Why does AI give me a different answer every time?
Why asking twice can change everything.
Word for word, the same question, the same tool, the same afternoon.
You get one answer. They get a different one. Sometimes slightly different. Sometimes the opposite.
Neither of you did anything wrong. This is built in.
Remember how it writes: one likely word at a time (see "AI isn't looking anything up").
AI doesn't always pick the most likely word. It picks from the likely ones, with a bit of chance. Usually that's what makes the writing sound natural rather than robotic.
But every small choice changes what comes next. Two runs drift apart word by word until they end up in different places.
Temperature: a setting that controls how adventurous the word choices are. Low temperature means it nearly always picks the top word. High temperature means it takes more chances.
Chat apps don't usually let you change it. Even at the lowest setting, answers can still vary a little.
In February 2026, the AI company Opper asked 53 models this:
I want to wash my car. The car wash is 50 meters away. Should I walk or drive?
You have to drive. The car needs to be at the car wash.
On a single try, only 11 of the 53 models got it right. Most said walk: "it's a short distance, saves fuel."
Of those 11 models, only 5 got it right all ten times. The others were right on some runs and wrong on others.
So "it got it right when I tried it" doesn't tell you much. The next person might get the other answer.
When a question has a clear, well-known answer, runs mostly agree.
When it's unsure, the dice matter more and the answers scatter.
So the disagreement is useful. If you ask three times and get three different answers, that's a sign it doesn't really know.
For anything important, ask more than once. Use the regenerate button, or open a fresh chat and ask again.
Treat disagreement as a warning. Different answers mean it's unsure, even if each one sounds confident.
Don't judge a tool on one try. One good or bad answer is an example, not a verdict (more in "One good AI answer doesn't prove it works").