Audits are critical mechanisms for identifying the risks and limitations of deployed artificial intelligence (AI) systems. However, the effective execution of AI audits remains incredibly difficult. As a result, practitioners make use of various tools to support their efforts. Drawing on interviews with 35 AI audit practitioners and a landscape analysis of 390 tools, we map the current ecosystem of available AI audit tools. While there are many tools designed to assist practitioners with setting standards and evaluating AI systems, these tools often fell short of supporting the accountability goals of AI auditing in practice. We thus highlight areas for future tool development beyond evaluation -- from harms discovery to advocacy -- and outline challenges practitioners faced in their efforts to use AI audit tools. We conclude that resources are lacking to adequately support the full scope of needs for many AI audit practitioners and recommend that the field move beyond tools for just evaluation, towards more comprehensive infrastructure for AI accountability.
翻译:暂无翻译