Ethics in AI Research Seminar: Trustworthy Recommendations

As AI becomes more powerful and enmeshed in our lives, the issue of trust is gaining prominence.

We need to be able to trust artificial systems to operate in ways that are beneficial to their users and society more broadly, yet this trust is often broken. An interesting example are recommender systems: while we need to rely on them to discover relevant content, it is often unclear to what extent they serve our true interests.

Implicit in many approaches to improve the trustworthiness of recommendations is what I call a needs-based account of trust, according to which we trust recommenders to correctly identify and cater to the needs of their users. However, user needs can be highly variable and context-dependent, so accurately predicting them is a difficult task that requires intrusive collection of personal data. Moreover, the narrow focus on user needs does not address the wider issue of systemic effects that undermine trust in recommendations, such as their potential to manipulate audiences or amplify harmful content.

Building on Katherine Hawley’s and C Thi Nguyen’s work on the philosophy of trust, I explore an alternative account of trustworthy recommendations. On the commitment-based account, we trust recommenders to deliver recommendations based on an agreed set of criteria. After discussing the merits and identifying some limitations of this account, I end by sketching ways in which it can translate into practical research directions in delivering trustworthy recommendations.