OpenAI Accuses DeepSeek of Distilling US Models for Advantage
Narrative
DeepSeek is using distillation techniques and obfuscated methods to extract outputs from leading US frontier models (including OpenAI's) to train its next-generation systems, as part of "ongoing efforts to free-ride on the capabilities developed by OpenAI and other US frontier labs." OpenAI detected new programmatic access attempts by DeepSeek-linked accounts bypassing safeguards and using third-party routers to hide activity.
OpenAI Memo to US House Select Committee on China (reported via Bloomberg/Reuters)
Reality
Memo sent February 12, 2026; widely reported February 13–14. No public response from DeepSeek yet. Accusation focuses on preparations for next model (likely V4). DeepSeek recently expanded context window to >1M tokens (from 128k) and updated knowledge cutoff to May 2025 (from July 2024) in ongoing V3 iterations, fueling speculation on V4 readiness.
Implication
Escalates US-China AI tensions and "free-riding" debates amid export controls and chip restrictions. Highlights distillation as a growing competitive threat to US labs' moats (high R&D/compute costs vs. low-cost replication). Note: While OpenAI characterizes this as "intellectual property theft," the AI research community has a long history of using distillation for efficiency; the legal debate centers on whether DeepSeek violated Terms of Service (ToS) by using model outputs to train a competing commercial product. Could accelerate policy responses (e.g., tighter API safeguards, further GPU export limits). Adds pressure on DeepSeek ahead of anticipated V4 release (mid-Feb, coding-focused), potentially amplifying market reactions if V4 lands strong despite controversy.