U.S. warns allies about DeepSeek distillation
- The U.S. State Department sent embassies a global cable on April 24 directing diplomats to warn allies about alleged AI theft by DeepSeek. - The cable also named Moonshot AI and MiniMax, saying mass querying and proxy use can distill U.S. models into cheaper benchmark-matching products. - The warning expands earlier OpenAI and Anthropic complaints into diplomacy across allied capitals. (reuters.com)
The U.S. State Department told embassies worldwide on April 24 to warn foreign governments about alleged artificial-intelligence theft by Chinese firms including DeepSeek. (reuters.com) The diplomatic cable, reviewed by Reuters, said staff should raise concerns about “extraction and distillation” of U.S. AI models. It named DeepSeek, Moonshot AI and MiniMax. (reuters.com) Model distillation is a shortcut: a company asks a stronger model huge numbers of questions, collects the answers, and trains a new model to imitate them. U.S. officials say that can reproduce performance without paying the original developer’s full training cost. (anthropic.com) (semafor.com) The cable said the goal was to warn partners about the risks of using models distilled from proprietary U.S. systems and to prepare for later U.S. outreach. It said such models can match selected benchmarks while losing safety controls built into the originals. (reuters.com) (anthropic.com) This was not Washington’s first complaint. In January 2025, OpenAI said it had seen signs that groups in China were trying to use distillation to replicate advanced U.S. models, which would violate its terms of service. (semafor.com) The campaign widened again in February 2026, when Anthropic said three China-based labs — DeepSeek, Moonshot AI and MiniMax — generated more than 16 million Claude exchanges through about 24,000 fraudulent accounts. Anthropic said the activity violated its terms and regional access rules. (anthropic.com) China rejected the latest U.S. accusations. Reuters reported that the Chinese Embassy in Washington called the claims “groundless.” (reuters.com) DeepSeek did not immediately respond to Reuters’ request for comment on the cable. Around the same time, the company launched a preview of its V4 model adapted for Huawei chips, underscoring how quickly Chinese AI firms are moving despite U.S. pressure. (reuters.com 1) (reuters.com 2) The immediate fight is over copying answers at scale. The larger fight is whether the U.S. can keep frontier-model safeguards, chips and intellectual property from being turned into lower-cost rivals abroad. (reuters.com) (anthropic.com)
