OpenAI at Congress: "DeepSeek continues to ‘distill’ patterns"

OpenAI sent to House Select Committee of the House of Representatives, in the American Congress, a document on strategic competition Between United States e China. The company reports risks related to “model distillation” and points energy and infrastructure as key factors in the global race forartificial intelligence.

The company led by Sam Altman writes that he has observed “indicative” activities on his platform attributable to continuous attempts to distill OpenAI models and other US laboratories. In AI jargon, “distill a model” means using a larger and more powerful system such as “teacher” to train a smaller one, the “student”, making it imitate the behavior of the first.

The ‘masked’ accesses

In practice, many responses, examples or signals produced by the reference model are collected (even in an automated way) and used as training data to obtain a lighter and less expensive model that reproduces part of the capabilities of the original model. The technique is common and legitimate but becomes controversial if the output is pulled at scale from proprietary services to replicate capabilities without permission.

OpenAI claims to have identified accounts associated with employees of DeepSeek engage in evading access restrictions, including through “obfuscated third-party routers” and other methods to mask the origin of requests. The company adds that it believes that employees DeepSeek have developed code to access models and obtain output programmatically, with the goal of using them for distillation.

The Chinese challenge and ‘democratic AI’

Second OpenAImost of the observed adversary activity of this type appears to come from the China and, “occasionally”, from Russia. While citing the membership of DeepSeek to Chinese voluntary commitments on AI safety, OpenAI writes that there is no published clear picture with robust evidence from independent tests and evaluations. The document also dedicates a passage to sensitive content. OpenAI states that DeepSeek applies restrictions on topics considered politically sensitive in China and reports examples in which generated responses were later removed. The note frames the challenge onartificial intelligence as the advancement of a “democratic AI” and claims that the Chinese Communist Party aims to become a global leader by 2030. In this context it defines the release of the model DeepSeek R1which occurred on the Lunar New Year a year ago, a useful indicator for measuring the state of competition. The company also recalls that in March 2025, at the request of the Commission itself, it had already sent an evaluation of the distillation Of DeepSeek and explains that this note, dated February 12, 2026, updates that analysis to include a reconstruction of evolving tactics in view of a new model believed to be more powerful expected for the next Lunar New Year, along with other elements on China’s progress towards the 2030 goal and the measures taken by OpenAI to strengthen the protection of your models.

Infrastructure and energy capacity

On the front infrastructural, OpenAI cites data attributed to Bloomberg according to which in 2025 the China would have added 543 gigawatts again electrical capacitycompared to 51 gigawatts in the United Statesand in 2024 another 429 gigawatts. It then adds internal numbers on its own computing ability powered by energy: 0.2 gigawatts in 2023, 0.6 in 2024 and around 1.9 in 2025 and writes that in two years the capacity has grown 9.5 times while still remaining low compared to demand. This is one of the reasons why the release of new features slows down. Finally, the memo mentions the Stargate projectwith the aim of bringing AI infrastructure to the United States to 10 GW by 2029, indicating that after a year the path would be “over halfway”.

Countermeasures and ecosystem security

As for countermeasures, OpenAI describes a strengthening of defenses against distillation: removal of suspicious users, detection and review systems, and training to reduce the leakage of “reasoning traces”. The company calls for an approach of “ecosystem security” to limit the movement of abuse towards less protected suppliers and says he is available for a briefing with the Commission.