Introducing the EDI Research Spotlight on Recommender Systems for Social Good.

Author Dr Silvia Milano is EDI Head of Research at TUM.

Recommender systems shape an increasing share of our informational, cultural, and social environments online. From news feeds and streaming platforms to scientific discovery and public services, algorithmic recommendations influence what we see, what we choose, and ultimately how we understand ourselves and the world. At the Ethical Data Initiative (EDI), this semester’s Research Spotlight focuses on a central question: what would it mean to design and evaluate recommender systems explicitly for social good?

Personalisation is often presented as an unambiguous good. Accurate, tailored recommendations are said to improve user satisfaction, engagement, and even access to information. These claims are now common not only in commercial settings, but also in public service media and other domains traditionally associated with social value. Yet, research has repeatedly shown such claims to be rarely substantiated. In practice, recommender systems tend to be optimised for business objectives such as retention, growth, and monetization, which are not often aligned with personal flourishing or public benefit.

This gap between stated aims and operational realities has given rise to many concerning consequences. Personalisation has been linked to well-documented harms, including polarisation, the amplification of misinformation, and anti-competitive dynamics. It also incentivises the large-scale collection of personal data, which increasingly becomes a secondary asset, as it is repurposed to train generative AI models far beyond the original context in which the data was collected. Acknowledging these risks does not mean denying that recommender systems can be used for socially positive ends. Rather, it highlights how applications of “Recommender Systems for Good” (RS4Good) remain comparatively underexplored.

One reason lies in the current structure of the field itself. Research on recommender systems is sharply divided between academic and commercial domains. Academic work often focuses on narrowly defined problems, benchmark datasets, and incremental improvements in accuracy metrics. Meanwhile, real-world systems operate under business incentives, opaque objectives, and proprietary data that are largely inaccessible to independent researchers. As a result, questions that matter most for social good—What problem is this system meant to solve? Whose values are embedded in its design? How should success be measured?—are frequently sidelined.

This “accuracy-first” culture is increasingly difficult to justify. Accuracy metrics presume a clear and stable target, usually framed as predicting individual preferences. But preferences are neither fixed nor value-neutral. They are shaped by context, power, and prior recommendations themselves. Optimising for accuracy alone risks reinforcing existing inequalities, narrowing exposure, and obscuring the normative choices built into system design. Moreover, accuracy is often not even the main goal of actual recommendations. From an ethical perspective, the problem is not merely that accuracy is insufficient, but that it often distracts from more fundamental questions about purpose and responsibility.

EDI’s vision is to help reorient research and practice around these deeper issues, by examining the epistemological assumptions and ethical implication of common recommender systems evaluation metrics and design choices. This involves foregrounding approaches that question and reconfigure the goals of recommender systems, rather than treating them as given. It also means expanding the range of datasets and evaluation methodologies used in RS4Good research—prioritising openness, auditability, and contextual relevance over convenience.

Recent work associated with EDI reflects this commitment. Ongoing and recent outputs have examined the societal harms and risks of recommender systems, including in the context of generative models, as well as the role of digital nudging and evaluation practices. These efforts were complemented by an interdisciplinary workshop which took place at LMU and TUM in May 2025, bringing together computer scientists, philosophers, economists, and social scientists to interrogate how recommender systems shape agency, autonomy, and knowledge across domains.

Looking ahead, EDI aims to foster a broader cultural change in how recommender systems are studied and evaluated. As predictive accuracy is based on shaky epistemological foundations, and tends to obscure the more important ethical issues at play in recommendations, we advocate for greater attention to problem formulation and evaluation methods aligned with social objectives. In practical terms, this includes supporting community-driven efforts to share information about datasets suitable for RS4Good applications, and to exchange best practices for evaluation beyond accuracy—such as measures related to diversity, fairness, interpretability, user empowerment, and socially good outcomes.

This is not a challenge EDI can—or should—address alone. Progress toward recommender systems that genuinely serve social good depends on collaboration across research communities and disciplines. To this end, we are working to build connections with partners such as the Research Data Alliance and other EDI affiliates, with the goal of creating shared platforms and resources that lower barriers to RS4Good research and practice.

We invite researchers, practitioners, policymakers, and infrastructure providers who are interested in these questions to engage with us. Expressions of interest, ideas for collaboration, and contributions of datasets or evaluation approaches are particularly welcome. These discussions will feed directly into the next EDI Town Hall and into a joint EDI–RDA meeting planned for October 2026, where we aim to move from reflection to coordinated action.

If recommender systems increasingly function as “algorithms of choice,” then the ethical stakes could not be higher. Reimagining how these systems are designed and evaluated is not simply a technical challenge, but a collective responsibility. 

Declaration on Generative AI

In the preparation of this blog post, the author used ChatGPT5.2 for spell checking and rephrasing. After using this tool, the author reviewed and edited the text as needed and takes full responsibility for the blog’s content.