AI Chatbot's False Alarm Nearly Sparked US-China Conflict

Key Takeaways

• An AI chatbot hallucination about a Chinese ship nearly triggered a US military attack, with troops and warplanes deployed before the error was caught
• AI referral traffic to retail pages grew 304 percent in 2025, far exceeding the 40 percent growth from other sources
• Google, Google DeepMind, and MIT report that AI speeds up research but not lab experiments due to verification burdens and real-world complexity
• OpenAI is reportedly pursuing the Hodge Conjecture while facing criticism over prioritizing announcements over rigorous verification
• PrismML's Bonsai 2 27B model runs on consumer hardware with Nvidia GeForce GTX 5090 or Apple M5 Max chips and uses 40% less energy than other 8B models
• Palantir CEO Alex Karp argues frontier AI labs push regulation to avoid legal liability
• Congress has introduced multiple AI regulation measures but progress is slow, with action possibly delayed until after the midterms
• 2028 presidential candidates including Elizabeth Warren and Bernie Sanders are debating AI policy, covering national strategy, job loss, and bias
• A study of 14,767 LLM benchmark papers from 2022 to 2026 found AI models are increasingly involved in creating and scoring tests, raising bias concerns
• South Korea's AI startup competition awarded Stillcut for deepfake detection, with Quantum Hightech and Motifdrive also receiving recognition

AI chatbot error nearly sparked US war with China

A US military analyst used an AI chatbot to study a cargo ship. The chatbot made up a false claim about nuclear smuggling. Armed troops and warplanes were nearly sent to intercept the Chinese vessel before officials discovered the AI-generated report was entirely false. The incident almost started a war and raised concerns about how the Pentagon uses AI.

Military AI use nearly caused diplomatic crisis with China

The Pentagon's experimental use of AI nearly caused a major diplomatic incident with China. An analyst asked a chatbot to analyze a ship's manifest, and the AI mixed public and secret data to reach a wrong conclusion. Military planes were already in the air before officials caught the error. The Pentagon is still working on rules for AI use in military operations.

AI hallucination almost triggered US military attack

An AI chatbot created false intelligence about a Chinese ship carrying nuclear weapons parts. A Special Operations Command analyst used the chatbot to combine open source and classified data. The chatbot misidentified the cargo, and the analyst then formatted the wrong findings into an official report. The operation was stopped at the last moment, highlighting serious risks of AI in military planning.

AI referral traffic to retail pages surged in 2025

Traffic from generative AI platforms to retail product pages grew 304 percent in 2025. This growth far outpaced the 40 percent increase from all other sources. The shift shows that AI tools are becoming a major way shoppers discover products. Retailers now need to adapt their strategies to keep up with this change.

OpenAI reportedly pursuing Hodge Conjecture amid math feud

OpenAI is reportedly trying to solve the Hodge Conjecture while facing criticism from mathematicians. Critics say AI firms use mathematicians' work to quickly solve unsolved problems without truly advancing understanding. OpenAI is said to care more about announcing a solution than doing careful verification. The company is trying to find a way to announce results without angering the math community again.

Study maps how LLM benchmarks are designed and evolving

Researchers analyzed 14,767 papers about LLM evaluation from 2022 to 2026. They found that benchmarks increasingly focus on action, interaction, and professional use. The study also notes that AI models are increasingly involved in creating and scoring tests, raising questions about independence and bias.

2028 presidential candidates debate AI policy and regulation

Democratic presidential hopefuls are making AI a major campaign issue. Some, like Elizabeth Warren, call for a national AI strategy. Others, like Bernie Sanders, focus on risks like job loss and bias. The debate reflects growing public concern about AI's impact on the economy and society.

Korea hosts competition to find promising AI startups

South Korea's Ministry of Science and ICT held a final competition for AI startups in Seoul. Stillcut won for deepfake detection technology. Quantum Hightech and Motifdrive also received awards. The finalists received prize money and investment opportunities.

Palantir CEO says AI safety talk is about avoiding liability

Palantir CEO Alex Karp said frontier AI labs push regulation to avoid legal responsibility for harm. He argues companies should first be accountable for their technology. Karp also warned that regulating AI is difficult because experts are often employed by the industry itself.

Congress proposes AI regulation but quick action is unlikely

Congress has introduced multiple AI regulation measures, but progress is slow. The House and Senate are both working on bills. Lawmakers disagree on how strict regulations should be. Rapid AI development makes it hard for Congress to keep up, and action may wait until after the midterms.

AI speeds up research but not lab experiments

A new report from Google, Google DeepMind and M.I.T. finds that AI is helping scientists in some fields but not in lab work. Researchers spend a lot of time checking AI outputs instead of trusting them. Lab experiments are harder for AI because they involve real-world materials, safety checks and hard-to-find training data. Clinical trials also move slowly due to regulatory reviews that AI cannot speed up.

PrismML launches tiny Bonsai 2 27B AI model

PrismML launched Bonsai 2 27B, a high-intelligence AI model small enough to run on consumer hardware. The model uses ternary compression with three-bit weights to keep accuracy while shrinking memory use. On benchmarks it performed close to Qwen3.8 in agentic tasks, coding and reasoning. It can run on an Nvidia GeForce GTX 5090 or Apple M5 Max chip and uses 40% less energy than other 8B models at full precision.

Sources

• US military chaos as AI chatbot 'almost started a war' inventing nuclear smuggling threat
• The Military's AI 'Almost Started a War' With China: Report
• AI Hallucination Nearly Triggers US Military Operation
• AI referral traffic quadrupled in 2025 as product discovery shifts
• OpenAI Reportedly Trying to Solve Hodge Conjecture Amid Feud With Math Community
• What Do We Expect from LLMs? Mapping the Design of LLM Benchmarks
• Newsom's big AI swings put 2028 in focus
• [Science Bulletin] MSIT Hosts Competition to Discover AI Startup Investment Targets
• Palantir’s Alex Karp Says Frontier Labs’ AI Safety Stance Is To Get Out Of Liability Of The Harm Their Products Might Cause
• Congress floats raft of measures to regulate AI, but quick action unlikely
• Why AI is speeding up scientific research but not lab experiments
• PrismML launches Bonsai 2 27B, a high-intelligence AI model so small it fits on consumer hardware