Why do I get a different answer each time?

You and a friend ask a chatbot the same question, word for word. Then you compare screens. The two answers start differently, run to different lengths, and one mentions something the other leaves out. Same question in, different answers out. It can feel as if something's broken. Nothing is. It comes from the way these artificial intelligence (AI) tools write.

A chatbot writes one small piece at a time, guessing what should come next. At each step, the model scores many possible next words, from very likely to very unlikely. After "The weather today is," words like "sunny" or "cold" might score high, and "purple" very low. It's a ranking. Then it has to choose.¹

That's where the variety comes in. The model doesn't always take the top word on the list. It picks from among the likely ones, with a little chance mixed in, so a strong second choice sometimes wins. Now and then, so does a third. One different word early on changes what fits next. From there, the two answers can drift further apart with every step.¹²

People who build with these models can adjust that chance with a setting called temperature. A low temperature makes the model stick closely to the most likely words, so its answers come out more predictable. A higher temperature lets less likely words through more often. The writing gets more varied, and sometimes stranger.¹³

So why not always take the top word? Because the safest choice at every step tends to make writing flat and predictable. Lower settings suit tasks that need precision, like technical writing. Anthropic's documentation says higher settings bring more creative and varied results. A little chance makes a reply read more like something a person might write. It's the reason a list of birthday ideas can come out fresh each time you ask for one, even though the question you typed was exactly the same.¹³

Chance isn't the only reason your screen and your friend's look different.

• You may be using different versions of the model, or different apps built on top of it.
• Your settings may differ, such as a chosen style or a request for shorter replies.
• One of you may have a memory feature switched on, so the app quietly adds what it knows about you.

So two screenshots can disagree before chance even comes into it.

That variety can work for you. If you ask the same factual question twice and the two answers disagree on a fact, they can't both be right, and that fact needs checking before you rely on it. If the wording changes but the facts hold steady, that's a good sign, though it's still no proof. Two answers, one quick check. Next time an answer matters, try asking a second time in a fresh chat, and compare the two.