KI-Vertrauenskalibrierung

Too much trust can hide errors; too little can prevent useful assistance. The aim is an informed, revisable judgment tied to the task and operating conditions.

Wichtige Erkenntnisse

• Base reliance on task-specific evidence.
• Evaluate confidence rather than accepting it at face value.
• Measure how people use and override recommendations.

Distinguish statistical calibration from a person’s trust. A calibrated probability score has an empirical relationship to how often predictions are correct across comparable cases. Human trust also depends on interface design, past experience, explanations, and the consequences of mistakes. A model’s verbal confidence is another output to evaluate. It may sound certain without reliable evidence. Even a statistically calibrated model can perform poorly, and good calibration on one dataset does not guarantee calibration after a shift in language, domain, or task. Give users information that supports independent checking. Show sources, describe relevant limitations, and separate verified results from inferred conclusions. Avoid decorative confidence badges that suggest more precision than was measured. Test how people use the system, including whether they notice errors and exercise overrides appropriately. Measure both overreliance and unnecessary rejection. Update guidance when behavior changes, and preserve a straightforward way to report mistakes or complete the task without the model.

1. Imagine 100 predictions assigned approximately 80% probability of being correct. Only 55 are correct in a representative held-out sample.
2. The group is overconfident under this evaluation; the displayed 80% should not be treated as established reliability.
3. Repeat with enough examples across probability ranges and important subgroups before changing how scores are shown to users.

The invented counts illustrate calibration assessment and its dependence on the evaluation sample.