Human writing carries a weak but real signal: people hedge when they are unsure. We have spent a lifetime calibrating to it, which is why a confident paragraph reads as a reliable one.
That signal does not survive contact with a language model. A model writes with the register of its training data, and financial commentary is written confidently. The tone is copied from the genre, not generated from any internal assessment of how well-founded the claim is.
What to substitute
Read for attachment rather than tone. Which specific figures appear, and where did each come from? A paragraph full of adjectives and no numbers has told you nothing. A paragraph with four numbers and no sources has told you something unverifiable, which is worse, because it feels like content.
The most useful thing in an AI market answer is usually the least dramatic sentence in it: the one saying what was read and when.
Hedging is not the same as honesty
The other failure runs the opposite way. A model that hedges everything — could go either way, depends on conditions — is not being careful, it is being useless. Calibration means being specific when the evidence supports it and explicit about the gap when it does not.
The goal is an answer that says what the map shows, what has historically followed that configuration, and where the evidence stops. All three, in that order, without the last one being used to excuse vagueness in the first two.
A practical test
Pick one number from any AI market answer and try to verify it. If the answer told you its source, this takes thirty seconds. If it did not, you have just learned what the whole answer is worth.