Artificial intelligence experts reported on Friday that they have removed over 2,000 web links containing suspected child sexual abuse imagery from a database utilized to train popular AI image-generation tools. The LAION research database is a substantial collection of online images and captions that has been a resource for prominent AI image-generating systems like Stable Diffusion and Midjourney.
A study conducted last year by the Stanford Internet Observatory revealed that the database included links to sexually explicit images of children, facilitating the creation of photorealistic deepfakes depicting children by certain AI tools. Following the release of this report in December, LAION swiftly eliminated its dataset. Eight months later, LAION announced in a blog post that it collaborated with the Stanford University watchdog group and anti-abuse organizations in Canada and the United Kingdom to address the issue and publish a refined database for future AI research.
Despite these efforts, an older and minimally filtered version of Stable Diffusion, identified by Stanford as one of the most popular models for generating explicit imagery, remained accessible until Thursday. It was only then that the New York-based company Runway ML removed it from the AI model repository Hugging Face, citing a planned discontinuation of research models and code that were not actively maintained.
The revised LAION database release coincides with increased global scrutiny on how certain technology tools are being utilized to produce or circulate illegal images of children. Earlier this month, the city attorney of San Francisco filed a lawsuit to shut down a cluster of websites permitting the creation of AI-generated naked images of women and girls. Moreover, the alleged sharing of child sexual abuse images on the messaging platform Telegram prompted French authorities to press charges against the platform’s founder and CEO, Pavel Durov, on Wednesday.