Audit, Don't Explain -- Recommendations Based on a Socio-Technical Understanding of ML-Based Systems

Audits may be more effective than explanations: The paper argues that systematic audits of ML systems are preferable to explainable AI (XAI) approaches, suggesting that understanding how a system works matters less than verifying that it works fairly and safely in practice.

ML systems require socio-technical evaluation: The paper emphasizes that ML-based systems cannot be evaluated purely on technical grounds—their real-world impacts depend on how they interact with organizational practices, human decision-makers, and social contexts, requiring audits that examine these connections.

Independent institutional oversight is needed: The paper recommends establishing independent auditing bodies similar to Germany's TÜV (technical inspection association) and Stiftung Warentest (consumer testing organization) to systematically verify that ML systems serve the public interest rather than relying on companies' own explanations.

Public law governance should guide ML deployment: The paper advocates for institutions governed by public law to oversee ML systems, suggesting that regulatory frameworks similar to those for consumer products and industrial safety should apply to algorithmic systems affecting the public.

Transparency alone is insufficient for accountability: By positioning audits as preferable to explanations, the paper implies that simply making ML systems more interpretable does not guarantee they operate responsibly—external verification and testing are necessary safeguards.