A precise number can make uncertainty easier to grasp while also concealing which parts of the probability model itself remain uncertain.
Ellsberg distinguishes risk with known probabilities from ambiguity about the probability model itself. A single persuasive AI probability estimate can obscure uncertainty across several plausible models.
A precise number can make uncertainty easier to grasp while also concealing which parts of the probability model itself remain uncertain.
In Ellsberg’s famous urn choice problems, people often prefer a situation with known probabilities to one in which the distribution of colours is unknown, even when a single conventional subjective probability distribution struggles to explain the combination of choices.
Ambiguity aversion is robust on average but not universal; ambiguity seeking also occurs, and behaviour varies with the source, probability and context. The case therefore does not support a general rule that ‘people avoid uncertainty’.
Uncertainty about the probabilistic model is itself relevant to decisions. Compressing this second-order uncertainty into a single probability may remove relevant epistemic structure.
AI transfer to investigate: AI may change how uncertainty is represented by presenting advice as a single precise probability, a probability alongside confidence in the evidence, or several plausible models. The mechanism is old; the representation changes.
Proposed test, not yet conducted: Present identical ambiguous evidence under four conditions: no AI; one precise AI probability; probability plus confidence in the evidence/model; several plausible scenarios. Later add confirming or falsifying information and measure confidence, choice and updating.
The research aim is not less trust, but better-calibrated trust. Risk, ambiguity and model uncertainty should remain separately visible where relevant.
Updating a judgement after new information and calibration are different outcomes. Calibration compares stated probabilities with observed outcomes across suitable repeated tasks.
Ellsberg (1961), Risk, Ambiguity, and the Savage Axioms
Ellsberg (1961, pp. 650–653 and 657–660) discusses urn choices and distinguishes probability judgements from confidence in them. His original response account was non-experimental. This discussion is not a comprehensive assessment of modern follow-up research or of the proposed AI test.
Help improve this case
We review your contribution before updating the text or sources.