US-China AI Tensions Rise Over Copying Claims

Anthropic, OpenAI and Google have accused Chinese companies of unfairly copying their artificial intelligence technologies over the past two years. These claims became a central point of discussion as Chinese leader Xi Jinping visited President Donald Trump in Washington this week. American firms like Anthropic and OpenAI lead the artificial intelligence race, but Chinese startups like Z.ai and Moonshot are growing fast. Silicon Valley argues that Chinese companies are copying U.S. technologies using a technique called distillation. Many experts note that claims from Anthropic and OpenAI are overblown. Charles O'Neill, a head of model training at Baseten, stated that the narrative that Chinese tech capabilities come entirely from an Anthropic model is not as true as people say. First developed in the early 2010s, distillation helps researchers build efficient technologies to run on cheaper hardware. Geoffrey Hinton, a former Google researcher who helped create the technique, recently explained that one model acts as a teacher and another as a student. Recently, firms have used this technique to collect data from competitor systems by setting up multiple accounts on services like Claude or OpenAI's GPT to generate data for training new systems. Legal experts say distillation might violate the Defend Trade Secrets Act, but U.S. courts have not yet ruled on this. Copyright law does not clearly apply because distillation copies general behavior rather than exact text. Critics point out that American firms have faced similar issues. Anthropic settled a major lawsuit last year by agreeing to pay a large amount to a group of authors and publishers after a judge found the company illegally used copyrighted books to train its systems. In June, Anthropic sent a letter to U.S. lawmakers accusing Chinese tech giant Alibaba of distilling its technologies. Chinese firms have not responded publicly to these accusations. Distillation does not give companies access to underlying computer code, only output data like words and characters. Rehaan Ahmad, co-founder of alphaXiv, noted that firms using outputs must still build a powerful base system first, which requires extra work, money and computing power. Companies try to curb distillation by shutting down suspect accounts, but doing so risks blocking real users. Lino Le Van, another researcher at alphaXiv, stated that stopping distillation is basically impossible.