Value-Aligned Evaluation for Everyday AI
Mar 9, 2026·
·
1 min read
Xia Ruize
Abstract
This paper proposes an evaluation framework for everyday AI systems that supplements accuracy metrics with measures of explanation quality, contestability, accessibility, and institutional fit. The argument is that systems used in education, communication, and public decision support must be evaluated in terms that remain legible to the humans affected by them.
Type
Publication
Artificial Minds, Human Values Working Paper Series
Standard evaluation pipelines are often optimized for benchmark performance. That works well when the question is purely technical. It works less well when the system is embedded in social institutions.
This paper argues that value alignment must become operational. The proposed framework therefore asks evaluators to score a system along four additional dimensions:
- Explanation quality — Can a non-specialist understand what the system is doing?
- Contestability — Is there a meaningful route for correction or appeal?
- Accessibility — Can diverse users interact with the system without avoidable exclusion?
- Institutional fit — Does the surrounding organization have the practices needed to use the tool responsibly?
The broader claim is that AI should be tested not only in the lab, but also against the moral and organizational conditions of its use.
