Just ahead of December’s second annual EBU AI Summit, the Technology & Innovation department took the opportunity to hold a workshop dedicated to explore the extent to which EBU Members could build these tools themselves. The psmGPT workshop featured contributions from Austria, Belgium, Germany, France, Slovenia, Sweden and Switzerland, providing a rich tapestry of ideas and inspiration.
The work presented fell into three major categories. The first was the use of LLMs – the large language models that power generative AI tools – to assist with content creation. Examples here included the Mediagen platform that France Télévisions has developed for all employees and Fenix AI, which provides a chatbot to assist journalists at RTV Slovenia. AltoMagic at Svergies Radio does something similar, but with integrated tools.
The second category was that of platforms and integration. Here, examples came from Germany’s Deutsche Welle and SWR, with the former putting in place a framework that will allow the gradual addition of more AI tools while the latter is moving towards “agentic” processing, where the tools should be able to perform actions, rather than just providing answers.
The third category of work presented is the use of LLMs for audience engagement. This is where RAG – retrieval-augmented generation – systems provide more user-friendly ways for audiences to engage with the huge volumes of content held by EBU Members. Examples included ORF’s AI Search and the EBU’s own Neo project. Also noteworthy was the Study Buddy tool from ARD, being a RAG focused on education.
The workshop identified three top concerns for EBU Members as they embrace generative AI tools: balancing innovation with credibility, having appropriate security measures, and ensuring processes are scalable.
The presentations from the psmGPT workshop are available to EBU Members.
Who's doing what?
EBU Technical Report 083, published in January 2025, explores the use and evaluation of LLMs in media organizations, focusing on applications, evaluation methods and integration challenges. It also addresses the issue of scalability and suggests cost-effective strategies for certain tasks such as classification or tagging.
This article first appeared in the March 2025 issue of tech-i magazine.