Authors recommend guidelines for designing and deploying audits of opaque AI systems to mitigate algorithmic harms to users and society.
WASHINGTON – A new report, “AI Audit-Washing and Accountability,” finds that auditing could be a robust means for holding AI systems accountable, but today’s auditing regimes are not yet adequate to the job. The report assesses the effectiveness of various auditing regimes and proposes guidelines for creating trustworthy auditing systems.
Various government and private entities rely on or have proposed audits as a way of ensuring AI systems meet legal, ethical and other standards. This report finds that audits can in fact provide an agile co-regulatory approach—one that relies on both governments and private entities—to ensure societal accountability for algorithmic systems through private oversight.
But the “algorithmic audit” remains ill-defined and inexact, whether concerning social media platforms or AI systems generally. The risk is significant that inadequate audits will obscure problems with algorithmic systems. A poorly designed or executed audit is at best meaningless and at worst even excuses harms that the audits claim to mitigate.
Inadequate audits or those without clear standards provide false assurance of compliance with norms and laws, “audit-washing” problematic or illegal practices. Like green-washing and ethics-washing before, the audited entity can claim credit without doing the work.
The paper identifies the core specifications needed in order for algorithmic audits to be a reliable AI accountability mechanism:
- “Who” conducts the audit—clearly defined qualifications, conditions for data access, and guardrails for internal audits;
- “What” is the type and scope of audit—including its position within a larger sociotechnical system;
- “Why” is the audit being conducted—whether for narrow legal standards or broader ethical goals, essential for audit comparison, along with potential costs; and
- “How” are the audit standards determined—an important baseline for the development of audit certification mechanisms and to guard against audit-washing.
Algorithmic audits have the potential to increase the reliability and innovation of technology in the twenty-first century, much as financial audits transformed the way businesses operated in the twentieth century. They will take different forms, either within a sector or across sectors, especially for systems that pose the highest risk. Ensuring that AI is accountable and trusted is key to ensuring that democracies remain centers of innovation while shaping technology to democratic values.
But as algorithmic audits are encoded into law or adopted voluntarily as part of corporate social responsibility, the audit industry must arrive at shared understandings and expectations of audit goals and procedures. This paper provides such an outline so that truly meaningful algorithmic audits can take their deserved place in AI governance frameworks.
The report is from the German Marshall Fund’s Digital Innovation and Democracy Initiative (GMF Digital), and is by former GMF Visiting Senior Fellow and Professor at Rutgers Law School (on leave) Ellen P. Goodman—now Senior Advisor for Algorithmic Justice at the National Telecommunications and Information Administration (NTIA)—and Program Manager and Fellow Julia Tréhu. It follows on GMF Digital’s “Auditing Algorithms for Accountable Tech” webinar as well as expert workshops convened on the subject over recent months.