A hotline for rogue military AI, before Trump meets Xi
Two governments that agree on almost nothing are being handed a narrow set of rules for the machines both are wiring into their militaries — and the proposal lands days before their presidents sit down in Washington.
US and Chinese security experts are proposing nuclear-style red lines around military AI, plus a dedicated hotline for incidents involving autonomous systems, according to recommendations published last week ahead of planned government-level AI talks and an expected September 24 meeting in Washington between Donald Trump and Xi Jinping. The authors are Melanie Sisson of the Brookings Institution and Tianjiao Jiang of Fudan University, both participants in the US-China AI and national security dialogue that Brookings and Tsinghua University's Center for International Security and Strategy have convened since 2019. Sisson argues humans should keep sole authority to launch AI-enabled cyberattacks against another country's nuclear command, control and communications systems or other strategic infrastructure; Jiang wants explicit bans on AI independently deciding to use nuclear weapons or autonomously attacking nuclear command systems, protections for energy, finance and healthcare, and a shared definition of "meaningful human control" so both capitals cannot use identical language to justify different levels of machine autonomy.
The hotline is the tell. A defensive AI that answers suspicious network traffic on its own could be read as an attack by the other side before any human establishes what happened, which is why Jiang wants a direct channel to say an operation was accidental, unauthorized or still under investigation. The idea builds on the Biden-Xi principle from November 2024 that humans, not machines, control nuclear decisions — and it walks straight into the doubts Carla Freeman of Johns Hopkins has raised about the existing military crisis line, which China did not answer during the February 2023 spy balloon incident. Neither government has endorsed the proposals. China has folded AI into the Foreign Ministry's arms-control department; Washington has no single arms-control home for it at all. We covered the run-up to these talks earlier — US, China schedule first dedicated AI safety talks of Trump's second term.
Nine voice models read the same 274 drug names in clinical sentences, and the best general-purpose systems mispronounce up to one in three newly approved names. The DOSE benchmark published by Synthio Labs scores each pronunciation from 0 to 5 against verified reference readings, counting 4 or higher as a pass; the company's own RxPronounce led at 91.2% overall and 87.0% on new names, 10.9 points clear of Cartesia's sonic-3.6 at 80.3%. Every general model fell on newly approved names — ElevenLabs' eleven_v3 dropped from 93.0% to 67.1%, Google's Gemini TTS from 89.1% to 61.6%, Azure below half of generic names — and xAI's Grok TTS read Xofluza out letter by letter. Two caveats worth keeping in view: the vendor's model is the winner, and scoring is automated audio comparison with no human reviewers in the loop, though the full audio set is on Hugging Face for anyone who wants to hear it.
Amazon took warrants for up to $340 million of Generac stock in a data-center backup-power deal that the backup provider says could be worth as much as $8 billion, sending Generac shares up more than 40%. Generac issued warrants for up to 1.69 million shares at $200.93 each, roughly 3% of outstanding stock, with about 308,000 vesting immediately and the rest tied to payments; the filing puts initial generator deliveries at $2.4 billion across 2027 and 2028. It is Amazon's second supplier stake in as many weeks, after warrants for up to $4 billion of Qualcomm, and it says where the bottleneck lives now: the buildout is being financed through the power and silicon underneath it, not just the chips.
What to watch: whether the AI hotline reaches the September 24 agenda in any form, and whether either capital signs onto a definition of human control it can live with.
Would you trust a hotline to de-escalate a machine-speed incident, or is the phone the weakest link in the chain? Tell us in the comments.
