Error Rate

Error rate refers to the proportion of an AI's answers that turn out to be wrong.

Think of making an "error notebook" after a school exam. If you solve 100 problems and get 3 wrong, your error rate is 3 percent. But in AI services, this number doesn't always carry the same weight. Getting 3 percent wrong in casual small talk is a completely different kind of accident than getting 3 percent wrong about facility locations or administrative boundaries on a map. That's because in the latter case, users take the answer at face value and act on it in the real world.

So even the same error rate can call for different acceptable thresholds depending on the screen or situation it appears in. A chatbot meant for casual conversation can tolerate a few percentage points of error, but in domains like healthcare, law, or disaster information—where users act on the answer immediately—much stricter standards are needed. In other words, what matters for safety isn't the number itself, but where that number shows up.

Discussing the challenge of layering generative features onto maps, the article notes that a 3 percent error rate can be fine on one screen and a disaster on another. While people often assume a lower error rate is always safer, the reality is that the standard depends on where that answer gets used.

Try asking a chatbot 10 questions in a row for which you already know the correct answers—for example, asking about the capital cities of various countries one by one—then compare its answers against the facts and count how many it got wrong. Divide the number of wrong answers by 10 to get a rough sense of that chatbot's error rate. Repeat with different difficulty levels or topics to see in which areas the error rate tends to rise.