On April 24, the US State Department sent a diplomatic cable about AI.
The recipients were US embassies and consulates worldwide. The mission was clear: go to your host country's relevant authorities and explain that Chinese AI companies are systematically "extracting and distilling" models from US laboratories.
Three companies were named: DeepSeek, Moonshot (the company behind Kimi), and MiniMax.
What makes this different
Anthropic had previously accused these three companies of using thousands of fake accounts to scrape 16 million conversations from Claude and GPT, using the high-quality reasoning outputs as training data. That was a commercial accusation — people could choose to believe it or not.
The State Department's move elevates the same issue to a diplomatic matter.
The cable included a sharp statement:
AI models obtained through unauthorized distillation allow foreign actors to release products that appear competitive on select benchmarks at a fraction of the cost of the original system — but they cannot replicate the full original performance.
In other words: benchmark scores can fool you, but the underlying capability isn't there.
More critically, the cable added that these models "deliberately strip away safety protocols and undermine the mechanisms that keep AI models ideologically neutral and truth-seeking."
"Ideologically neutral" — coming from a State Department document, that phrase carries weight far beyond a technical accusation.
Why now
The timing is unlikely to be coincidental. DeepSeek's V4 launched on April 24 — a 1.6 trillion-parameter MoE model running on Huawei chips, priced an order of magnitude lower than GPT-5.5. On OpenRouter, V4 and Kimi K2.6 have occupied the top two spots for usage over the past two weeks.
The State Department cable is dated April 24.
China's embassy in Washington responded the same day:
Allegations that Chinese entities have stolen US AI intellectual property are unfounded and constitute a deliberate attack on the development and progress of China's AI industry.
DeepSeek previously stated that V3 used "data naturally generated and collected through web crawling," not synthetic data from OpenAI. That statement contradicts Anthropic's claim of 16 million scraped conversations, but DeepSeek has never directly responded to Anthropic.
What happens next
The cable's bottom line is a "risk warning" — alerting foreign governments to be cautious about using models developed from "stolen" US technology. This sets the stage for more concrete actions.
Possible next steps:
- Linking distillation practices to existing AI chip export controls and blacklisting the named companies
- Pushing allies to adopt the same stance, restricting procurement or use of related models
- Advancing new "AI model source transparency" rules in multilateral forums (OECD, G7)
Distillation is technically difficult to prove — how do you demonstrate that one model's outputs were trained on another model's outputs? It's a statistical and linguistic gray area. But diplomatic warfare doesn't require courtroom-level evidence — it just needs to push the narrative.
The State Department has already pushed it.
An awkward reality
The three named companies, along with Qwen and Zhipu AI, represent the entire backbone of China's open-source AI ecosystem over the past six months. If the US uses diplomatic channels to convince Europe, Japan, and Southeast Asian countries to sideline them, the impact won't be limited to a few companies — it will affect the global distribution of the entire open-source large model approach.
But open source is the most international form of distribution — the download numbers on Hugging Face speak for themselves. Llama 4 Scout reached 1.2 million downloads in its first week; DeepSeek V4 is expected to reach a similar scale. Whether it can be blocked, the market will decide.
At least in this week of April, Anthropic, the State Department, and former White House AI advisors are all pushing the same story on three parallel tracks. That rhythm doesn't look accidental.
Sources: CocoLoop, US State Dept orders global warning about alleged AI thefts by DeepSeek, other Chinese firms (Reuters/CNBC); Extraction and distillation: US State Department upgrades AI theft accusations (TechRadar)