AIWiki
Malaysia
Back to all articles
AI Foundationslaiondatasetsopen data

LAION

4 min readUpdated November 2026
LAION
Type
Non-profit open AI organisation (Germany)
Founded
By volunteer researchers; active from 2021
Best known for
LAION-400M and LAION-5B image-text datasets
Licence
Datasets released for research; code and models per project
Key models trained
Stable Diffusion, Imagen (and successors)
Related
Common Crawl, Stable Diffusion, AI and Copyright
LAION (Large-scale Artificial Intelligence Open Network) is a German non-profit organisation that releases open datasets and models for AI research. It is best known for publishing large collections of image–caption pairs scraped from the public web, which became the training foundation for several high-profile text-to-image models.[1]

History

LAION-400M appeared in August 2021 with 400 million image–caption pairs, an attempt to reproduce the dataset OpenAI used to train CLIP while releasing OpenAI's code and weights but not its data. A far larger successor, LAION-5B, was released in March 2022: 5.85 billion CLIP-filtered image–text pairs, then the largest freely available dataset of its kind, funded by Doodlebot, Hugging Face and Stability AI.[2][3]

The organisation also released OpenAssistant in April 2023, an open chatbot assistant built with more than 13,500 volunteers contributing around 600,000 human-written data points; that project has since been shut down, while datasets and models remain available on Hugging Face.[1]

How the Dataset Works

LAION does not host images. Its pipeline reads the HTML of the Common Crawl web archive, extracts tags and treats their alt text as captions, then uses CLIP similarity filtering to discard pairs whose image and caption do not match. What is published is a list of URLs and captions — researchers download the images themselves.[1][4] LAION-5B's NeurIPS paper describes the pipeline in detail.[3]

Controversy

The datasets have drawn significant criticism. Studies found pairs depicting pornographic, hateful and stereotypical content. In December 2023 the Stanford Internet Observatory reported 3,226 suspected links to child sexual abuse material in LAION-5B (1,008 externally validated), prompting LAION to remove LAION-5B and LAION-400M under its zero-tolerance policy for illegal content; a cleaned release, Re-LAION-5B, followed in August 2024.[1][5] LAION was named as a non-party in Getty Images' lawsuit against Stability AI and was separately sued by a German photographer seeking removal of his work — a case the Regional Court of Hamburg dismissed in September 2024 in a ruling described as a landmark for text-and-data-mining exceptions in the EU.[1][6]

Why It Matters for AI

LAION demonstrated that large, usable training corpora could be built and published openly by volunteers, lowering the barrier to image-model research outside large companies. It also made the provenance problem visible: web-scraped datasets inherit copyright, privacy and content-safety liabilities that downstream models carry with them — a debate reflected in the site's coverage of AI and copyright and data contamination.[7]

>See Also

🇲🇾Malaysian Context

🇲🇾 Relevance to Malaysia: open datasets are how Malaysian universities, startups and independent researchers can train or fine-tune vision models without buying expensive proprietary data — relevant to MDEC-backed AI initiatives and local model efforts. But LAION's controversies are a practical lesson in data governance: Malaysia's Personal Data Protection Act 2010 restricts the use of personal data scraped from the web, and the Malaysia AI Governance Framework expects organisations to document data sources and quality.[8] For Malaysian content creators — illustrators, photographers, designers — LAION illustrates why provenance and opt-out mechanisms matter, since images published online can enter training corpora. Local AI teams are increasingly expected to keep dataset documentation (sources, licences, filtering steps) as part of responsible-AI practice.

References

  1. ↑Wikipedia. LAION. https://en.wikipedia.org/wiki/LAION
  2. ↑LAION. (2021). LAION-400M announcement. https://laion.ai/blog/laion-400m/
  3. ↑LAION. (2022). LAION-5B: a new era of open large-scale multi-modal datasets. https://laion.ai/blog/laion-5b/ — see also NeurIPS paper, arXiv:2210.08402
  4. ↑Android Police. Open data: how Common Crawl and LAION shape AI training. https://www.androidpolice.com/open-data-ai-training-common-crawl-laion/
  5. ↑Stanford Internet Observatory. (2023). Analysis of LAION-5B dataset security issues (as reported by LAION). https://laion.ai/
  6. ↑Bloomberg News. (2023). The future of AI relies on a high school teacher's free database. https://www.bloomberg.com/news/features/2023-04-24/the-future-of-ai-relies-on-a-high-school-teacher-s-free-image-database-powers-ai-unicorns
  7. ↑Hugging Face. LAION dataset collection. https://huggingface.co/datasets?search=laion
  8. ↑Malaysia AI Governance Framework. https://www.mdec.my/
Back to all articles