When using AI, it’s imperative to know your subject matter because AI sometimes gets things wrong. The problem is, it spits out information with total confidence because it doesn’t know it’s wrong. How do we break AI of its unfounded overconfidence? Add randomness.

A new study in Nature Machine Intelligence says briefly training AI on random data before real learning leads to a better match between confidence and accuracy and, ultimately, more reliability.

Giving AI a short “warm-up” using noisy, meaningless data before real training might sound odd, but it teaches the important lesson that it doesn’t always know the answer — a simple fix to a bad habit.

After the “warm up,” the AI becomes more adept at judging its own confidence with less guessing and more honesty.

The solution is quick, low-cost and works across different AI systems.

Inspired by early brain activity, this small modification could make AI much safer in critical areas like healthcare and autonomous technology.

More like this: AI viruses join the fight against bacteria