News recommendation algorithms often reinforce filter bubbles, but AI may also help surface a broader range of viewpoints. This is one of the findings from A Puzzle of Perspectives, the PhD dissertation of computational linguist Myrthe Reuver at Vrije Universiteit Amsterdam.

Reuver explored how Natural Language Processing (NLP) methods can support more diverse news recommendations. She focused on stance detection, training large language models to determine whether texts argue for or against certain topics, such as abortion laws or climate policy. While the models could identify simple stances, they struggled with nuance and topic complexity.

A key insight: AI works best in collaboration with human experts. In one experiment on sexism detection, a model trained with input from gender studies experts outperformed both AI-only and human-only approaches. Similar results emerged in a climate stance dataset, where social science theory improved model understanding.

The experiments ran on supercomputers from SURF, the Dutch infrastructure for academic computing. Reuver emphasises the importance of such independent resources in light of Europe’s growing concerns about Big Tech dominance in AI.

Her findings align with broader debates about AI and democracy. “This research shows the limits of current models — but also the opportunities for AI to support better public discourse when paired with human insight.”