OpenAI has accused Chinese AI startup DeepSeek of using “distillation” to extract outputs from leading US artificial intelligence models and use those results to train its own systems, according to a memo sent to US lawmakers.
The claims center on what OpenAI described as increasingly sophisticated attempts to evade safeguards and access restrictions, including using third-party routing methods that mask where requests are coming from.
OpenAI’s memo was sent on Feb 12 to the US House Select Committee focused on strategic competition between the United States and the Chinese Communist Party, and it was reviewed by Bloomberg News and seen by Reuters, according to published reports.
DeepSeek did not immediately respond to requests for comment in the reports, and OpenAI declined to comment on the memo, according to the same coverage.
What OpenAI alleges
In the memo, OpenAI said DeepSeek has used distillation techniques as part of “ongoing efforts to free ride on the capabilities developed by OpenAI and other US frontier labs.”
The reports describe distillation as a technique where one AI model relies on another model’s outputs to help train and improve a newer system, effectively transferring learnings and producing similar capabilities.
OpenAI said it detected “new, obfuscated methods” intended to evade its defenses against misuse of model outputs, as described in the Bloomberg-reviewed memo.
OpenAI also said an internal review suggested that accounts associated with DeepSeek employees sought to bypass guardrails by accessing models through “obfuscated third-party routers” and other methods that hide their source.
The memo further alleged that DeepSeek employees developed code to access US AI models and obtain outputs for distillation in “programmatic ways,” according to the reports.
Business and safety concerns
OpenAI argued that distillation has continued and become more sophisticated despite efforts to crack down on users who violate its terms of service, based on activity it said it observed on its platform.
The company also warned that when model capabilities are copied through distillation, safeguards can be lost or weakened, increasing the risk of misuse in high-risk areas such as biology or chemistry.
The memo framed the practice as a potential business threat, pointing to how DeepSeek and many other Chinese models do not have a monthly subscription cost, while US companies such as OpenAI and Anthropic charge fees for premium services and have invested billions in AI infrastructure.
OpenAI said this imbalance could erode a US advantage in AI, according to the same memo.
DeepSeek’s profile and market impact
DeepSeek is based in Hangzhou and is backed by parent company High-Flyer, according to the reporting.
The company drew attention early last year when it released a set of AI models that were described as rivaling some top US offerings, fueling concerns in Washington about China closing the gap in AI despite restrictions.
The coverage noted that Silicon Valley executives have previously praised DeepSeek models named DeepSeek-V3 and DeepSeek-R1, and that these models are available globally.
Bloomberg’s account also said OpenAI began raising concerns privately not long after the R1 model’s release in 2025, including opening a probe with partner Microsoft into whether DeepSeek obtained data in an unauthorized manner.
Wider US policy spotlight
OpenAI’s warning came as Washington has also focused on advanced AI chips and whether chip access could speed DeepSeek’s progress, according to the Bloomberg-reviewed memo.
The report said that at the end of 2025, US President Donald Trump moved to ease chip restraints and allow Nvidia to sell H200 processors, chips described as about 18 months behind Nvidia’s Blackwell versions.
Bloomberg’s account also said US authorities opened a probe shortly after the R1 release into whether DeepSeek bypassed US export controls by buying chips via Singapore.
In addition, records obtained by the House China committee were described as showing Nvidia provided technical support to help DeepSeek improve and co-design its R1 model.
Those records were also described as stating that DeepSeek’s V3 base model required 2.8 million H800 GPU hours for full training, and that the H800 processors were permitted for sale to China for a few months in 2023 before rules later that year halted sales.
