Two Dimensions of Subjective Uncertainty: Clues From Natural Language

We argue that people intuitively distinguish epistemic (knowable) uncertainty from aleatory (random) uncertainty and show that the relative salience of these dimensions is reflected in natural language use. We hypothesize that confidence statements (e.g., I am fairly confident, I am 90% sure, I am reasonably certain) communicate a subjective assessment of primarily epistemic uncertainty, whereas likelihood statements (e.g., I believe it is fairly likely, I'd say there is a 90% chance, I think there is a high probability) communicate a subjective assessment of primarily aleatory uncertainty. First, we show that speakers tend to use confidence statements to express epistemic uncertainty and they tend to use likelihood statements to express aleatory uncertainty; we observe this in a 2-year sample of New York Times articles (Study 1), and in participants' explicit choices of which statements more naturally express different uncertain events (Studies 2A and 2B). Second, we show that when speakers apply confidence versus likelihood statements to the same events, listeners infer different reasoning (Study 3): confidence statements suggest epistemic rationale (singular reasoning, feeling of knowing, internal control), whereas likelihood statements suggest aleatory rationale (distributional reasoning, relative frequency information, external control). Third, we show that confidence versus likelihood statements can differentially prompt epistemic versus aleatory thoughts, respectively, as observed when participants complete sentences that begin with confidence versus likelihood statements (Study 4) and when they quantify these statements based on feeling-of-knowing (epistemic) and frequency (aleatory) information (Study 5).