Uncertainty should be expressed when present

What it is

Calibrated uncertainty expression tells people when perception, interpretation, or planning lacks sufficient evidence and whether to wait, confirm, or intervene. It translates internal uncertainty into consequence-appropriate action rather than exposing a raw score.

Why it happens

Autonomous systems encounter distribution edges. Certain-looking output turns fluency into assumed confidence; universal warnings create fatigue. Calibrated levels combine evidence, consequence, and intervention availability so reliance changes with actual reliability.

Studying it

Confidence, expression, and error cost can be varied while measuring calibrated choices to confirm, take over, or continue, plus time, over-reliance, and unnecessary intervention. Model confidence must first predict real correctness. Numeric probability, categories, and action guidance should be compared separately.

Where it stops holding

Some model scores are not probabilities and must not be labelled success rates. Rare high-consequence events may lack calibration data. Expression cannot transfer responsibility to an uninformed person, and precise numbers without an available action merely add anxiety.

Applying it

• Calibrate confidence against outcomes before mapping ranges to continue, slow, confirm, or safe stop.
• Identify the uncertainty source and information a person can change, such as occlusion or two referents.
• Measure appropriate reliance across consequences while counting extra interventions caused by low-confidence false alarms.

Related

• Same group: X3.03.1 Anthropomorphic expression must not exaggerate understanding · X3.03.2 Pretending to understand is worse than admitting failure
• Adjacent: X3.07 Explainable decision basis · X3.08 Communicating uncertainty
• Search terms: uncertainty communication · confidence calibration · appropriate reliance