User explainability score

Description

Measures whether real users understand, trust appropriately and can act on an AI system's explanations, combining comprehension tests with validated rating scales.

Technical explanation methods do not guarantee that users understand anything. A user explainability score quantifies explanation quality from the user's side: comprehension (can users predict or describe why the system acted), calibration (does trust track actual model accuracy), actionability (can users change outcomes) and subjective satisfaction. Hoffman and colleagues proposed explanation satisfaction and goodness checklists, and forward-prediction tasks, where users predict what the model will do, are an objective measure. Rigour involves comparing explanation designs experimentally, measuring calibrated reliance rather than raw trust, and combining objective and subjective components into a scored index with known reliability.

Attributes

Loss type
Latent
Stage
Evaluation
Data
Mixed
Effort
Weeks

Steps

  1. Define the explanation goals: understanding, calibrated trust, actionability, contestability.
  2. Build comprehension tasks, such as forward prediction ('what will the system recommend for this profile?') and counterfactual tasks ('what would change the outcome?').
  3. Administer a validated explanation satisfaction scale and a trust scale after exposure.
  4. Measure reliance calibration: agreement with the AI when it is right versus when it is wrong.
  5. Combine component scores into an index (e.g. 0-100), checking reliability of the composite.
  6. Compare explanation designs in a between-subjects study with at least 30-50 participants per condition.

Sources

  • The Alignment Problem (Brian Christian)
  • Artificial Intelligence for Humans (Jeff Heaton)
  • Methods Atlas (Vivek Gopinath)
  • Hoffman, Mueller, Klein & Litman (2018) Metrics for explainable AI: challenges and prospects
  • The Alignment Problem (Christian)
  • AI for Humans (Heaton)
  • Methods Atlas (Gopinath)
  • The Inquiry Series, Ch 58 — AI Explainability & Value Alignment Audit

Also known as

  • Explanation satisfaction measure
  • Explanation quality evaluation

Categories

  • User Research
  • Usability & Evaluation