OpenAI raises concerns over Deepseek’s AI training practices - alleges IP theft
- Marijan Hassan - Tech Journalist
- 6 minutes ago
- 2 min read
Formal complaint to Congress alleges Chinese startup "Free-Rides" on American R&D.
The AI "Cold War" has reached a boiling point. Last week, OpenAI sent a high-stakes memorandum to the House Select Committee on China, formally accusing the Chinese AI startup DeepSeek of systematic intellectual property theft.
The charge? Using a technique called "distillation" to siphon the intelligence of American frontier models, including GPT-4 and o1, to build its own rival systems at a fraction of the cost.
The memo marks the first time OpenAI has explicitly asked U.S. lawmakers to intervene in what it calls "sophisticated efforts to free-ride" on American innovation.
What is "distillation" in this context?
In AI research, distillation is a legitimate technique where a smaller "student" model learns from the outputs of a larger "teacher" model. However, OpenAI argues DeepSeek’s application of it is predatory and violates its terms of service.
By querying OpenAI’s models with millions of complex prompts and using the high-quality answers as training data, DeepSeek can effectively "clone" the reasoning capabilities of GPT-4 without spending billions on the original R&D or trial-and-error phases.
The "evidence"
OpenAI claims its internal forensic data shows statistically impossible similarities in reasoning structures and error patterns between DeepSeek-R1 and its own proprietary models.
OpenAI’s memo describes a persistent, cat-and-mouse game between its security teams and DeepSeek employees. The US AI pioneer reports detecting accounts linked to DeepSeek employees using masked third-party routers and unauthorized resellers to hide their Chinese origin while pinging OpenAI servers.
Despite OpenAI’s "anti-distillation" guardrails, the company alleges DeepSeek developed automated scripts specifically designed to programmatically extract model knowledge in ways that mimic normal user behavior to avoid detection.
Safety erosion
A major point of concern for lawmakers is that when a model is "distilled," the original safety safeguards often don't transfer. This means DeepSeek may have captured the intelligence of U.S. models while stripping away the restrictions that prevent the AI from helping create biological weapons or malware.
Geopolitical and economic fallout
The timing of the memo is critical. DeepSeek shocked the industry in 2025 by releasing models that rivaled GPT-4 while costing 80% less to train.
Because DeepSeek offers its models for free or at a near-zero cost, OpenAI warns that "unregulated distillation" could bankrupt Western AI firms that have invested tens of billions in physical infrastructure.
Representative John Moolenaar, chair of the House China committee, echoed OpenAI’s frustration, stating, "This is the CCP playbook: steal, copy, then dominate."
The memo highlights that software-level "piracy" is effectively neutralizing U.S. hardware sanctions. Even if China can't get the latest Nvidia chips, they can simply "query the brains" of American models running on those chips in U.S. data centers.
DeepSeek’s defense (and Silence)
While DeepSeek and its parent company, High-Flyer, have not formally responded to the memo, the startup has previously argued that its efficiency comes from algorithmic breakthroughs and better utilization of "older" chips like the H800. Proponents of open-source AI have also accused OpenAI of trying to "pull up the ladder" behind them to maintain a monopoly on high-end intelligence.
