I am a full professor of Human-Centred Transformative AI at the Hasso Plattner Institute (HPI) in Potsdam and a Research Consultant at Google DeepMind.
My lab works on the foundations and safety of frontier AI, with a focus on multi-agent systems that operate in the real world. People increasingly work alongside AI agents, and groups of agents are starting to carry out tasks autonomously in high-stakes domains. We develop methods that make these systems safe and dependable. Our research areas are:
- Multi-agent systems, including self-evolving systems, and the environments needed to train and evaluate them
- Reinforcement learning, post-training and reasoning
- Monitorability, interpretability and early error detection
- Privacy-preserving and trustworthy machine learning (methods and theory)
- Data attribution, memorisation and unlearning
- Continual learning and test-time adaptation
- Training dynamics and learning theory
How we work. We are a small, tight-knit research team. Our aim is for open-source software, models and reusable research results to become our main output. We still publish papers when a result warrants it, but conference deadlines play a shrinking role in how we plan our work. We are fortunate not to depend on third-party funding, so we can choose problems by how much they matter. I do research myself every day and work directly with everyone in the lab, which is why I take on very few students. My teaching is equally hands-on: we build intuition and understanding by working through problems and discussing them together, then take our results to AI agents, develop them further and bring them back to the group.
Our name. The name of our lab, Human-Centred Transformative AI Lab (HTAIL), reflects two convictions: frontier AI is transformative in both its promise and its risks, and it must be developed with a human-centred approach, as defined by Ben Shneiderman. Despite the name, we are not experts in human-centred design or human–computer interaction.
What we do not work on. I trained as a physician and radiologist before moving to AI full-time, and earlier in my career I worked extensively on medical AI and federated learning. I no longer initiate research in medical AI, medical data science or federated learning, and I do not supervise students or theses on these topics. I occasionally contribute to medical AI projects led by collaborators. I am also not a regulatory, policy, legal or copyright expert or an ethicist, and I do not comment on these issues.
Working with us. Open positions in my lab are advertised exclusively on the HPI Job Portal. Neither my office nor I respond to unsolicited applications or enquiries by email, without exception.
My name. My legal name, which I use in publications, is Georgios. I go by George in English and Georg in German. If you are unsure how to address me, use George.