Matilde Barbini

Where Editorial Authority Goes When AI Enters the Newsroom

Matilde Barbini of EPFL studies how AI can be aligned with the values of journalists and readers.

Artisans of Words
Journalists value attribution and verification and care deeply about their own style. One local reporter described AI search summaries using his work while readers no longer visit his site: "My words are being used to inform people, but I'm not making any profit of it."

A Feedback Loop of Deference
Without a newsroom policy, editorial authority migrates towards models and platforms. The more a model sounds like the newsroom's voice, the more journalists defer to it.

Alignment Is Not a One-Off
For Matilde, alignment is not a property optimised at a given moment and decided by whoever built the model. "It's something that should be undergoing continuous challenges and discussions" with the people who actually use the system in a specific context.

Personalisation: Benefit and Risk
Research shows that immigrants in Switzerland benefit from news adapted to their language or cultural background, which helps them understand their new country. Framing or changing the news itself according to who you are is a different matter. Navigating that trade-off remains an open question.

Questioning the Benchmarks
Many benchmarks come from the same industry that builds and trains the models. Academia, she observes, is shifting towards open models such as Apertus because they allow research into explainability rather than a race for performance.

Who Holds the Power?
Her closing question: who decides what a value is, how values are prioritised, who shapes AI, and who is affected by it?

Alignment is a continuous process, and the people affected should help define the values and guidelines.



The research focuses on understanding the values and emerging practices surrounding the use of large language models (LLMs) in news production. Using mixed-methods, the study targets two main groups: 1. News creators – examining values such as transparency, efficiency, and explainability in LLM-generated news. 2. News audiences – exploring expectations around personalization, algorithmic choices, and value alignment with content creators. The research will include diverse socio-economic communities and address themes such as local vs. wider-scope news and sensitive content.