You are lost in an unfamiliar city. You ask two passers-by for directions.
The first answers with confidence, without the slightest hesitation: left, then straight on. He is completely wrong.
The second frowns: "I'm not sure. Ask at the bakery, they should know."
Which of the two really helped you?
Two promises that contradict each other?
This week, Google presented Gemini 4 Argon with a promise: a model that answers more directly, with fewer rhetorical precautions. And an independent test found a rare quality in it: it is the large model that makes up the fewest answers when it does not know.
At first glance, these two qualities pull in opposite directions. Answering more directly means hesitating less. Making things up less means knowing when to stop when you do not know. How do you do both at once?
Why AI struggles to doubt
An AI is trained to produce an answer. Faced with a question, its reflex is to complete it, the way you finish a sentence. Saying "I don't know" does not come naturally to it.
Worse: for a long time, the very way they were evaluated pushed them to guess. Researchers showed this in 2025: as in a multiple-choice exam, answering at random sometimes earned points, whereas admitting ignorance never earned any. So machines were taught, unintentionally, to bluff rather than to doubt.
Doubt is not hesitation
We need to distinguish between two things that are often confused.
Hesitation is unnecessary precautions: the "it is important to note that", the "it depends", the warnings in every sentence. It is a way of covering yourself. It brings the reader nothing.
Doubt is something else: knowing where the limits of what you know lie. Answering frankly when you are sure, and saying so clearly when you are not. That kind of doubt is not a weakness. It is a form of honesty.
Seen this way, the contradiction disappears. A good AI should hesitate less and doubt better.
In 1576, Michel de Montaigne had a medal struck with his motto. It shows a set of scales with perfectly balanced pans, and a question: "What do I know?". For Montaigne, this question was not an admission of weakness, but the starting point of all honest thought: before asserting, ask yourself what you really know. Two thousand years earlier, Socrates already said that his only wisdom was knowing that he did not know.
What we should expect from an AI
Specialists have a word for this quality: calibration. A well-calibrated AI is confident when it is right, and cautious when it risks being wrong. Its confidence follows what it knows.
It is what we expect from a good weather forecaster. When they announce a 70% chance of rain, it should rain on roughly seven days out of ten in those conditions. No more, no less. They do not know everything, but they know exactly what they do not know.
That is why the hallucination rate perhaps matters more than performance scores. An assistant that knows a lot but makes things up when it does not know is dangerous, because you can never tell when to believe it.
What to take away
The passer-by who sends you to the bakery did not know the way. But they saved you more time than the one who thought they did.
An intelligence is not measured only by what it knows. It is also measured by how accurately it knows what it does not know.