LAION-400M

Transparency and accountability 🙈

The LAION-400M dataset is seen to have several important transparency limitations.

• Source data. LAION provides little information on the websites and sources used to create its dataset.

• Scraping methodology. Details about LAION's web scraping techniques and filtering processes are not fully disclosed.

• Copyright and privacy. It is unclear if proper consent was obtained for all images and text, or how copyright issues were addressed.

• Demographic representation. There's limited information on the demographic distribution and potential biases in the dataset.

• Content moderation. The extent and methods used to remove inappropriate or offensive material are not fully detailed.

• Quality assessment. Criteria for assessing the quality and relevance of image-text pairs are not thoroughly explained.

• Downstream impact. There is little analysis on how using this dataset impacts the models trained on it.

Risks and harms 🛑

The LAION-400M dataset is accused of violating privacy, enabling the generation of offensive, hateful, explicit and derogatory content, and perpetuating biases due to unfiltered, large-scale web-scraped data.

Research, advocacy 🧮

• Birhane A., Luccioni A.S. et al (2023). Into the LAIONs Den: Investigating Hate in Multimodal Datasets

• Thiel D., Hancock J. (2023). Identifying and Eliminating CSAM in Generative ML Training Data and Models

• Birhane A., Prabhu V.U., Kahembwe E. (2021). Multimodal datasets: misogyny, pornography, and malignant stereotypes