US Government Unveils Measures to Counter Chinese AI Model 'Exploitation'

The White House Office of Science and Technology Policy announced measures to combat unauthorized distillation of US AI models by Chinese developers. Distillation involves training a new AI system using outputs from a parent model, like OpenAI's GPT-4o, to replicate capabilities at far lower cost—avoiding expensive research and processor expenses. The US government will promote wider information sharing among US developers, help detect extraction via proxy accounts (tens of thousands used), and work with industry to hold bad actors accountable. Officials estimate this costs Silicon Valley billions in annual profit.

Key targets include Chinese firms DeepSeek, Moonshot, and MiniMax, which use jail-breaking techniques to strip safety protocols and create cheap, open-weight models that mimic closed US versions. The memo clarifies support for open-source ecosystems but deems industrial-scale distillation unacceptable, as it undermines US R&D investments and produces unreliable models. Top US developers like OpenAI, Anthropic, and Google have raised alarms, sharing intel on extraction tactics. The move addresses economic challenges as Chinese open-source models undercut proprietary US systems, which rely on user fees to offset billions in infrastructure costs.