As others have noted, there is a “temperature” parameter that defines how much randomness gets injected in its process. With a temperature of 0 it’s results would be repetitive and lack any type of creativity. With a temperature set to max it would hallucinate a ton and have wild answers.

I just tested it in the playground using gpt4 and when I used the default temperature of 1, it’s answers carried from like 303k to 305k where each answer was slightly different. Then I tested it several times with temperature of 0 and it gave the same exact result every time 303820. So the algorithm thinks that 303820 is the most correct answer, but the injected randomness makes it pick slightly less optimal answers. I don’t know if the other answers are completely out of thin air or if it’s training data just had a bunch of different sources that varied slightly.