Episode 8: Why Language Models Get Things Wrong

Artificial intelligence has become part of everyday life for many people. Whether at university, in research, or at work, language models such as ChatGPT can provide answers to complex questions within seconds. But those answers are not always correct.

In the latest episode of Welthandelsplatz 1, Karin Heitzmann talks with Axel Polleres, Professor of Data and Knowledge Engineering at WU Vienna, about how modern language models work, why they sometimes hallucinate, and how we can build more reliable AI systems.

Why language models hallucinate

The problem is familiar: When we use AI systems, they invent sources, attribute publications to the wrong authors, or present inaccurate information with remarkable confidence. Axel Polleres put the reliability of AI to the test by asking a language model about his most important research publications. The result? Alongside several correct answers, the model listed papers that were incorrectly attributed to him and even cited a publication that doesn’t exist at all.

The reason why we keep seeing such inaccurate results lies in the way language models work. Rather than storing knowledge like a traditional database, they look for statistical patterns across vast amounts of text. This enables them to generate impressively fluent responses, but it also means they can produce answers that sound convincing but are simply wrong.

How can we make AI more reliable?

Our latest podcast episode also explores how AI systems could become more trustworthy in the future. Polleres explains how researchers are combining language models with knowledge graphs – structured networks of information that can help assess plausibility and reduce errors.

One example is a research project involving the analysis of accident reports from the European rail sector. In this project, language models are used to extract information from texts, while knowledge graphs help verify and classify the results.

Critical thinking and dialogue remain essential

The episode also takes a broader look at the history of artificial intelligence, showing that technological progress is rarely linear. Phases of great optimism have repeatedly been followed by periods of disappointment, so-called AI winters.

One question remains essential for Axel Polleres: Who can we turn to for reliable information? His answer: AI can be a useful tool in many situations, but critical thinking and dialogue with other people remain essential.

Further information

• Axel Polleres: “Was kann Künstliche Intelligenz (noch) (nicht)?” (article in German)

• Sohi, S., Balan, D., Anjomshoaa, A. & Polleres, A. (2025): Towards Harmonised Rail Safety Knowledge: LLM Techniques for EU Accident Report Processing. Available at: https://ceur-ws.org/Vol-4079/short4.pdf

• BilateralAI project