Tensions between the United States and China over artificial intelligence development flared up again in late April 2026, when the Trump administration formally pledged to curb what it labels industrial-scale, unauthorized intellectual property extraction from leading U.S. AI models, just one day before Chinese AI developer DeepSeek launched its latest high-performance frontier model, DeepSeek V4.
In an official memorandum issued by the White House Office of Science and Technology Policy (OSTP) on Thursday, April 23, OSTP Director Michael Kratsios detailed allegations that foreign entities, primarily based in China, have been running coordinated, large-scale campaigns to distill proprietary capabilities out of cutting-edge U.S.-developed AI systems. According to Kratsios, bad actors leverage networks of tens of thousands of fake proxy accounts to bypass platform detection, while using jailbreaking methods to access protected proprietary information hidden within these models. This systematic siphoning, Kratsios argued, lets foreign competitors exploit billions of dollars in U.S. research investment and innovation to build competing products.
While Kratsios noted that distilled models do not match the full performance of the original U.S.-built systems, he emphasized that they can achieve comparable results on common industry benchmarks at a tiny fraction of the development cost for their creators. Beyond commercial unfairness, Kratsios claimed these unauthorized distillation campaigns also let bad actors deliberately remove built-in safety protocols and alignment safeguards that ensure original AI models operate responsibly and remain ideologically neutral and fact-based.
To counter these alleged activities, the memorandum outlined a four-pronged strategy for the Trump administration: first, it will share timely intelligence with U.S. AI development companies detailing attempted large-scale unauthorized distillation, including specific tactics used and the actors behind them; second, it will facilitate tighter cross-sector coordination between private AI firms to collectively disrupt these campaigns; third, it will partner with industry to develop standardized best practices for detecting, mitigating, and remediating large-scale distillation attacks, while helping firms strengthen their defensive systems; and finally, the administration will explore new enforcement and policy measures to hold bad actors accountable for these activities.
The White House’s public warning landed just 24 hours ahead of DeepSeek’s April 24 launch of DeepSeek V4, a moment that highlights growing anxiety in Washington over the rapid pace at which Chinese AI developers are closing the performance gap with the United States’ top frontier models. Unlike many firms that keep training methods private, DeepSeek has been transparent about its use of knowledge distillation, a common training technique where a smaller “student” model learns from the outputs of larger, more capable “teacher” models. When the company launched its V3 model in January 2025, it openly confirmed it used knowledge distillation during training. For its new V4 model, DeepSeek says it has advanced the technique with a new method called On-Policy Distillation (OPD), which draws on outputs from 10 separate teacher models to refine its student model’s outputs, speeding up the training cycle significantly.
According to DeepSeek’s published research, its top-tier DeepSeek-V4-Pro-Max outperforms OpenAI’s GPT-5.2 and Google’s Gemini-3.0-Pro on standard industry reasoning benchmarks, while its lighter, more efficient DeepSeek-V4-Flash-Max matches the performance of those two leading U.S. models at a far lower development and inference cost. The company also noted that its V4 line’s overall performance is only 3 to 6 months behind the very latest frontier models from U.S. developers, OpenAI’s GPT-5.4 and Google’s Gemini-3.1-Pro.
The current friction around AI distillation is not new. The launch of DeepSeek V3 in 2025 sent ripples through global financial markets, as investors reacted to the emergence of a low-cost Chinese AI model that could compete head-to-head with leading U.S. offerings. During a Senate hearing that January, U.S. Commerce Secretary Howard Lutnick claimed DeepSeek had built its model “dirt cheap” by sourcing Nvidia AI chips through third-party countries and leveraging open training data from Meta. That said, then-recent comments from President Trump struck a different tone: in February 2025, he argued that the development of lower-cost AI was an inevitable technological shift that could ultimately benefit the United States, calling falling AI development costs “a very good development.”
Criticism of Chinese AI firms’ distillation practices faded for nearly a year before resurfacing in early 2026, when U.S. AI leaders published formal accusations of so-called “distillation attacks” that reignited the political debate. In a February 12 memo to the U.S. House Select Committee on China, OpenAI accused DeepSeek of using distillation techniques to “free-ride on the capabilities developed by OpenAI and other U.S. frontier labs,” adding that it had detected new obfuscated methods designed to bypass existing safeguards against misuse of model outputs, and that past attempts to stop the activity had not been fully successful.
Weeks later, leading U.S. AI developer Anthropic published its own report detailing similar alleged activity by three major Chinese AI labs: DeepSeek, Moonshot, and MiniMax. Anthropic claimed the three firms had generated over 16 million interactions with Anthropic’s Claude model across roughly 24,000 fraudulent accounts, a clear violation of the company’s terms of service and regional access restrictions. “Distillation can also be used for illicit purposes: competitors can use it to acquire powerful capabilities from other labs in a fraction of the time and at a fraction of the cost that it would take to develop them independently,” the report noted, adding that these attacks follow a consistent pattern: bad actors use proxy services to scale access, build networks of fake accounts to avoid detection, then send massive volumes of structured prompts to extract capabilities and build training datasets, with thousands of nearly identical prompts across coordinated accounts targeting high-value model functions.
The allegations moved to formal congressional scrutiny on April 16, when the House Select Committee on China held a hearing titled “China’s Campaign to Steal America’s AI Edge,” where lawmakers repeated accusations that Chinese firms source high-end Nvidia chips through third countries and use unauthorized distillation to extract proprietary data from U.S. AI models. “Chinese labs are resorting to unauthorized distillation attacks to extract information from our best AI models. Since they don’t have enough AI chips to develop the models on their own, they prefer to simply steal them from their American competitors. Anthropic, OpenAI, and Google have all verified that this is happening,” committee chairman John Moolenaar said, adding that Congress must pass new legislation to block what he called China’s multi-pronged effort to acquire U.S. AI technology through both legal and illegal means.
China has pushed back firmly against these allegations. In response to the White House’s claims of AI IP theft, Chinese Foreign Ministry spokesperson Guo Jiakun called the accusations completely groundless, saying they amount to a deliberate attack on China’s legitimate AI industry development. “We urge the U.S. to respect facts, discard bias, stop its containment of China’s sci-tech development, and choose the course of action conducive to sci-tech exchanges and cooperation between China and the U.S.,” Guo stated.
The escalating focus on AI IP comes as U.S. lawmakers have already ramped up restrictions on Chinese access to advanced semiconductor manufacturing technology. Earlier in April, a bipartisan group of House and Senate lawmakers introduced the Multilateral Alignment of Technology Controls on Hardware (MATCH) Act, which aims to block Chinese chipmakers from purchasing ASML’s deep-ultraviolet immersion lithography systems, critical equipment for producing advanced logic chips. Lawmakers are now expanding that regulatory push beyond semiconductor hardware to target AI model distillation.
These latest developments also come against a shifting backdrop of AI chip policy. Just one day before the OSTP memorandum, on April 22, Commerce Secretary Lutnick confirmed that no Nvidia H200 AI chips have yet been sold to Chinese companies, after Beijing urged domestic tech firms to avoid purchasing the U.S.-made chip, three months after the Trump administration approved H200 exports to China. Industry analysts note that Chinese AI firms still want access to H200 chips but remain wary of potential abrupt shifts in U.S. trade policy that could disrupt supply chains. Against this backdrop, analysts say DeepSeek V4 and Huawei’s recently launched Ascend 950PR AI chip are positioned to become the core of China’s emerging domestic AI ecosystem. Huawei launched the Ascend 950PR in late March, marketing the chip as delivering 2.87 times the performance of Nvidia’s earlier H20 chip and approaching the performance of the H200, with plans to ship roughly 750,000 units of the new chip in 2026.