The LLM4News project is built on a central question: how can we personalise news content using AI without compromising on reliability and neutrality? While generative models are powerful in creating tailored experiences, their tendency to produce errors or biased phrasing makes them risky in the sensitive context of journalism.

The BBC’s recent research reinforces these concerns. Their study demonstrated that even leading AI assistants regularly provide incomplete, misleading, or skewed information. This creates a significant challenge for media organisations seeking to maintain public trust while experimenting with AI-driven personalisation.

Our project investigates Retrieval-Augmented Generation (RAG) as a potential solution. Unlike standard LLM responses that rely solely on pre-trained knowledge, RAG combines generative capabilities with retrieval from trusted, up-to-date sources. This means an LLM can draw on curated, verified information before producing personalised outputs, thereby reducing hallucinations and bias.

Additionally, the Flemish media context struggles with the fact that LLM's are dominantly trained with English data and in the case of Dutch LLM's, with data from the Netherlands. This is a critical barrier for Flemish content creation.

LLM4News aims to build prototypes of RAG-powered systems tailored for Flemish journalism.