对AI评估方法的批判性审视报告

Executive Summary Policymakers frequently invoke explainability and interpretability as keyprinciples thatresponsible and safe AI systems should uphold. However, it is unclear how evaluationsof explainability and interpretability methods are conducted in practice. To examineevaluations of these methods, we conducted a literature review of studies that focuson the explainability and interpretability of recommendation systems—a type of AIsystem that often uses explanations. Specifically, we analyzed how researchers (1)describe explainability and interpretability and (2) evaluate their explainability andinterpretability claims inthe context of AI-enabled recommendation systems. Wefocused on evaluation approaches in the research literature because data on AIdevelopers’ evaluation approaches is not always publicly available,and researchers’approaches can guide the types of evaluations that AI developers adopt. We find that researchers describe explainabilityand interpretability in variable waysacross papers and do not clearly differentiate explainability from interpretability. Wealso identify five evaluation approaches that researchers adopt—case studies,comparative evaluations, parameter tuning, surveys, and operational evaluations—andobserve that research papers strongly favor evaluations of system correctness overevaluations of system effectiveness. These evaluations serve important but distinctpurposes. Evaluations of system correctness test whetherexplainablesystems arebuilt according to researcher specifications, and evaluations of system effectivenesstest whetherexplainablesystems operate as intended in the real world.If researchersunderstand and measure explainabilityor other facets of AI safetydifferently, policiesfor implementing or evaluatingsafeAIsystemsmay not be effective.Although furtherinquiry is needed to determine whether these results translate to other research areasand the extent to which research practicesinfluence developers, these trends suggestthat policymakers would do well to invest in standards for AI safetyevaluationsandenablea workforce that can assess the efficacy oftheseevaluations in differentcontexts. Table of Contents Executive Summary................................................................................................................................1Introduction...............................................................................................................................................3Background...............................................................................................................................................6Recommendation Systems..............................................................................................................6Explainability and Interpretability.................................................................................................7Methodology............................................................................................................................................9Findings...................................................................................................................................................11Explainability Descriptions...........................................................................................................11Descriptions That Rely on the Use of Other Principles...................................................12Descriptions That Focus on an AI System’s Technical Implementation....................12Descriptions That State the Purpose of Explainability Is to Provide aRationale for Recommendations............................................................................................12Descriptions That Articulate the Intended Outcomes of Explainable orInterpretable Systems...............................................................................................................13Explainability Evaluation Approaches......................................................................................14Case Study....................................................................................................................................14Comparative Evaluation............................................................................................................15Parameter Tuning.......................................................................................................................16Survey.............................................................................................................................................16Operational Evaluation..............................................................................................................17Evaluations of System Correctness and Effectiveness........................................................19Policy Considerations.........................................................................................................