OpenAI Accuses DeepSeek of "Free-Riding" on US AI Innovation Through Model Distillation

San Francisco, CA — In a significant escalation of the technological rivalry between the United States and China, OpenAI has formally warned US lawmakers that Chinese AI startup DeepSeek is systematically utilizing "distillation" techniques to replicate the capabilities of proprietary US artificial intelligence models.

According to a memo sent to the House Select Committee on the Strategic Competition Between the United States and the Chinese Communist Party, and subsequently reported by Bloomberg and Reuters on February 12, 2026, OpenAI alleges that DeepSeek is employing "sophisticated" and "obfuscated" methods to extract data from OpenAI’s servers. This data is then allegedly used to train DeepSeek’s own models, including the recently popularized DeepSeek-R1, effectively allowing the Chinese firm to bypass the immense research and development costs incurred by American laboratories.

This development marks a pivotal moment in the global AI landscape, shifting the focus from hardware export controls to the intangible—yet highly valuable—flow of model weights and algorithmic logic.

The Mechanics of "Model Distillation"

At the heart of the controversy is a technique known in machine learning as "knowledge distillation." While the term sounds abstract, the process represents a tangible threat to the competitive moat of leading AI labs.

In a standard training scenario, an AI model learns from raw datasets—trillions of tokens of text, code, and images. This process requires massive computational power and months of processing time. Distillation, however, shortcuts this process. A "teacher" model (in this case, presumably OpenAI’s GPT-4 or o1 series) is queried extensively. The "student" model (DeepSeek’s architecture) learns not just from the correct answers, but from the probability distributions and reasoning traces provided by the teacher.

OpenAI’s memo contends that DeepSeek is not merely using public outputs but is actively circumventing safeguards to harvest these high-quality training signals at scale. By doing so, DeepSeek can purportedly achieve near-parity performance with a fraction of the compute resources and financial investment required by its US counterparts.

Comparisons of AI Training Methodologies

To understand the economic and technical disparity cited by OpenAI, it is essential to compare the two primary approaches to model development.

Table 1: Native Training vs. Model Distillation

Evidence of "Obfuscated" Extraction

The allegations go beyond simple usage violations. OpenAI claims to have detected specific, adversarial patterns of behavior linked to DeepSeek employees. The memo outlines how these actors allegedly utilized disguised third-party networks to mask the origin of their queries, thereby evading OpenAI’s geographic and volume-based blocks.

"We have observed accounts associated with DeepSeek employees using methods to circumvent access restrictions," the memo states. OpenAI characterizes this activity as an attempt to "free-ride" on the technological breakthroughs of US labs. The implication is that DeepSeek’s vaunted efficiency—often cited as an engineering marvel—may be partly attributed to this unauthorized transfer of intelligence rather than solely architectural innovation.

National Security and Safety Guardrails

Beyond the commercial implications, OpenAI raised a red flag regarding national security. The company warned lawmakers that when capabilities are copied via distillation, the safety alignment and ethical guardrails built into the original model are often lost or discarded.

DeepSeek’s models are known to comply with strict Chinese internet regulations, censoring topics such as the status of Taiwan or the 1989 Tiananmen Square protests. However, OpenAI argues that the danger lies in what is not filtered: the raw capability to generate cyber exploits or design biological agents.

"When capabilities are copied through distillation, safeguards often fall to the wayside," OpenAI noted. This creates a scenario where a distilled model possesses the dangerous capabilities of a frontier US model but lacks the "refusal" mechanisms designed to prevent misuse in high-risk domains like biology or chemistry.

The Economic Threat to Western Labs

The rise of DeepSeek has already sent shockwaves through the stock market, impacting the valuations of US chipmakers and AI firms alike. By offering high-performance models for free or at significantly lower API costs, DeepSeek challenges the business model of companies like OpenAI, Anthropic, and Google, which rely on subscription revenues to fund their multi-billion dollar infrastructure projects.

If distillation becomes a normalized route for competitors to catch up, the incentive for private capital to fund expensive "frontier" research could diminish. OpenAI’s appeal to Congress suggests they view this not just as a terms-of-service violation, but as a systemic threat to the US innovation ecosystem that requires legislative or regulatory intervention.

Industry Reaction and Future Outlook

The accusations have sparked a fierce debate within the technical community. Proponents of open-source AI argue that analyzing model outputs is a standard practice and that "learning from the best" is a fundamental driver of scientific progress. However, critics point out that automated, large-scale extraction violates the contractual terms of service of almost all commercial AI providers.

DeepSeek has not yet issued a detailed public rebuttal to these specific claims, though the company has previously attributed its success to efficient coding and novel architecture specifically designed for inference optimization.

As the US House Select Committee reviews these allegations, the industry anticipates potential policy shifts. These could range from stricter "Know Your Customer" (KYC) requirements for AI API access to new trade restrictions aimed at preventing the digital export of model weights and reasoning traces.

For Creati.ai, this unfolding story underscores the critical importance of intellectual property protection in the age of generative AI. As models become more capable, the line between inspiration and theft is becoming the new frontline of global technological competition.