The report released by the NGO Human Rights Watch, this Monday (10), points out that photos of Brazilian children and adolescents are used to create artificial intelligence tools without their knowledge or consent. The analysis concluded that LAION-5B, a dataset used to train popular AI tools and built from scraping information from much of the internet, contains links to identifiable photos of children.
In addition to the images, the platform also provides the names of some children in the captions or in the URL where the image is stored. The investigation also concluded that, in many cases, identities are traceable, with information about when and where the child was at the time the photo was taken. Among the examples, the organization says that one of the modified photos shows a two-year-old girl with her newborn sister. The image caption details the names of the children, the location of the hospital, and the date the baby was born. In total, 170 photos of children from ten Brazilian states were found.
However, the organization states that this number is a small sample of the problem, as LAION-5B contains 5.85 billion images and captions, and Human Rights Watch analyzed 0.0001% of the total.
The images show newborn babies, small children blowing out candles, dancing at home, students in the middle of dance presentations, and teenagers celebrating Carnival. There are images from different dates and photos from up to ten years before the creation of LAION-5B. The research was conducted by Hye Jung Han, a child rights and technology researcher at the NGO. She says she decided to investigate AI platforms after discovering that more than 80 young people in different states of Brazil were targets of sexual harassment after having their photos modified into deepfakes by schoolmates. From the searches, Han noticed that the images used by the tools are not from image banks or linked to famous celebrities. The researcher states that the images found in the tools were published on school websites, personal blogs, and YouTube channels.