AI Just Caught a Hidden Flaw in Engineering Math

Engineers rely on math to guarantee that complicated machines won't spin out of control. Planes, drones, chemical plants, and self-driving cars all have one thing in common: they react to information that's a little bit old. A sensor reads the temperature, and by the time the computer responds, a fraction of a second has passed. Those small delays can add up and cause wild, dangerous swings. So for decades, engineers have used a mathematical checklist to prove a system will stay calm even with delays.

The checklist works like this: if you can write down a certain kind of formula that always shrinks over time, you're allowed to declare the system safe. It's like saying a ball rolling in a bowl will always slow down, so it can't escape. That shortcut has been treated as reliable. Vittorio De Iuliis and Pierdomenico Pepe now present a single case where the checklist passes but the system still misbehaves — a counterexample. If it holds up, the shortcut isn't always valid.

What makes this unusual is how they found it. The authors say the counterexample came out of back-and-forth conversations with large language models — the same AI chatbots anyone can use. That's notable because it shows these tools doing genuine, original mathematical work rather than just summarising what's already published. It's a hint that AI is becoming a research assistant, not just a search engine.

One honest caveat: the paper is labelled a 'candidate' counterexample, posted online before formal peer review. Other mathematicians need to check the reasoning before anyone rewrites textbooks. And this is deep, technical work with no immediate product or price change. Still, it matters because the math underneath everyday technology should be right — and because AI just showed it can help find where it isn't.