Generative AI Safety: AI Leaders Commit to 'Safety by Design' Principles for Child Protection

We stand at a critical juncture with the rapid advancement of generative AI. While this technology offers immense potential for innovation and creativity, it also presents a significant challenge to child protection. The misuse of generative AI has already begun to manifest within our communities, creating a pressing need for proactive measures. This moment offers a rare opportunity to steer the development of generative AI down a path that prioritizes the safety and well-being of children. The choices made now will shape the future of online safety and the digital environment for generations to come.

Generative AI significantly lowers the barriers to creating and distributing harmful content, including CSAM. A single predator can now generate vast quantities of abuse material with unprecedented speed. This technology enables the manipulation of existing images and videos of children into new forms of abuse, or the creation of entirely fabricated CSAM. The influx of AIG-CSAM poses a severe threat to an already strained child safety ecosystem. It complicates the already difficult task for law enforcement to identify and rescue children in active harm's way, and it opens new avenues for victimization. Furthermore, generative AI can be used to scale grooming and sextortion efforts, reduce social and technical barriers to sexualizing minors, and even facilitate information sharing among child sexual predators by providing instructions and advice on abuse and evading detection. While the prevalence of AIG-CSAM is currently small, its rapid growth necessitates an immediate and proactive response.

The 'Safety by Design' principles, adopted by leading AI companies, address child safety risks across the entire AI lifecycle. These commitments include: * **Development and Training:** Building generative AI models that proactively address child safety risks and responsibly sourcing training datasets to prevent the inclusion of CSAM and CSEM. This involves detecting and removing such content and mitigating risks associated with depictions of children alongside adult sexual content. * **Testing and Evaluation:** Incorporating iterative stress-testing and feedback loops to understand and address a model's capability to produce abusive content. This includes structured, scalable testing throughout development and integrating findings back into model improvement. * **Deployment and Distribution:** Releasing models only after thorough evaluation for child safety, and safeguarding products and services from abusive content and conduct. This also involves responsibly hosting models, whether first-party or third-party, and implementing clear policies against child safety violations. * **Maintenance and Evolution:** Continuously understanding and responding to evolving child safety risks, investing in research and future technology solutions, and actively fighting CSAM, AIG-CSAM, and CSEM on platforms. This includes preventing services from scaling access to harmful tools and removing models or services designed for abuse.

Beyond data integrity, the 'Safety by Design' principles emphasize rigorous testing and content provenance. Iterative stress-testing strategies are crucial to continuously assess a model's potential to generate abusive content. If developers don't proactively test for these capabilities, malicious actors will exploit them. Furthermore, content provenance solutions are vital for distinguishing AI-generated content from authentic material, which is essential for combating AIG-CSAM. Companies are working to develop state-of-the-art media provenance or detection solutions, potentially incorporating imperceptible watermarking or other techniques. Models are released only after being trained and evaluated for child safety, with protections implemented throughout the process. Products and services are safeguarded against abusive content and conduct, and user feedback mechanisms are incorporated to empower users and identify misuse.

The collective commitments made by these AI leaders serve as a powerful call to action for the entire industry. Thorn and its partners urge all companies involved in the development, deployment, maintenance, and use of generative AI technologies and products to adopt these 'Safety by Design' principles. Demonstrating a dedication to preventing the creation and spread of CSAM, AIG-CSAM, and other acts of child sexual abuse and exploitation is not just an ethical imperative but a crucial step towards forging a safer internet and a brighter future for children. By working together, the industry can ensure that the transformative potential of generative AI is realized responsibly, with child protection at its core.