Human Rights Watch (HRW) has raised alarms about the unauthorized use of Australian children’s personal photos to train artificial intelligence (AI) tools. These images are scraped from the internet without the consent or knowledge of the children or their families, leading to significant privacy violations and potential exploitation.
The HRW report highlights that photos of Australian children are included in LAION-5B, a vast dataset compiled by scraping internet content. This dataset is utilized to train various AI models, which can then be used to generate deepfakes, posing serious risks to the children’s safety and privacy. Hye Jung Han, a researcher and advocate for children’s rights and technology at Human Rights Watch, emphasized the urgent need for the Australian government to implement laws protecting children’s data from misuse by AI technologies.
HRW’s analysis discovered that LAION-5B contains identifiable photos of Australian children, often with detailed captions or URLs revealing their names and locations. For instance, one image depicts two young boys holding paintbrushes, with accompanying information disclosing their full names, ages, and the preschool they attend in Perth, Western Australia. Such personal data can be traced back to the children’s real-life identities, heightening their vulnerability.
The report identified 190 photos of children from across Australia, but this figure likely underrepresents the true extent of children’s data in the LAION-5B dataset. These images span various stages of childhood, capturing intimate moments such as births, school activities, and cultural events. Some photos, particularly those of First Nations children, include identifiable details and culturally sensitive contexts, further complicating the privacy concerns.
Many of these images were initially shared in limited, private settings and are not easily discoverable through standard online searches. They originated from personal blogs, social media, school websites, and professional photography sessions. Some were uploaded years before the creation of LAION-5B, reflecting an era when online privacy considerations were less stringent.
Additionally, HRW found instances where LAION-5B contained photos from sources that had taken measures to protect privacy. For example, a photo of two boys making funny faces was extracted from an unlisted YouTube video intended to safeguard the individuals’ privacy. This action contravenes YouTube’s terms of service, which prohibit the unauthorized scraping of identifiable personal information.
The incorporation of children’s data into AI training sets like LAION-5B poses ongoing privacy threats due to the persistent nature of AI models. These models can reproduce exact copies of the training data, including sensitive information such as medical records and personal photos. Despite some companies’ efforts to prevent data leakage, these safeguards have repeatedly failed, exacerbating the risks.
Indigenous Australians face additional risks, as many First Nations communities restrict the reproduction of images of deceased individuals during mourning periods. The perpetual retention of such data in AI systems undermines these cultural practices and poses significant ethical concerns.
The misuse of children’s photos to train AI models facilitates the creation of realistic deepfakes, which can be weaponized to exploit and harm minors. A recent incident in Melbourne saw approximately 50 girls’ social media photos manipulated into sexually explicit deepfakes, which were then disseminated online. The ease and speed with which AI tools can generate lifelike deepfakes exacerbate the potential for widespread abuse.
In response to these findings, LAION, the organization behind LAION-5B, acknowledged the presence of children’s personal photos in the dataset and committed to removing them. However, they argued that the responsibility for safeguarding children’s data rests with the individuals who upload it online.
Australia’s Attorney General, Mark Dreyfus, has proposed legislation banning the nonconsensual creation or distribution of sexually explicit deepfakes. Nevertheless, Human Rights Watch contends that broader measures are necessary to protect children’s personal data from any form of misuse, including the unauthorized digital manipulation of their likenesses.
The Australian government plans to introduce reforms to the Privacy Act, including the development of the Children’s Online Privacy Code, which aims to safeguard children’s rights in data collection and usage. Human Rights Watch advocates for this code to prohibit the scraping of children’s data for AI training and to offer robust mechanisms for addressing and remedying harm.
In conclusion, Human Rights Watch calls for comprehensive AI regulations that prioritize data privacy protections for all individuals, especially children. By implementing these safeguards, the development of AI technology can be steered towards upholding, rather than violating, children’s rights.
