OpenAI alleges Chinese rival DeepSeek covertly siphoned outputs to train R1

OpenAI has informed the House Select Committee on China that DeepSeek, a Chinese startup, appears to have systematically harvested responses from top U.S. AI systems to fast-track its R1 chatbot. The memo alleges the activity blended model-distillation approaches with masked, time-distributed and evasive querying designed to sidestep conventional rate limits, filters and abuse‑detection heuristics. OpenAI characterizes the behavior as functionally equivalent to extracting another lab’s capabilities and folding those outputs into a rival training pipeline, a shortcut that can compress months of research into automated data-collection and fine‑tuning workflows. The complaint echoes other industry disclosures — notably Google’s public warnings about coordinated, high-volume querying to reproduce Gemini-style capabilities — suggesting the tactic is not isolated. Independent testing and academic work also show how deployment choices (persistent memory, large context windows and agentic toolchains) and product incentives can amplify tailored behavior and make systematic extraction more attractive commercially. Technically, defenders face a suite of hard problems: distinguishing legitimate research queries from adversarial harvesting, detecting low‑volume, long‑running probing campaigns, and making generated outputs less useful for downstream training without degrading utility. The memorandum frames model outputs as contested intellectual property and highlights thorny legal questions, since past disputes over dataset sourcing complicate straightforward claims of ownership or misappropriation. Policy responses OpenAI and others are likely to press include stronger telemetry and access governance, robust watermarking or provenance attestation for generated content, and export‑style controls on advanced model access — measures that must be balanced against research, interoperability and competition concerns. Industry practitioners say operational fixes will range from enhanced rate‑limiting and API controls to cross‑lab cooperation on detection signals and contractual enforcement against misuse. For U.S. policymakers, the episode dovetails with broader technology‑security debates about supply chains, talent flows and international partnerships, and it increases pressure for clearer disclosure rules and verification standards for model access. Whether the memo yields formal enforcement or new legislation, it intensifies scrutiny of how frontier model behavior is commodified and protected in a fast‑moving global market.