Abstract: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.
What problem does this paper attempt to address?
### Problems the Paper Attempts to Address
The paper "Towards an AI Accountability Infrastructure: Gaps and Opportunities in AI Audit Tools" attempts to address the following issues: Despite the increasing attention AI audits receive at the policy level, effectively conducting AI audits remains very challenging. Specifically, the paper focuses on the following aspects:
1. **Inadequacies of Existing Tools**: Although many tools aim to help practitioners set standards and evaluate AI systems, these tools often fail to support the accountability goals in AI audits. For example, many tools focus solely on assessing the performance, fairness, and safety of AI systems, lacking functionalities to identify harms, communicate audit results, and advocate for subsequent changes.
2. **Challenges in the Audit Process**: Auditors face various challenges in their work, including difficulty finding suitable tools, obtaining high-quality and untampered data, applying consistent and comprehensive standards and methods, ensuring the integrity of the audit, and interdisciplinary collaboration.
3. **Incomplete Tool Ecosystem**: The current ecosystem of AI audit tools is incomplete, lacking all the resources needed to support the entire audit process. In particular, there are relatively few tools for harm discovery, data collection, and transparency infrastructure.
4. **Lack of Accountability**: Existing tools and methods are often insufficient to hold the builders and operators of AI systems accountable for their actions. Audits need not only to assess system performance but also to identify potential harms and develop effective strategies to address these harms or hold relevant stakeholders accountable for the audit results.
### Main Research Methods
To explore these issues, the authors conducted the following research:
- **Tool Classification**: By interviewing 35 AI audit practitioners and analyzing 390 tools, the authors mapped the current ecosystem of AI audit tools and divided it into seven main stages: harm discovery, standards identification and management, transparency infrastructure, data collection, performance analysis, audit communication, and advocacy.
- **Interview Survey**: Through semi-structured interviews, the authors gathered insights into the tools used by audit practitioners in their work and the challenges they face. These interviews revealed that many tools, although available, often fail to meet all the needs of auditors in practice.
### Main Findings
- **Uneven Tool Distribution**: Most tools are concentrated in the evaluation stage, particularly in standards identification and management and performance analysis, while other critical stages (such as audit communication, harm discovery, and advocacy) have relatively fewer tools.
- **Challenges in Practical Application**: Despite the availability of many tools, auditors still face numerous challenges, including difficulty involving affected stakeholders, applying consistent and comprehensive standards and methods, obtaining high-quality data, ensuring audit integrity, and interdisciplinary collaboration.
- **Diverse Tool Needs**: Auditors need more tools that support the entire audit process, especially in harm discovery, data collection, and transparency infrastructure.
### Conclusions and Recommendations
The authors believe that current AI audit tools are mainly concentrated in the evaluation stage, neglecting the needs of other critical stages in the audit process. Therefore, they suggest:
- **Developing More Comprehensive Tools**: In addition to evaluation tools, there should be tools that support harm discovery, data collection, transparency infrastructure, audit communication, and advocacy.
- **Promoting Interdisciplinary Collaboration**: Strengthening collaboration between different disciplines to ensure the comprehensiveness and effectiveness of audits.
- **Improving Tool Accessibility and Quality**: Providing more high-quality, open-source tools so that auditors can more easily access and use them.
Through these improvements, the entire AI audit process can be better supported, achieving true accountability.