Machine learning models have limitations. That is why it is important to critically question the results of applications based on large language models (LLMs) such as ChatGPT instead of trusting them blindly. In the first funding phase of the project, a common language for trust and distrust was developed. A study also showed that simple warnings (disclaimers) are not sufficient to promote a healthy level of skepticism. Therefore, the researchers from psychology and computer science developed new approaches to explain model uncertainties in an understandable way. Building on this, the project team is now developing targeted measures to promote healthy distrust, especially when LLMs are used as support in scientific writing. New forms of explanations, dubbed “perplexing explanations,” play a central role in this. These explanations are intended to show the unreliability of LLMs to students and will be used as an automatic method in the TRR 318 to strengthen human autonomy in dealing with AI systems.

Research areas: Computer science, Psychology