Fixing Recommender Systems: From Identification of Risk Factors to Meaningful Transparency and Mitigation

Published as the first VLOPs/VLOSEs were preparing their inaugural DSA risk assessments in August 2023, this briefing argues that recommender systems deserve special regulatory scrutiny because their design choices are a direct driver of the 'systemic risks' Article 34 of the DSA requires platforms to identify and mitigate.

The brief compiles six mechanisms by which recommender-system design can produce such risks: amplifying borderline content because it drives engagement, rewarding provocative engagement with greater reach, making opaque editorial choices that boost or suppress users, exploiting behavioural data to personalise content in ways that harm wellbeing, building in addictive features, and using personal data in ways that enable discrimination. For each, it cites available independent research while acknowledging that outside researchers remain constrained by platforms' unwillingness to share the data needed to verify causal claims.

It closes with a concrete technical checklist, covering the recommender system's algorithmic architecture, input features, training data and interpretability tooling, that it argues the European Commission, independent auditors and researchers should be able to demand from platforms in order to verify risk-assessment claims rather than take them on faith.