On June 15, 2026, Kai Eckert returns to the Zentrum für Lehre und Lernen (ZLL) at Mannheim Technical University with a follow-up talk in the lecture series on AI-based tools in teaching, three years after the first installment.
Back in 2023, the original talk in this series — “Chancen, Grenzen und Risiken durch ChatGPT und andere KI-basierte Tools für die Hochschullehre” — offered a handful of theses, a lot of speculation, and very little empirical grounding. This second installment, “Drei Jahre später…”, revisits those theses against the backdrop of three years of research and puts them to the test.
The talk works through several strands of the discussion:
- Standortbestimmung — AI use among students has gone from a niche behavior to the norm, with usage rates climbing from roughly 63% in 2023 to nearly 92% in 2025 across German higher education.
- From overestimating AI to overestimating ourselves — the original worry about blind trust in plausible-sounding AI output has gained a new layer: fluent, fast answers can create an illusion of competence that outpaces actual understanding.
- Detection, plagiarism, and institutional grey zones — AI-text detectors remain unreliable, unintentional plagiarism has expanded to include hallucinated facts and invented citations, and many institutions still lack binding examination rules to match their updated honesty declarations.
- Effects on learning itself — newer findings on cognitive offloading, metacognitive laziness, and a “competence paradox” suggest that AI tends to help most those who need it least, while tools designed to scaffold thinking (rather than hand over answers) can offset some of the negative effects.
- Practical guidance, updated — for students: understand the technology, document AI use transparently, and stay able to explain your own work, not just present it; for teaching staff: rethink assignments and exam formats so that what is trivially solved by AI is no longer what gets graded.
A Note on Teaching and Assessment
A recurring theme of the talk is that detection is the wrong lever — what matters is redesigning tasks and exams so that the sense of doing them remains visible, even when AI could produce the end product on its own. The talk closes with open questions for the audience: which assignments have become trivial with AI, whether we should be assessing products or processes, and whether grades themselves still serve their purpose in an AI-saturated learning environment.
Update (2026-06-17): Here is the link to the news article at ZLL.