Ask a chatbot for the birthday of someone obscure and it may give you a date, with total confidence, even though it cannot know. The authors of this paper tried exactly that. A leading open model answered three times with three different wrong dates, even though the prompt said to answer only if it knew.
People call this hallucination. It can feel like a mysterious flaw. The paper argues it is not mysterious at all. It has two plain causes, and one of them is how we grade the models.
Hallucinations come from two things. First, facts a model saw only once are statistically hard to tell from plausible falsehoods. Second, most tests give a point for a right answer and nothing for "I don't know", so guessing always scores at least as well as admitting doubt. Models trained and ranked on those tests learn to bluff.
Cause one: facts the model barely saw
During pretraining a model reads huge amounts of text. Some facts show up thousands of times and are learned well. Others, such as one person's birthday, appear in the data exactly once, or never.
The paper shows that a model cannot reliably separate such a rare fact from a convincing wrong one. Mathematically it turns text generation into a yes or no question: "Is this a valid answer?" The authors prove that the rate of generated errors is at least twice the rate at which a classifier gets that question wrong.
There is a simple rule of thumb. If 20 percent of the birthday facts appeared exactly once in training, expect the model to hallucinate on at least 20 percent of such birthdays. The paper calls the share of once-seen facts the singleton rate.
Imagine memorizing a phone book by reading each page one time. You will know the famous numbers you saw many times. For a name you saw once, you can produce a believable number, but you cannot check it. Without a way to check, a believable number and a true number feel the same.
