At the Research Data Alliance (RDA) Organisational Assembly of 23^rd June 2026, Dr Silvia Milano, Senior Lecturer and EDI Head of Research at the Technical University of Munich (TUM) presented a webinar sharing her research on recommender systems. The research supports the EDI’s core position that ethics is a practice, not an abstract exercise and frameworks for responsible data use must always be situated in the specific context of any given application.

Recommender systems are everywhere, used in entertainment platforms like Netflix and Spotify, social media, online shopping, news, and scholarly research. They are algorithms that filter information from vast catalogues of options based on inferences about individual user preferences and shape how people access information, opportunities, and services.

While these systems are designed to keep us engaged, they tend to prioritise business goals (such as clicks and revenue) over what is good for users or society. Dr Milano argues that we need to move beyond simply making these systems more accurate, and instead design them to better support fairness, diverse viewpoints, and people’s ability to make their own informed choices.

Despite their scale and influence, how recommender systems should operate ethically is poorly understood. Most systems prioritise user engagement, retention, or monetisation, which frequently diverges from socially beneficial goals. Dr Milano’s research aims to move beyond an accuracy-first model and reframe recommendations in terms of social value and public interest.

A Brief History of Recommender Systems

Early research into using algorithmic systems for user modelling began in the late 1970s and has gained momentum since the 1990s, driven by the publication of large user interaction datasets. The Netflix Prize is a landmark example. Netflix, originally a DVD subscription service, published a large dataset of user ratings and rental histories and offered a prize for improving recommendation algorithms. This generated widespread research interest.

Dr Milano draws a critical observation from this history: the field was built from the start on behavioural data arising from commercial interactions. Surveillance was therefore embedded in how the field developed.

How Recommendation is Currently Approached

Recommendation is typically broken into two subtasks: a Prediction task (primary), estimating how a user will interact with a given item – the core of any recommender system and Ranking and display (secondary), presenting a ranked list of items, potentially incorporating side constraints such as business incentives or coverage goals.

The field is currently characterised by a highly individualistic, accuracy-centred framework. The dominant goal is to accurately predict the relevance of individual items to individual users within a largely commercial context.

Problems with the Accuracy-First Paradigm

Dr Milano outlined key reasons why this framing is problematic:

• Offline experiments don’t translate to real-world effectiveness. Even practitioners acknowledge that user studies show offline accuracy results are not indicative of how algorithms perform in practice.
• Higher prediction accuracy does not necessarily mean a better system. For example, a social media recommender that accurately predicts what content a user will engage with may be more addictive but not socially better.
• While personalised recommendations are credited with increasing business performance and user satisfaction, these claims are seldom backed by empirical research.
• Algorithms used in practice are often not the most accurate, but those that better serve business goals.
• Personalisation creates multiple negative effects: spread of misinformation, anti-competitive practices, and encouragement of personal data collection and surveillance, which is increasingly being monetised in problematic ways.
• The focus on predicting user preferences obscures the interests of other stakeholders and makes it harder to justify alternative approaches towards Recommender Systems for Social Good.

Proposed Changes

To address these problems, Dr Milano proposes a conceptual shift: from an individualistic framework centred on accurate prediction, to treating recommendation as a lens for social good. Achieving this shift requires addressing three fundamental questions:

1. How can recommender systems account for multiple stakeholders?
2. How can social good be defined for specific applications?
3. How can systems be designed to support objectives such as fairness, diversity, autonomy, well-being, and judicious use of resources given the significant environmental and operational costs of running such systems?

Dr Milano acknowledges there are significant challenges. Defining “good” is often controversial. Measuring influence across stakeholders is difficult and the impact at an individual level can conflict with impact at a social level. However, there are pathways for addressing these challenges.

Current benchmarks derive from commercial datasets with specific characteristics which are not suited for application to the social good. Using data from non-commercial, multiple stakeholder contexts would address this bias. Furthermore, the concept of “accurate prediction” should be abandoned, replacing fixed preferences with context-specific recommendations.

Developing tools for tracking the individual, social and environmental impacts of recommender systems and incorporating these into evaluation frameworks could re-define recommender systems as a force for social good.

The recording of the session, presentation slides, summary of the Q&A session and further details about Dr Milano’s research can be found on the RDA website here and on the EDI Zenodo page here.

Work on this project has been partly funded by the Alexander von Humboldt Foundation through a Humboldt Fellowship for Experienced Researchers.

For more information please contact s.milano[at]tum.de.