-5B: a German suit against LAION itself and two suits against a developer of image models. The table lists them alphabetically as of 2026-11-10.
| Case | Court and docket | How LAION-5B appears | Status, as of | Source |
|---|---|---|---|---|
| Andersen v. Stability AI | N.D. Cal. 3:23-cv-00201-WHO, filed 2023-01-13 | artists allege infringement through LAION-based training of image models | DMCA claims dismissed, copyright claims survive (Dkt. 223, 2024-08-12); discovery, as of 2026-10-19 | archive.org docket |
| Getty Images v. Stability AI (UK) | High Court [2025] EWHC 2863 (Ch); Court of Appeal CA-2026-000007 | judgment cites LAION-5B at 5.85 billion URL-text pairs | training claim abandoned for lack of evidence of UK training; secondary infringement dismissed 2025-11-04; appeal heard, judgment reserved 2026-11-12, as of 2026-11-16 | caselaw.nationalarchives.gov.uk |
| Kneschke v. LAION | LG Hamburg 310 O 227/23; OLG Hamburg 5 U 104/24; BGH I ZR 281/25 | reproduction of a photograph to build LAION-5B | heard 2026-09-03; pronouncement scheduled 2026-12-17, as of 2026-11-10 | bundesgerichtshof.de |
In the US case, artists allege that image models trained on LAION-based data infringe their works. The court dismissed the DMCA claims on 2024-08-12 (Dkt. 223) and the Copyright Act claims survive, and the case is in discovery as of 2026-10-19. The page on Andersen v. Stability AI carries the claims and rulings.
In the UK case, the High Court judgment of 2025-11-04 cites LAION-5B at paragraph 47, and the training claim was abandoned for lack of evidence that training took place in the UK.
Secondary infringement was dismissed, and the Court of Appeal case tracker shows the appeal (CA-2026-000007) part heard on 2026-11-11 and judgment reserved on 2026-11-12, as of 2026-11-16. The record sits on the page about Getty Images v. Stability AI (UK).
A photographer sued LAION e.V. over the reproduction of a photograph to build LAION-5B. The claim failed at the Regional Court of Hamburg (310 O 227/23, 2024-09-27) and the Higher Regional Court of Hamburg dismissed the appeal (5 U 104/24, 2025-12-10).
The Federal Court of Justice heard the revision (I ZR 281/25) on 2026-09-03 and scheduled its pronouncement for 2026-12-17, so no Federal Court ruling exists as of 2026-11-10; claims and status sit on the Kneschke v. LAION case record.
No court has ruled that LAION's license makes training on the linked images lawful, and none of the three records decides that question for the dataset as a whole.
The German case concerns the act of building the set, the UK case concerns where model training took place, and the US case concerns the models built on LAION-based data.
Each court asks a different question under a different law, so a result in one does not transfer to another. The status lines above are the only ones the page states, and each carries its date. This page is information, not legal advice.
An Image Linked in LAION-5B: Benefits, Risks, Downsides and Alternatives for the Holder
For a photographer or image archive whose pictures are linked in LAION-5B, the dataset brings no payment and no contract, and removal from later versions does not reach copies already downloaded.
The four lists below separate what the link set does for a holder from what it costs, and they end with the options the holder has, including leaving things as they are.
Benefits
- The image itself was never copied into the set, only its link and caption.
- The copyright stays with the holder, unaffected by LAION's license.
Risks
- Links and captions persist in copies made before 2023-12-19.
- The German case on dataset building is not final until the Federal Court of Justice rules.
- Image-model suits that rely on LAION-based training are unresolved.
Downsides
- Images reachable for free through a public link set can be harder to license for a fee.
- EU and German opt-outs for online works must be machine-readable under Article 4(3) of Directive (EU) 2019/790 and section 44b(3) UrhG, so a plain-text notice may not suffice.
Alternatives
- Reserve rights in machine-readable form for future crawls, as the page on machine-readable TDM reservations describes.
- License images directly with captions and metadata, as described in selling photos and images for AI training.
- Ask the publisher to drop links from later versions.
- Not license at all and leave the images as they are.
None of these changes the status of copies already made. The list is neutral: the benefits are real but narrow, and the risks depend on court outcomes that are not yet final. The decision to reserve rights, to license directly or to do nothing stays with the holder, and doing nothing is a valid choice.
Image Licensing, Opt-outs and Dataset Law Beyond This Page
This page covers LAION's two link sets; removal requests, opt-outs, other image corpora and the wider law each have their own page.
- Machine-readable TDM reservations: how to reserve rights a machine can read.
- Removing content from AI training datasets: removal from later versions.
- Opting out of AI training: platform settings.
- Licensed vs scraped training data: the licensed alternative.
- Creators selling content for AI training: photographers and other creators.
- EU text and data mining exception: Art. 4 and the opt-out.
- Germany: text and data mining under UrhG sections 44b and 60d: the German provisions.
- Copyright and AI training across jurisdictions: fair use and TDM exceptions.
- AI copyright lawsuits over training data: the case tracker.
- DataComp and CommonPool: another image-text pool.
- Alternatives to selling data, including not selling at all: not licensing at all.
Frequently Asked Questions About LAION-5B
What is the LAION-5B dataset?
LAION-5B is a set of 5.85 billion image-text pairs, published on March 31, 2022 by LAION e.V. Each entry holds a link, a caption and metadata, not an image file. LAION states that 2.3 billion of the pairs are English, and its later Re-LAION page gives 5.8 billion for the original size.
When was LAION-5B released, and is it still available?
LAION-5B was released on March 31, 2022 and taken down on December 19, 2023. LAION released the successor Re-LAION-5B on August 30, 2024, with 5,526,641,167 pairs. The Re-LAION release carries the Apache-2.0 license and a research and a research-safe variant, as the Re-LAION section above sets out.
Does LAION-5B contain copies of my photos?
No, it contains links and captions, not image files. The image stays on its host website and under its own copyright. A row in the set names where the picture is and what text accompanied it, so a photo can be linked without being copied into the dataset.
Is it legal to train AI on LAION-5B?
No final ruling on record decides it. The German case awaits the Federal Court of Justice pronouncement of 2026-12-17, the UK training claim was abandoned, and the US claims are in discovery. The wider question sits on the page is web scraping for AI training legal.
Agents and advisers can use the introducer form.