Stanford researchers found links to child abuse, hundreds of them, in the artificial intelligence training visual set called LAION-5B used by Stable Diffusion.
Very large data sets are needed to train artificial intelligence. The larger the data set, the better the artificial intelligence performs. LAION It also creates data sets for artificial intelligence developers. Stanford Internet Observatory, LAION-5B It revealed hundreds of links to child abuse in the data set.
LAION-5B, creator of Stable Diffusion Stability AI It was also used by. Stanford researchers, who started examining the data set in September 2023, aimed to find out whether there was child abuse content in this data set and, if so, how many there were. According to the study results, at least 1679 contents Links to images containing child abuse were found. This information was also shared with institutions such as PhotoDNA and the Canadian Child Protection Centre.
This data set was also used in Stable Diffusion
According to the information on LAION’s website, the dataset does not store images, but creates an internet index that includes text descriptions of the images and links to the images. At Google, Imogen is an older version of LAION-5B for training generative artificial intelligence. LAION-400M had used it. While the company said that 400M was not used in later versions. Imogen researchers are also included in the data set. “There is a lot of inappropriate content, including child abuse, racist slurs, and harmful social stereotypes.”” he stated.
Stanford researchers say the existence of these contents does not directly affect the outputs of the data set While LAION stated that they implemented a zero tolerance policy against such harmful content and that they would temporarily withdraw the data set from publication. On the other hand, retraining artificial intelligence trained with this data poses a bigger problem.
Previously, in the USA, state prosecutors submitted to Congress the use of artificial intelligence in child abuse and with productive artificial intelligence He called for the convening of a committee to prevent the production of such content.
Source :
https://www.theverge.com/2023/12/20/24009418/generative-ai-image-laion-csam-google-stability-stanford