Child abusers are increasingly using AI-generated "deepfakes" to produce simulated child abuse imagery, initiating a cycle of sextortion that can persist for years. This alarming trend has led to calls for stricter regulations and global cooperation to combat the issue.
Both the Labour and Conservative parties in the UK are advocating for a ban on all explicit AI-generated images of real people. However, there is a lack of global consensus on regulating this technology, and the ease of creating such images remains a significant challenge.
A recent discovery by researchers at Stanford University highlighted the presence of hundreds, possibly thousands, of instances of child sexual abuse material (CSAM) within the Laion dataset, one of the largest training sets for AI image generators. Despite the dataset containing about 5 billion images, researchers employed automated scanning to identify questionable content and reported it to law enforcement.
Although Laion's creators removed the dataset from public access and emphasized that they had not distributed the explicit images, the illicit training data has already been integrated into AI systems worldwide. This poses a serious risk as AI image generators can create explicit content due to their exposure to such material.
The issue is further complicated by the open-source nature of datasets like Laion, which are widely used for independent AI research. Unlike proprietary datasets from companies like OpenAI, these open-source datasets are freely accessible, making it challenging to ensure they are free from explicit content.
OpenAI, for example, has implemented measures to filter explicit content from its Dall-E 3 model, but the effectiveness of these efforts remains to be verified. Companies like OpenAI also maintain control over their AI models by requiring requests to be processed through their systems, allowing for additional filtering and oversight.
AI safety experts argue that a multi-layered approach, combining content filtering with purpose-built tools, is more effective than relying solely on AI models trained to avoid creating explicit images. Moreover, completely clean training data may not always be beneficial, as AI models need exposure to real-world scenarios to recognize and report explicit content accurately.
Kirsty Innes, the director of tech policy at Labour Together, emphasized the importance of maintaining space for open-source AI development, suggesting that innovative solutions to address these challenges could emerge from such initiatives.
While short-term measures focus on targeting purpose-built "nudification" tools and their creators, long-term strategies must grapple with the complex task of regulating and understanding AI technologies. The overarching question remains: how can we effectively limit a technology that continues to evolve and challenge our understanding?
