Tournesol: Permissionless Collaborative Algorithmic Governance with Security Guarantees

Recommendation algorithms play an increasingly central role in our societies. However, thus far, these algorithms are mostly designed and parameterized unilaterally by private groups or governmental authorities. In this paper, we present an end-to-end permissionless collaborative algorithmic governance method with security guarantees. Our proposed method is deployed as part of an open-source content recommendation platform https://tournesol.app, whose recommender is collaboratively parameterized by a community of (non-technical) contributors. This algorithmic governance is achieved through three main steps. First, the platform contains a mechanism to assign voting rights to the contributors. Second, the platform uses a comparison-based model to evaluate the individual preferences of contributors. Third, the platform aggregates the judgements of all contributors into collective scores for content recommendations. We stress that the first and third steps are vulnerable to attacks from malicious contributors. To guarantee the resilience against fake accounts, the first step combines email authentication, a vouching mechanism, a novel variant of the reputation-based EigenTrust algorithm and an adaptive voting rights assignment for alternatives that are scored by too many untrusted accounts. To provide resilience against malicious authenticated contributors, we adapt Mehestan, an algorithm previously proposed for robust sparse voting. We believe that these algorithms provide an appealing foundation for a collaborative, effective, scalable, fair, contributor-friendly, interpretable and secure governance. We conclude by highlighting key challenges to make our solution applicable to larger-scale settings.