OpenAI Accuses DeepSeek of Siphoning Model Outputs
OpenAI has formally accused its Chinese rival DeepSeek of covertly siphoning outputs from its models to train their own. The allegations, which highlight growing concerns over data leakage and model security, were submitted to the U.S. House Select Committee on China.
- The core allegation involves "distillation," a technique where a model is trained on the outputs of another, more capable model. OpenAI claims DeepSeek developed code to programmatically access US AI models and harvest their outputs for this purpose. - To mask their activities, DeepSeek employees allegedly used obfuscated third-party routers and other methods to circumvent access restrictions put in place by OpenAI. OpenAI stated it observed accounts associated with DeepSeek employees developing these circumvention methods. - This issue extends beyond a business dispute, with OpenAI raising national security concerns. The company warned that safety guardrails embedded in models may not transfer during distillation, potentially enabling misuse in sensitive fields like biology or chemistry. - OpenAI's memo also highlighted that DeepSeek's models appear to censor topics sensitive to the Chinese government, such as Taiwan and Tiananmen Square, and exhibit a pro-CCP bias. - The practice of distillation poses a significant business threat to companies like OpenAI and Anthropic, which invest heavily in infrastructure and charge for premium services, as competitors could "free-ride" on their capabilities to offer similar, often free, services. - This accusation is part of a larger pattern of concern in Washington regarding China's rapid AI advancements. US authorities had previously investigated whether DeepSeek bypassed export controls to acquire processors, with documents indicating that Nvidia provided technical support for the development of DeepSeek's R1 model. - The controversy also brings to light the broader issue of "model collapse," where models trained on synthetic data from other AIs can degrade in quality over time, leading to less diverse and potentially nonsensical outputs.
