Trusted grounded retrieval-augmented generation for journalism
Abstract
The rapid evolution of generative artificial intelligence (AI) presents significant evaluation challenges. As large language models are increasingly exposed to misinformation, synthetic content, and imperfect web-scale training data, grounding model outputs in trusted, verifiable contexts has become essential. Conventional benchmarking based on static datasets does not adequately capture factual grounding and contextual relevance in operational settings. To address these limitations, the European Broadcasting Union adopted a use-case-driven framework focused on retrieval-augmented generation (RAG) for public service media (PSM). The paper examines how curated, high-quality PSM archives can anchor model behaviour and improve factual reliability. Four contributions are presented. First, the paper describes a practical RAG architecture covering query formulation, document chunking, vector retrieval, reranking, and contextual generation, with attention to reranking strategies that constrain model improvisation. Second, it proposes a multilayered evaluation framework combining information retrieval metrics, factual consistency checks, answer coverage assessment, and expert human judgement. Third, it introduces a smart writing assistant that uses trusted context to support journalistic ideation and drafting while preserving full editorial control. Finally, it explores bounded agentic workflows in which an additional layer intervenes only when retrieval quality is insufficient. Together, these components show how trusted corpora, transparent evaluation, and controlled automation can improve reliability while keeping journalists in command of the editorial process. This article is also included in The Business & Management Collection which can be accessed at https://hstalks.com/business/.
The full article is available to subscribers to the journal.
Author's Biography
Alexandre Rouxel is a Senior Project Manager for Data and AI at the European Broadcasting Union (EBU), within the Technology and Innovation Department. His work focuses on the development and evaluation of artificial intelligence (AI)-driven solutions for media content analysis, including tagging, enrichment, information retrieval, and data modelling. A Society of Motion Picture and Television Engineers Fellow, he co-chairs a drafting group focused on defining metadata models and registry frameworks for AI systems. Prior to joining the EBU, Alexandre spent over 20 years as an algorithm and systems engineer in Nasdaq-listed companies, contributing to the design of advanced technologies and industry standards. He holds master’s degrees in signal processing (University of Rennes 1) and data science (Télécom Paris) and is the author of 12 international patents and several scientific publications.
Alberto Messina leads RAI’s research and development activities on artificial intelligence systems for content management at the Centre for Research and Technological Innovation. His work spans archive information management, automated production, intelligent systems, digital accessibility, and publishing services. Alberto is the author of more than 100 scientific publications and holds a degree in electronic engineering as well as a PhD in business and management with a specialisation in computer science. A former contract professor at the Polytechnic of Turin, he contributes actively to international standardisation bodies, particularly the EBU and MPEG.
Lubos Steskal is a Strategic AI Adviser at TV 2, Norway’s largest commercial broadcaster and streaming service. He obtained his PhD In computer science from Comenius University in Bratislava, Slovakia. His research interests primarily focus on natural language processing and AI applications in media and journalism. He has co-authored peer-reviewed papers addressing topics such as event detection in Norwegian news, news recommender systems, and computational analysis of online discourse. Lubos serves as the industry leader for the Norwegian language technology work package in the mediafutures research consortium and actively collaborates with the EBU on AI-related topics. His work at TV 2 includes exploring AI-driven solutions for content analysis and generation, semantic search, and interactive AI news anchors. Prior to TV 2, Lubos held positions as a Postdoctoral Researcher at the University of Bergen, where he taught courses on AI, and as a Senior Data Scientist in the financial services industry.
Pierre Fouché is a Project Engineer in Machine Learning within the Technology and Innovation Department at the EBU. He holds an engineering degree in data science and focuses on the research and development of data and AI systems tailored to the needs of the media industry. Pierre’s work involves collaborative projects and proofs of concept to design, experiment with, and evaluate solutions for media applications.
Anthony Dekerckheer is a Lead Data Scientist with over 10 years’ experience specialising in data analysis. After earning his master’s degree in pure mathematics from Brussels University, he pivoted to tech, applying his skills to design and develop back-end services to extract and structure data, thereby building strong foundations with relational databases and computer vision techniques. In 2021 Anthony joined the data team of the French-speaking National Broadcaster of Belgium handling projects in various domains including Enterprise AI and explainability, data visualisation, metadata modelling, AI ethics, and personalised recommendations systems.
