EPIC Urges the NTIA to Tackle Privacy Harms, Bias, and Regulatory Hurdles in New Comment on AI Model Openness

Yesterday, EPIC submitted comments to the National Telecommunications and Information Administration (NTIA) in response to its request for comments on the risks and benefits of dual-use foundation models with widely available model weights. President Biden’s recent AI Executive Order tasked the NTIA with issuing a report on dual-use foundation models—sophisticated AI systems designed to be used for a wide variety of purposes—and the potential impacts of making these systems more transparent.

The debate over AI model openness is nuanced, so rather than recommend either closed AI models or open AI models, EPIC highlighted how key privacy risks, bias issues, and regulatory hurdles manifest under closed versus open AI models. On data privacy, EPIC underscored the deep privacy risks inherent to most commercial AI training data, described common data security vulnerabilities in both closed and open AI systems, and discussed how model openness and privacy-preserving techniques like differential privacy intersect. On AI bias, EPIC described common AI debiasing techniques and highlighted how model openness brings both benefits (e.g., better independent assessments of AI bias) and risks (e.g., malicious actors using greater model transparency to circumvent or exploit debiasing techniques). Finally, EPIC urged the NTIA to consider the relative capacity of different actors to oversee AI systems when weighing the benefits and risks of model openness. With closed AI systems, independent assessments become harder but assigning liability for harm becomes easier; under open systems, the opposite is true. Ultimately, where the NTIA lands in this debate must depend, at least in part, on whether independently assessing AI models or assigning liability for AI harms is a greater regulatory hurdle.

EPIC has long been a leading advocate for stronger oversight and regulation of AI systems, including efforts to ban government use of biometric surveillance tools and extensive research on AI risks, including reports on generative AI harms and government AI procurement. EPIC looks forward to working with federal agencies like the NTIA to meet the expectations of President Biden’s AI Executive Order and ensure that AI systems serve the public interest.