Supervisors

Suitable for

Abstract

Background

Currently algorithmic platforms are designed to promote user engagement and addiction. Users’ agency of interacting these platforms is increasingly valued, including being able to opt out personalised advertisement or provide explicit user feedback to control recommendations. However, there is limited evidence showing effective such mechanisms work in practice. The risks of neglecting users' agency in algorithmic designs are exacerbated when much more sophisticated AI models are deployed on these platforms, underpinned by complicated feedback loops or increasingly human-like responses. 

In this project, we will explore the transparency of recommendation algorithms on leading social media platforms and their ability to align with users feedback, i.e. respecting their user agency. 

Objectives

1. Design a set of avatars of different demographic features, such as age, gender, or personal interests, in order to create the simulation data for assessing algorithmic impact. 
2. Define impact benchmarks, drawing on existing child-centred recommendation algorithm impact benchmarking

1. Produce a reproducible algorithm benchmark dataset and pipeline that others can extend to new models and domains.

Methodology

• Avatar Development: Generate avatars by drawing on existing methodologies and clear research goals
• Benchmark Design: Operationalize dimensions into computational checks (e.g., factual consistency for accuracy and user satisfaction, harmful manipulation detection for safety, presence of response to user feedback for user autonomy).
• Algorithm Testing: Collect responses from simulated algorithmic interactions
• Evaluation: Score outputs using a mix of automated methods and annotations.

Expected Contributions

• A tested avatar-based algorithm auditing methodology
• A multi-dimensional evaluation framework to assess support for user autonomy and agency
• Resources for researchers and policymakers working on responsible AI in family and education contexts.

References

Wood S (2024). Children and Social Media Recommender Systems: How Can Risks and Harms be Effectively Assessed in a Regulatory Context? https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4978809