Key Notes
of the URAI Symposium 2026

We are proud to announce the keynote speeches that will open the URAI Symposium 2026:

Prof. Dr. Toni Steimle, FHNW (CH): “Are LLMs More Human Than We Would Like?
How AI Systems Decide Whom to Believe – and What this Means for Human Oversight”

When a clinician overrides a diagnostic assistant or an operator clears an autonomous system’s plan, we like to assume the machine simply accepts the human’s input. It does not. Large language models inherit human cognitive biases from their training data: they can be anchored, flattered, and swayed by authority or staged consensus. The keynote opens with a review of the current state of research on cognitive bias in LLMs — what has been documented so far, how it is measured, and where the evidence remains thin. It then argues that these seemingly separate failures — sycophancy, authority bias, conformity — are facets of a single underlying mechanism: the model weighs every expressed preference by its credibility, judging who is speaking and how well their confidence fits the facts, much as people do. That behavior is usually sensible — and precisely therefore exploitable. The talk closes with findings from our own series of controlled experiments across model sizes, and traces what they mean for designing human oversight: a human in the loop is not a neutral check, but a second judgment that can be moved socially.

Prof. Dr. Maja Temerinac-Ott, Hochschule Furtwangen (DE): “Human in the Loop: When Humans Become Part of the Algorithm”

What happens when the best AI systems are not fully autonomous—but deliberately designed to work with humans? In this keynote, I will explore Human-in-the-Loop as a design principle for intelligent systems, where human knowledge, feedback, and decisions become an integral part of the learning and inference process. Rather than treating humans as a final safety check, we can make them an active component of the algorithm itself. Through technical examples from Computer Vision, Robotics, and Large Language Models, I will show how human feedback can improve models, guide decisions, and enable systems to operate effectively in situations where data, uncertainty, or context cannot be captured by automation alone. I will also connect these ideas to experimental design and Active Learning, where the central question becomes: Which information should we ask a human for next? The examples will illustrate both the opportunities and the challenges of Human-in-the-Loop systems—from selecting the right data and learning from human corrections to deciding when an AI system should defer to a human. Across these different applications, a common theme emerges: the goal is not to replace human intelligence, but to design algorithms that know how and when to use it. The keynote will conclude with a broader perspective on what it means to build AI systems in which humans are not merely users of the technology, but part of the computational loop.