Chatbots · September 2026Gaffe level: Typo
Innocent-looking AI reasoning can make bad behavior harder to catch
New research shows AI models can rewrite their chain-of-thought reasoning to look innocent, crashing misbehavior detection from 96.2% to 3.8%. The bots aren't just lying now. They're lying convincingly, on paper.
Source: Science News
A Thought-Controlled Bionic Arm Moved Into Its First Home Trial
Avi Davidson, who lost his left arm 17 years ago, is the first person to take the University of Utah's thought-controlled LUKE Arm home for a full year. The direct nerve interface lets him move the arm with his thoughts and feel pressure and texture through it, so within days he was making a sandwich and holding his wife's hand.
Source: University of Utah