Beyond blue links: Meet NEO, EBU’s project to reinvent news access
The European Broadcasting Union developed a generative AI tool that allows journalists and the public to interact directly with a trusted database of millions of articles.
The European Broadcasting Union (EBU) manages an immense flow of information. Its internal aggregator platform called “News Pilot” receives approximately 3,000 articles daily from public service media members across Europe. While this repository is a goldmine for editorial teams, such a big volume makes traditional keyword searching inefficient.
To address this, the EBU developed NEO, a generative AI tool grounded in Retrieval-Augmented Generation (RAG). NEO allows users to ask natural language questions and receive synthesised answers based solely on trusted content from EBU members, effectively turning a static database into an interactive dialogue.
Inspiration: The "chicken" problem
The project was born from an organic internal request shortly after the launch of ChatGPT, when a communications colleague asked if ChatGPT could simply be plugged into the EBU website to help users find information. Sébastien Noir, Deputy Director of the Technology and Innovation Department at the EBU, immediately spotted the risks of a raw integration: “If we just take the API, I could very well come to the EBU website and ask how to cook chicken, which is not exactly our core business,” Noir explains.
The team realised that while the conversational interface was desirable, the knowledge base had to be restricted to their own trusted data rather than the open web’s recipes and trivia.
The roadmap to the solution
The roadmap began with a "hacker" prototype built in two days. Initially, the team considered applying it to the corporate website, but internal hurdles regarding security and legal clearance slowed progress. They pivoted to the News Pilot, which they already controlled and hosted over 4 million articles.
By layering AI over this specific dataset, they solved a genuine user need: summarising cross-border perspectives on complex topics like the war in Gaza or the US elections without manual translation and curation.
The tech stack
The project relies on a sophisticated pipeline rather than a single off-the-shelf model:
News Pilot: The core database contains 4 million articles from EBU members.
EuroVox: An EBU tool used to translate incoming content (text and audio) into English and other languages instantly.
RAG (Retrieval-Augmented Generation): The architecture that restricts the AI to answering based only on retrieved articles.
Dynamic LLM Selection: The system swaps between different Large Language Models depending on the task, after comparing speed and precision.
Alexis Alleman, Data Scientist in Natural Language Processing, emphasises the strict boundaries set for the AI: “We force it to generate the response by extracting the content we gave it and not using the knowledge it acquired during its training.”
The team involved
The project was driven by the EBU's Technology and Innovation department, requiring a mix of data science for fine-tuning the retrieval algorithms and grading article relevance; prompt engineering to ensure the AI interprets ambiguous queries correctly and full-stack development: To integrate the tool into existing CMS workflows and public-facing websites.
The obstacles faced
According to Noir and Alleman, there are many challenges in projects like this. The first one is temporal relevance: standard LLMs struggle with the concept of "latest news". If a user asks about Donald Trump, the AI might provide a biography rather than today's headlines. The team had to build specific filtering steps to grade articles based on freshness when a query implies "news".
On the other side, there is also a challenge in understanding what kind of intention lies in the audience’s queries. Users often type short, unclear queries like "Donald Trump?". The system had to be trained to interpret intent or ask follow-up questions to clarify if the user wanted a biography or breaking news.
Finally, as this project involves translations, they need to be assessed, as nuances are often lost in the process. The team is currently researching how to detect when a translation alters the factual meaning of a source article, ensuring "facts in, facts out."
Opportunities and learnings
Bringing value to the audiences: Originally built for journalists, NEO has been successfully deployed by EBU members for the public. Swedish Radio, SwissInfo, and LSM (Latvia) have integrated versions of NEO on their websites, allowing citizens to query their specific news archives.
Understand the gaps: The tool offers a unique editorial metric: identifying what content is missing. If users keep asking questions that NEO cannot answer due to a lack of articles, the newsroom gets a signal. “With NEO, we can in a certain way measure the audience for content we don't have,” notes Noir.
Regional alliances to build and share technology: Noir envisions a future where European public broadcasters collaborate. If a Swedish user asks about Oktoberfest and Swedish Radio has no content, a federated NEO could seamlessly fetch trusted articles from a German broadcaster, strengthening the public service ecosystem against big tech.
What we can learn from the EBU
Context is king: Simply connecting an LLM to your data isn't enough. You must build a pipeline that understands time and intent to differentiate between an encyclopaedic query and a news query.
Strict RAG for trust: To maintain editorial integrity, the AI must be strictly confined to the provided data. "Facts in, facts out" is the only viable strategy for news organisations to avoid hallucinations.
AI changes user behaviour: Audiences are moving away from "blue links" and towards direct answers. Media organisations must prepare for "Generative Search" where the goal is to surface answers from content, not just drive clicks to pages.
Relive this session, part of JournalismAI Discovery course in French
This case study was produced as part of the 2025 edition of JournalismAI Discovery course in French. Access all sessions recordings here.
