The United States Navy has warned service members to avoid using a hugely popular Chinese generative artificial intelligence (AI) app. The service has essentially deep-sixed DeepSeek, and in an emailed memo told sailors and civilian personnel that the AI platform shouldn’t be used in any capacity citing “potential security and ethical concerns associated with the model’s origin and usage.”
DeepSeek recently surpassed ChatGPT as the Apple App Store’s top free app, and its rise caused shares of AI chipmakers Nvidia and Broadcom to see massive stock price drops on Monday – resulting in a combined $800 billion loss in market cap and a 3.1% drop in the Nasdaq, CNBC reported.
Although the main concern with DeepSeek is its ties to China, the U.S. military has cautioned personnel against the dangers of AI in general. A 2023 guidance from the U.S. Navy to servicemembers had warned against any commercial generative AI, highlighting that the technology poses a “unique security risk” and that there remains a threat of “data compromise” when using the technology.
“It is the right decision,” suggested technology industry analyst Roger Entner of Recon Analytics.
“Members of the U.S. military should not provide access to their data and especially questions that they ask to a foreign power that is our adversary,” Entner told ClearanceJobs. “This is providing the Chinese government and its military valuable insights about what is going on in the head of U.S. servicemembers. Being in front of a search engine or AI tool is the most honest moment people have during the day. People can lie with their answers, people can’t lie with the questions.”
Rolling in the Deep
DeepSeek may have just hit the market, but there is already speculation it could eventually face a ban not dissimilar from that of TikTok due to its Chinese ownership. There are also concerns it may have relied on the output from U.S.-made generative AI platforms to better train the Chinese platform.
The fact that the U.S. Navy was quick to caution against its use was thus seen as a no-brainer.
“DeepSeek as a model contains a set of biases that may run contrary to the objectives of United States military personnel-like any model, malicious patterns can be embedded into the end product that induces unwanted behavior when certain triggers are met,” explained David Brauchler, technical director & head of AI and ML security at cybersecurity consultancy NCC Group.
“For example, questions about U.S. military strategy could trigger the model to respond with poor-quality suggestions. It’s worth noting that all models have some form of bias and failure cases that are difficult or infeasible to detect in advance,” Brauchler told ClearanceJobs.
Then there is the issue that DeepSeek as a service submits data to China-hosted servers without security or data privacy guarantees. It was for similar reasons that U.S. lawmakers had called for the ban of TikTok.
“However, it is important to distinguish these two categories, as the DeepSeek model is open-weight and available to the public, and can only be polluted insofar as its output quality is impacted –e.g. censorship of responses – because it can be downloaded and run locally, but the DeepSeek service provides no such assurances,” Brauchler added.
Further Bans Coming – Almost Certainly
Even as the Trump Administration has done a reverse course on TikTok, it is unlikely the White House would throw DeepSeek a lifeline. Trump has called for the U.S. to maintain its lead in the development of AI. For those reasons, more restrictions could likely be imposed on DeepSeek and similar platforms coming from China.
“The U.S. government could easily justify a restriction on its departments leveraging DeepSeek’s services due to the risk of sensitive data being captured by foreign intelligence,” said Brauchler. “With more difficulty, the government may be able to justify banning the platform within the United States as a whole. Banning the model itself would require significant effort and would result in questionable implications for technological freedom and data imports. I anticipate that the latter move would be contested and adjudicated, and by the time the question is resolved, a new model may have captured the public’s attention.”