Overview
This 15-minute session explores a central problem of scientific AI use: plausible outputs can be wrong in ways that are difficult for researchers to detect. Using brief examples from literature retrieval, scientific writing, and statistical coding, the session contrasts surface plausibility with diagnostic evidence and highlights findings from recent AI benchmarks and human–AI collaboration research. It introduces the CALIBRATE framework, a risk-adjusted approach for deciding what to delegate, how to define correctness, and when stronger verification is required. A short interactive challenge asks participants to judge a convincing AI output before the underlying failure is revealed. The key message is simple: do not ask only whether an AI output looks right—ask what evidence could prove it wrong.

Target group
Anyone interested in the topic

Lecturer
Dr. Robin Segerer, Information specialist psychology

Register here

All Coffee Lectures Health & Science in the Autumn Semester 2026