PaperlayerView exact revision on arXiv

Source-linked AI summary

The Quest for Reliable Metrics of Responsible AI

Theresia Veronika Rampisela, Maria Maistro, Tuukka Ruotsalo, Christina Lioma

arXiv:2510.26007v1cs.CYcs.AIcs.IRcs.LG

TL;DR

Responsible-AI progress is often measured with evaluation metrics, but their reliability and robustness have received less scrutiny. The paper synthesizes prior work on recommender-system fairness metrics into guidelines for reliable metrics, including for AI in Science; its findings identify mathematical and empirical limitations that motivate careful metric evaluation.

  • Problem

    Few studies examine the reliability and robustness of responsible-AI evaluation metrics, even though their scores can be misleading and metrics guide early system development.

    Source: §1 · p. 2

  • Method

    The paper reflects on prior analyses of recommender-system fairness metrics and distills their findings into non-exhaustive guidelines for reliable responsible-AI metrics.

    Source: §1 · p. 2, §2.2 · p. 4

  • Results

    Prior analyses found fairness metrics with mathematical flaws or unreachable score bounds, scores near zero regardless of fairness, redundant measures, and group metrics that cannot estimate individual fairness.

    Source: §2.2 · p. 3

  • Takeaways & Limitations

    Metric developers and evaluators should examine valid inputs, score ranges and extrema, sensitivity, redundancy, and both group and individual fairness.

    Source: §2.2 · p. 4, §3 · p. 4

    The guidelines are explicitly non-exhaustive and described as a bare minimum for reliable AI metrics.

    Source: §4 · p. 5

Abstract

from arXiv · show

The development of Artificial Intelligence (AI), including AI in Science (AIS), should be done following the principles of responsible AI. Progress in responsible AI is often quantified through evaluation metrics, yet there has been less work on assessing the robustness and reliability of the metrics themselves. We reflect on prior work that examines the robustness of fairness metrics for recommender systems as a type of AI application and summarise their key takeaways into a set of non-exhaustive guidelines for developing reliable metrics of responsible AI. Our guidelines apply to a broad spectrum of AI applications, including AIS.

1 Introduction

Recommender systems are widely used across personal and professional settings, and fairness matters because unequal recommendations can affect people and fields differently. The paper addresses limited scrutiny of fairness metrics’ reliability and distills prior recommender-system work into guidelines for reliable responsible-AI metrics.

1 Introduction

  • Recommender systems help users find relevant items and serve both high- and low-stakes professional and personal purposes.Examples include recommending scientific papers, citations, and job positions.Source: §1 · p. 1
  • Unequal job recommendations may exacerbate gender pay gaps, while imbalanced paper recommendations can disadvantage disciplines or researchers and narrow science’s inclusiveness.The authors emphasize that fair recommender systems are scientifically and societally important.Source: §1 · pp. 1–2
  • Few studies assess the reliability and robustness of responsible-AI evaluation metrics, although unreliable scores can mislead and these metrics guide early development.The paper condenses prior recommender-system fairness work into non-exhaustive guidelines also relevant to AI in Science.Source: §1 · p. 2

2 Fairness Evaluation

Recommender-system fairness metrics vary by the subject and granularity they assess, and their limitations can make their scores difficult to trust or interpret. The authors summarize theoretical and empirical limitations, proposed approaches, and practical evaluation guidelines.

2.1 Evaluation of Fairness in Recommender Systems

  • User fairness concerns disparities in recommendation effectiveness, item fairness concerns exposure, group fairness compares subject groups, and individual fairness concerns utility variation across users or items [Wang et al., 2023].Source: §2.1 · p. 2
  • Some fairness metrics have limited or unknown meaning because more than 30 measures span different subjects and granularities, while some have not been analyzed for limitations [Wang et al., 2023].Fairness evaluation distinguishes user from item fairness and group from individual fairness.Source: §2.1 · pp. 2–3

2.2 Our Contributions to RS Fairness Evaluation

  • Fairness metrics may crash, have unknown or unreachable score bounds, or lack known inputs for their fairest and unfairest outcomes [Rampisela et al., 2024a, 2025a].Unreachable extrema make scores difficult to interpret because a nominal range may not match the scores any input can attain.Source: §2.2 · p. 3
  • Empirically, some metrics score close to zero regardless of fairness, similar metrics can be redundant, and group fairness cannot estimate individual fairness [Rampisela et al., 2025b].The authors recommend considering multiple granularities for a more comprehensive fairness view.Source: §2.2 · p. 3
  • The authors’ investigation raises awareness that metric limitations can affect usability and interpretability, so scores should not be trusted without understanding how a metric operates.Source: §2.2 · pp. 3–4
  • Corrected metrics avoid invalid mathematical operations and use min-max normalisation to map fairest and unfairest cases to 0 and 1; another proposed metric jointly evaluates effectiveness and fairness [Rampisela et al., 2024a, 2025a, 2025c].The authors also released public source code for both approaches.Source: §2.2 · p. 4
  • The guidelines recommend corrected metrics, careful interpretation of scores, avoiding redundant measures, and evaluating both group and individual fairness.Some measures can score close to the fairest case and overestimate fairness.Source: §2.2 · p. 4

3 Guidelines for Formulating Reliable Metrics

The authors generalize findings from recommender-system fairness evaluation into non-exhaustive questions for developing reliable responsible-AI metrics. The questions address valid inputs, score meaning and range, sensitivity, and overlap with existing metrics.

3 Guidelines for Formulating Reliable Metrics

  • Check whether any input cases must be excluded to prevent invalid mathematical operations.Source: §3 · p. 4
  • Specify the metric’s score range and how its values should be interpreted.Source: §3 · p. 4
  • Identify which inputs produce the minimum and maximum metric scores.Source: §3 · p. 4
  • Assess how sensitive the metric is to changes in its input.Source: §3 · p. 4
  • Check whether the metric reaches conclusions similar to an existing measure.Source: §3 · p. 4
  • Publicly released source code is available for the corrected metrics and the effectiveness-fairness joint evaluation approach.Source: §3 · p. 4

4 Conclusions

The guidelines are a non-exhaustive bare minimum for reliable responsible-AI metrics, and the authors call for broader collaboration to make them more concrete, especially for high-stakes AIS.

  • The guidelines are explicitly non-exhaustive and intended as a bare minimum for reliable metrics across AI technologies.Source: §4 · p. 5
  • The authors call for collaboration across users, technical actors, social scientists, ethics experts, policy makers, and government agencies to identify critical AIS aspects and audit metric reliability.Source: §4 · p. 5
  • They expect this collaboration to produce more concrete, measurable guidelines for responsible AI development, particularly in high-stakes contexts such as AIS.Source: §4 · p. 5
  • The authors urge greater attention to evaluation alongside technology development, noting that evaluation metrics are often absent from AI policy discussions despite quantification's role in regulation and policy-making.Source: §4 · p. 5
Loading 2510.26007v1…