OpenAI is asking U.S. policymakers to look harder at DeepSeek and other AI models it links to the People's Republic of China. In a policy proposal submitted for the Trump administration's "AI Action Plan" initiative, the company describes DeepSeek as "state-subsidized" and "state-controlled" and argues that certain models should face new limits.
The request marks a sharper turn in OpenAI's public position on the Chinese AI lab. OpenAI had already accused DeepSeek of "distilling" knowledge from OpenAI's models against its terms of service. The newer proposal goes further by tying DeepSeek to broader national security concerns, export policy and the handling of user data.
What OpenAI Is Asking Policymakers To Consider
The proposal recommends that the U.S. government consider banning models from DeepSeek and similar PRC-supported operations. OpenAI frames the issue around models that it calls "PRC-produced," especially where those models might be used in countries treated as "Tier 1" under the Biden administration's export rules.
OpenAI's central argument is that such limits would reduce privacy and "security risks," including the "risk of IP theft." The concern, as presented in the proposal, is that DeepSeek is subject to Chinese legal requirements that could force it to comply with demands for user data.
That argument puts AI model access into the same policy conversation as data security and compute access. Rather than treating AI models only as software products, OpenAI is presenting them as infrastructure that can raise geopolitical and commercial risks when deployed across borders.
Why DeepSeek Is At The Center Of The Fight
DeepSeek became more prominent earlier this year, and its R1 "reasoning" model is one of the models named in the source article. OpenAI's proposal treats DeepSeek as a case study for the risks it sees in PRC-linked AI systems.
At the same time, the source article notes an important uncertainty: it is not clear whether OpenAI's references to "models" mean DeepSeek's API, its open models, or both. That distinction matters because the technical and privacy questions are different depending on how a model is accessed and where it is hosted.
The article says DeepSeek's open models do not include mechanisms that would allow the Chinese government to siphon user data. It also says companies including Microsoft, Perplexity and Amazon host those open models on their own infrastructure. That complicates a broad reading of OpenAI's proposal, because the risk profile may differ between a hosted service and an open model running elsewhere.
The Evidence Question Remains Unsettled
The source article states that there is not a clear link between the Chinese government and DeepSeek. DeepSeek is described as a spin-off from a quantitative hedge fund called High-Flyer, not as a government agency or a clearly documented state-run project.
Still, the article also says the PRC has shown increased interest in DeepSeek in recent months. Several weeks before the article, DeepSeek founder Liang Wenfeng met with Chinese leader Xi Jinping. OpenAI's proposal appears to place that broader context beside its claims about legal exposure, data access and state influence.
That leaves the debate with two separate questions. One is whether Chinese law creates a real data-access risk for users of DeepSeek-related services. The other is whether DeepSeek itself should be treated as under the command of the PRC, which the source article says has not been clearly established.
OpenAI Later Softened The Message
After the story was published, OpenAI spokesperson Liz Bourgeois sent a statement on 3/15 8:38 p.m. Pacific. In that statement, OpenAI said it was not seeking restrictions on people using models like DeepSeek. Instead, the company described its position as a proposal to change U.S. export rules so more countries could access U.S. compute if their datacenters did not rely on PRC technology that presents security risks.
That follow-up changes the emphasis. The original submission, as described in the article, calls for the government to consider bans on "PRC-produced" models in certain countries. The later statement frames the goal as broader access to compute and AI, conditioned on datacenter choices.
The source article adds one more notable detail: OpenAI's own AI-powered deep research tool characterized the later statement as "equivocal" and said it used language that partially contradicted the stronger position in the original submission. That does not settle the policy question, but it underlines how sensitive the wording has become.
What The Proposal Signals For AI Policy
OpenAI's proposal shows how quickly AI competition is being pulled into export controls, datacenter policy and security screening. The dispute is not only about whether DeepSeek's models work well or compete with U.S. models. It is also about where models come from, who hosts them, what laws may apply to them and how governments should define risk.
For companies and policymakers, the hard part is precision. A broad phrase like "PRC-produced" can cover very different technical situations. An API operated by one company, an open model hosted by Microsoft, Perplexity or Amazon, and a datacenter technology requirement are not necessarily the same policy target.
OpenAI's submission brings those issues into one argument: restrict risky PRC-linked AI technology to protect privacy, security and intellectual property. The source article, however, shows that the details are still contested, including what OpenAI means by "models," how direct the government connection to DeepSeek is, and how far any proposed restriction should go.