The most dangerous AI isn't the one that becomes too powerful, but the one we trust too much

Artificial intelligence doesn’t need to reach superintelligence to pose a global threat. It only needs to make a mistake—and for humans to act on it.

According to an article in *The Conversation*, current AI systems, like large language models, can already influence critical decisions without being conscious or hostile. In Ukraine, over 70 computer vision systems are used to target military objectives, reducing human involvement to mere validation. In the U.S., an analyst used a chatbot in 2026 to assess intelligence on a Chinese ship, incorrectly identifying components of a nuclear weapons program. The error nearly triggered a military operation before being detected.

This risk stems from over-reliance on automated systems. Models produce fluent, convincing responses, even when wrong, lending unjustified authority to errors. Policymakers often claim humans remain in the loop, but automation bias skews judgments, as shown in 2003 when Patriot missiles shot down allied aircraft. The solution requires systematic verification of AI outputs, clear labeling, and challenging results before any action is taken.