three stories about verification losing the race against velocity. On September 8, the NSA,
CISA, and FBI jointly accused six Chinese AI companies of systematically distilling US frontier models since late 2024 -- and within two days, one of the named companies shipped a new model anyway while Beijing fired back with a mirror-image accusation of its own.
EcoGPT Goes Viral by Exaggerating What AI Actually Costs the Planet
EcoGPT, a chatbot app marketed as an "eco-friendly" alternative to mainstream AI assistants, has been downloaded more than 100,000 times this month on the strength of viral social videos claiming that AI data centers are draining the planet's fresh drinking water and that ordinary chatbot use is a hidden environmental catastrophe. Why it's taking off: The claims tap into something real -- AI's power and water footprint is a legitimate and growing concern -- while flattening it into a single terrifying number that's easy to share and hard to fact-check in a 30-second clip. A true underlying worry paired with a false, alarming headline figure is a reliable formula for virality, with or without an app attached. Worth knowing: The specific numbers behind the panic don't hold up. Google reports Gemini queries use roughly 0.26 milliliters of water each -- nowhere near the "half a litre per prompt" figure that circulated earlier and that EcoGPT's marketing leans on. The real environmental story is less dramatic per-query and more structurally serious: large data centers can consume millions of gallons of water a day, concentrated in specific drought-stressed regions rather than spread evenly across every person's individual prompt. EcoGPT's critics are calling it a greenwashing play that profits from environmental guilt using numbers that don't match the disclosed data.
1) US Spy Agencies Publicly Name Six Chinese AI Labs Over "Industrial-Scale"
On September 8, the NSA, CISA, and the FBI issued a joint cybersecurity advisory (AA26-251A) accusing six Chinese AI companies -- DeepSeek, Alibaba, Moonshot AI, MiniMax, StepFun, and Z.AI -- of running "aggressive, malicious, and targeted" distillation campaigns against American frontier models since late 2024, extracting what the agencies describe as billions of tokens across millions of exchanges. Distillation is the technique of feeding a large, expensive model's outputs into a smaller model as training data, letting the smaller model absorb much of the larger one's capability at a fraction of the cost. The advisory alleges DeepSeek specifically queried GPT-4, GPT-5, and multiple Claude versions to harvest reasoning traces, agentic behaviors, and coding optimizations used to train its R1 and V3 models -- and says the company's widely cited $5.6 million training-cost figure is misleading because it excludes the value of that harvested data. Moonshot AI is separately accused of drawing on Claude Fable 5 to train its Kimi K3 model. The agencies say the campaigns likely proceeded with the Chinese government's knowledge, and are urging US AI developers to deploy behavioral-monitoring tools and quietly degrade output for suspected accounts rather than blocking them outright.
Why it matters: For a year, DeepSeek's cheap training runs have been treated mostly as a technical parable about efficient architecture. This advisory reframes it as an intelligence story -- three of the US's own security agencies are saying, on the record, that the "efficiency miracle" runs partly on outputs quietly siphoned from the models it competes against. Whatever survives scrutiny, expect this to accelerate rate-limiting, output watermarking, and terms-of-service enforcement across every major US model API.
2) Two Days After the US Accusation, DeepSeek Ships a New Flash Model
Almost simultaneously with the advisory, DeepSeek opened a free test endpoint for V4.1 Flash, a faster successor to its V4 Flash line, with early benchmarks showing it matching Zhipu's GLM 5.3 Flash on raw speed -- up to roughly 427 tokens per second -- while claiming gains over its own V4 Pro on coding and agentic tasks; the free test window is scheduled to close today, September 10. China's government, meanwhile, has publicly rejected the NSA/CISA/FBI advisory, countering that US labs routinely distill from Chinese open-weight models like DeepSeek's and Alibaba's Qwen line, and signaling it could retaliate if Washington moves to restrict access to the named companies' models or APIs. Why it matters: Neither side slowed down to litigate the accusation first. DeepSeek shipped a new model into the middle of a US federal advisory naming it directly, and Beijing answered with a mirror-image accusation rather than a denial. Distillation claims are becoming a normal weapon in AI competition between the US and China, not an exceptional scandal -- which means the accusation itself may matter less to either country's roadmap than it does to how nervous US enterprises get about running Chinese open-weight models in their own stacks.
3) OpenAI Says Its Agents Now Out-Produce Humans 3-to-1 on Research Tasks
On September 6, OpenAI announced it had hit a goal it set last fall: fielding an "automated research intern," an agentic system that can take a bounded, well-defined research objective -- write and test a training script, run and analyze an experiment, refactor a codebase -- and carry it out mostly unsupervised over a multi-day span. OpenAI says its own research organization now gets roughly 3.1 "agent-workdays" of output for every workday of human labor. Crucially, OpenAI is careful to say this is not an autonomous scientist choosing its own research agenda: it's a supervised system that still needs a human to define the objective and evaluate the result, and the company's own data shows more than half of successful 4-to-8-hour agent tasks still require at least one human intervention along the way. OpenAI says it's targeting a fully autonomous "AI researcher" -- one that sets its own agenda -- by March 2028. Why it matters: "3.1x more research output" and "still needs a human to step in more than half the time" are both true at once, and the gap between those two facts is exactly where the real capability sits today. The useful skill for evaluating any agent claim like this isn't reading the headline multiplier -- it's asking what fraction of runs needed a human rescue, because that's the number that tells you whether you can actually remove a person from the loop or just make their job faster.
Today's stories all turn on the same problem: verification is losing the race against velocity. The NSA, CISA, and FBI are accusing six Chinese labs of a pattern that's genuinely hard to prove model-by-model, and DeepSeek answered not by contesting it but by shipping another model into the news cycle. Beijing didn't deny the accusation either -- it fired back with a mirror-image one. OpenAI's "3.1 agent-workdays" claim is real, but its own data shows the gap that headline number papers over. And EcoGPT is thriving precisely because its environmental claims are easier to feel true than to check. Across trade policy, model benchmarking, and consumer apps, the pattern is the same: whoever states a number first, loudest, and with the most confidence sets the frame, and everyone else is stuck fact-checking after the fact.
DeepSeek, Qwen, Kimi, and similar models are now explicitly named in a US federal security advisory. Even if your own usage is benign, expect procurement and compliance teams to start asking where a model's training data came from -- get ahead of that question rather than answering it after the fact.
OpenAI's 3.1x figure sounds like automation; its "human intervention needed on over half of long tasks" figure is the real constraint. When evaluating any agentic tool for your own workflow, ask what fraction of runs needed a human to step in -- that's the number that tells you whether you can actually remove a person from the loop.
EcoGPT's "AI is draining the water supply" framing spread faster than anyone could check it against real disclosed figures like Google's 0.26ml-per-query number. Skepticism of hype should cut both ways -- a scary number is just as worth verifying as an impressive one.