Reading Between the Lines: A Linguistic Perspective on Fake News and AI

Digital disinformation and AI-generated content are transforming the way we perceive and evaluate information. In cooperation with HKA – Karlsruhe University of Applied Sciences and supported by the ZKM as an educational partner, the symposium “The Fabrication of Truth” provides a platform to highlight current challenges and explore new perspectives.

The symposium is part of the project “Reality Check: Fake or Fact in the Age of Generative AI”, initiated by HKA’s Institute for Intelligent Interaction and Immersive Experience and funded by the Carl Zeiss Foundation. Its aim is to raise awareness—especially among young people—of the challenges posed by fake news, deepfakes, and digital manipulation.

The lecture outlines a linguistic perspective on the limits of current AI systems in handling fake news. The focus is on implicit assumptions that reflect the shared background knowledge between interlocutors—so-called presuppositions. These make communication efficient, but they become problematic when the presupposed content is false. False presuppositions are particularly effective at embedding misinformation unobtrusively: the false information is not stated explicitly but is instead taken for granted as true. Questions such as “Have voters found out that the Left Party does not want the EU to establish its own sea rescue operations in the Mediterranean?” suggest an incorrect state of affairs and risk silently establishing false information within the shared knowledge base (“common ground”).
Questions containing false presuppositions pose a communicative challenge not only for humans but also for large language models such as ChatGPT, which produce responses that often appear coherent and plausible at first glance. However, when language models answer such questions without scrutinizing the assumptions they implicitly contain, they subtly contribute to the spread of misinformation.
The lecture presents results from a recent experiment in which various language models were deliberately tested with false presuppositions about political facts. The analysis shows that factual knowledge alone is not sufficient to identify and correct implicit misleading content.