AI Express // Latest edition // nr. 01 // Sep 13The Synthesis - Sep 13: Who verifies the output?
Sep 13free stories unlockedfull issue for subscribers
Welcome to The Synthesis, the AI Express Sunday read on what the week meant. The running theme was trust: Washington named six Chinese companies it says distilled frontier models at industrial scale, a court fined a lawyer for filing an unverified AI brief, and Anthropic published nine months of blocked misuse. Below, the week in one sentence, the big story, the fine print, the security file, one bet, the deals, and a signal of hope.
The week in one sentence
The Big Story
PolicyIn a rare joint advisory, the three agencies accused DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI of extracting, since late 2024, billions of tokens from US frontier models (variants of Claude, GPT, Gemini and Grok), 'likely with Chinese government awareness.' Distillation is, they say, not a supplement but the core of these companies' strategy, and publicly quoted training costs hide the true price of the extracted data. Their suggested defenses are unusual: US providers should degrade responses to suspected campaigns, quietly move accounts to weaker models and share threat signals. China's foreign ministry rejected the claims. The advisory reframes the open-weight race: the cheapest path to a frontier model may be more queries, not more compute.
The Fine Print
AccountabilityThe New Mexico Supreme Court fined Stephen Aarons and held him in contempt after an appeal brief he signed and filed was built on ChatGPT fabrications: wholly invented witnesses (Officer Michelle Amarillo, Officer Sanchez, Manal Al-Jibury, Teresa Marquez) and false testimony about the shooter's clothing and appearance. Aarons said he expected a 'bulletproof summary.' The court found he never verified the factual claims, removed him from the murder appeal and referred him to the disciplinary board. Courts are becoming the enforcement mechanism for AI accountability, and the fine is the new cost of skipping verification.
For subscribers only
Biosecurity
Anthropic's nine-month log of blocked misuse, from grant applications to a dangerous virus
lockedPhishing
The new bait is trust in AI brands
lockedForecast
Distillation becomes a product problem before it becomes a policy one
lockedScience
Antibiotics: one lab's hunt gets a decade faster
lockedThe full brief, the tools, the GitHub radar and the context stay subscriber-only. Subscribers always read first, in full.