Title
How aligned language models break
Abstract
Alignment training of modern LLMs has made them much harder to break and less likely to engage in harmful behavior, but it hasn't made them perfectly safe - misalignment remains, and new failure modes emerge as models become more intelligent. This talk covers a set of these failure modes, specifically the ones I've discovered in my own work. I'll shed light on how the idea of doing machine psychology led me from deception in LLMs to a string of other safety-critical issues: emergent misalignment, evaluation and self-awareness, shutdown resistance, and more. I'll show what these failures look like, and what they tell us about the models we're building.
Bio
Thilo Hagendorff is an expert in AI safety, AI ethics, as well as machine behavior in generative models. He leads the AI Safety Research Group at the University of Stuttgart. Previously, he worked for the Cluster of Excellence “Machine Learning” at the University of Tuebingen. He was a visiting scholar at Stanford University, UC San Diego, as well as the European Laboratory for Learning and Intelligent Systems (ELLIS) in Alicante.
Hagendorff is a visiting researcher at the University of Oslo and Integreat during August and September 2026.
Speaker is presenting in person in Oslo.
Practical
The Tuesdays seminars series are devoted to various topics relevant for the Integreat´s research focus. Presenters from the Integreat community and beyond have 40 minutes to present, followed by a group discussion.
Seminars are open for attendance for everybody.
For those unable to attend in person: https://uio.zoom.us/j/65632027408
Point of contacts:
- Organiser and moderator: Arnoldo Frigessi, University of Oslo
- Integreat Administrative Leader Maria Dikova, University of Oslo