Work

A cross-domain framework for auditing algorithms: From harmful content to self-preferencing

Year: 2025
Type: report
Abstract: As online platforms increasingly rely on algorithmic systems to curate, rank, and remove content, concerns about the societal and economic consequences of algorithmic decision-making have intensified, and with them the importance of algorithmic auditing. However, there is a lack of universal, practical guidance on the specific steps involved in designing and implementing platform audits across different problem domains and under varying conditions of data access, methodological constraints and legal obligations. This paper aims to address this gap by deriving a generalised framework with fine-grained, concrete steps required for the practical implementation of algorithmic audits using a broad focus on potential methodologies and a broad focus on potential application scenarios. The framework is based on a structural analysis of empirical audit approaches in the two domains of moderation of illegal and harmful content and of potential self-preferencing in recommendation algorithms, combined with further methodological literature. On this basis, the paper specifies the sequence of steps involved in carrying out audits in practice, identifies recurring challenges encountered across settings, and proposes a classification of audit approaches, incorporating advantages and disadvantages, and concrete examples. The proposed unified process model refines and operationalises existing high-level frameworks, while generalising more fine-grained domain-specific approaches. Furthermore, it demonstrates that meaningful audits can be conducted under restrictive access conditions, but that platform support can substantially improve the feasibility of precise, generalisable and causally informative findings. Taken together, these contributions offer regulators, platforms, and independent auditors a practical basis for systematically scrutinising complex platform algorithms and for interpreting audit results with greater rigour.
Language: English
Open Access status: green
OpenAlex ID: W7123379655
Cites: 109
Cited by:
Related to: 10