US Report Flags Chinese AI Models as Security Threat

A recent US government report highlights that Chinese AI models are trailing behind American counterparts in performance, cost, security, and adoption. Released on Tuesday, the report describes Chinese models as 'adversary AI', citing security flaws and censorship as risks to AI developers, consumers, and US national security.

The report follows former President Donald Trump's AI Action Plan, which called for an evaluation of Chinese models' capabilities and alignment with state narratives. DeepSeek, a prominent Chinese AI company, has faced criticism in the US for data theft and amplifying state narratives.

According to the evaluation by CAISI, DeepSeek's models scored lower than US models across 19 public and internal benchmarks and were more susceptible to being compromised by malicious users for hacking and cybercrime activities.

The report also claims that Chinese state censorship is embedded in DeepSeek models, based on a new benchmark developed with the Department of State. DeepSeek's R1-0528 model was found to be 25.7% aligned with Chinese state narratives when prompted in Chinese.

Meanwhile, DeepSeek's 'open-weight' models have helped China catch up in the global AI adoption race. Downloads of DeepSeek models on the Hugging Face platform surged by 1,000%, with Alibaba Cloud's Qwen models increasing by 135%.

US Commerce Secretary Howard Lutnick stated on social media that the report aids in ensuring 'continued US leadership in AI'. He noted that 'DeepSeek lags significantly, especially in cyber and software engineering'.