bào dào报道：zhōng guó中国yán jiū rén yuán研究人员yòng用měi guó美国rén gōng zhì néng人工智能xùn liàn训练jūn shì军事xì tǒng系统
Reuters reviewed more than 80 Chinese research papers and patents.
The report said that some researchers linked to China’s military used AI models from U.S. companies to train smaller Chinese models.
This method is called “model distillation.” A powerful model first provides answers, and those answers are then used to train a new model.
Researchers have used this method to process military software code and to develop tools for monitoring social media.
Some small models have also been placed in drones and other military equipment.
In this way, even if communications are cut off, the equipment can analyze video, detect targets, and help with navigation.
U.S. AI companies said they do not allow China’s military to use their services and are concerned that the new models do not have the original safety protections.
The Chinese government denied improperly using U.S. technology and said that U.S. criticism is unfair.
Experts say that model distillation is common, but using someone else’s model without permission may create security and intellectual property problems.
At the same time, small models can usually learn only part of the original model’s abilities.