China and US plan September AI talks amid claims of stolen model 'watermarks'

This story captures the emerging second front of the AI race: not just who can build the best models, but who controls the intellectual property substrate they’re built on. If U.S. officials are prepared to say publicly that they see technical fingerprints of American LLMs inside Chinese models, it signals a move toward treating model weights and training tricks as strategic assets, not just trade secrets.

Strategically, bilateral AI talks framed around IP and ‘watermarks’ will shape how much cross-border collaboration remains feasible. If the U.S. pairs evidence of copied model internals with sanctions threats, Chinese labs will double down on domestic training stacks, and both sides will ramp up their compliance and counterintelligence operations around model training and evaluation.

For the AGI timeline, tighter IP enforcement and export controls could slow diffusion of the very highest-performing models while accelerating investment in homegrown alternatives. That might create a more fragmented global landscape where frontier capabilities are replicated regionally rather than shared globally. It also raises the odds that covert model exfiltration and forensic watermark detection become standard tools in statecraft, intertwining AI alignment debates with national security doctrine.