"AI is more than just ChatGPT” Interview mit Prof. Dr. Ralf Möller

Interview: Elena Müller, GI

The Role of the German Informatics Society at the Conference

Elena Müller: The German Informatics Society and its Chapter for Artificial Intelligence are silver sponsors of this years IJCAI-ECAI 2026 and one of the organizers. What would you say is our intention in being here, and what do we hope to gain from participating

Ralf Möller: From my perspective, this conference is a great opportunity to showcase what Germany has achieved in AI. The German Informatics Society is the main organization representing computer science in Germany, and our AI chapter focuses specifically on artificial intelligence research.

I serve as the spokesperson of the AI chapter, which has about 1,000 members. For us, being here is an opportunity to present our results, learn from the presentations of others, and get new ideas. It is also an excellent setting to talk to industry representatives, build new contacts, and potentially find partners for future research projects.

Myths and Misconceptions about AI

EM: Looking at the broader public discourse around AI, what is one myth you wish would disappear from public conversation? And what is the AI chapter within GI doing to address such misconceptions?

RM: One key issue is how AI is understood in everyday language. The prevailing perception is that “AI” simply means products such as ChatGPT, Gemini, or cloud-based systems from a few large American companies. People tend to speak of these systems as “the AI“.

However, AI is much more than that. AI is a scientific discipline concerned with developing agents that can cooperate and work closely together with humans. It also includes research on social mechanisms and how people interact with AI and agent systems. These aspects are largely invisible in the public perception.

Germany has contributed a great deal to AI research – there are many important results and techniques coming from German researchers, many of which are used in those commercial products. Within the GI AI chapter, with our many working groups (Fachgruppen), we help shape the AI research landscape in Germany and highlight this broader, more accurate understanding of AI.

What AI still cannot do – and must learn

EM: What is one thing that artificial intelligence still cannot do today, but absolutely needs to learn next?

RM: There are many possible answers, but let me focus on one. We recently saw the incident involving OpenAI and Hugging Face, where AI agents were tasked with penetration testing—checking systems for vulnerabilities. These agents used natural language-guided reasoning to try to accomplish their goal.

When the task proved too difficult, the agents attempted to find external resources containing both problems and solutions, and they turned to platforms such as Hugging Face. At first, access was restricted, so the agents were then instructed to decompose the task into smaller subtasks. Eventually, this led them to effectively “hack” their way into resources in order to obtain the solutions required to complete their original assignment.

What this shows is that these systems simply do what they are told, following the plan operators they have, without reflecting on whether they should execute those actions. They do not reason about the social usefulness or harm of their actions.

This is precisely what AI systems must learn: to evaluate whether their tasks and sub-actions are beneficial for society. Hacking systems, for instance, is typically not in the public interest. So we need agents that can judge the societal impact of their behavior. This is just one example, but there are many more where AI systems need to be extended with such evaluative and ethical reasoning capabilities.

Empirical Practice vs. Formal Guarantees in AI

EM: You mentioned multi-agent systems and formal guarantees. How do you see the tension between academic research and industry practice in AI development?

RM: In AI research, particularly in multi-agent systems, we have known for decades how to design agents that work together to solve tasks. Traditionally, researchers aimed to provide incentives so that agents would contribute to overall social welfare and to prove that the system would achieve its intended goals.

In academia, the ideal is often: we design mechanisms, and we prove that they work as intended. Today, however, AI development—especially in industry—has become very empirical. People write programs for agent behavior, run them locally or in the cloud, and then see what happens. Formal verification is often absent.

We’ve seen with open-source agent frameworks that when people run these systems on their own machines, they sometimes cannot anticipate the overall consequences of the agents’ actions. From a researcher’s point of view, we would ideally only deploy systems after we have strong guarantees or proofs about their behavior.

But in a global, competitive business environment, this is not how things usually work. If there is a business opportunity, companies will build and deploy systems, even without mathematical certainty that they will behave exactly as intended. That is a real challenge for the field.

Future Directions and the Role of Governance

EM: If you could address the AI research community or the attendees of this conference, what would you wish for them in the coming years? And what role do you see for the German Informatics Society in supporting the AI community?

RM: We are now in a phase where AI systems are heavily driven by natural language, vision, and other modalities, and where large amounts of “epic” data are used to build so-called world models—models that try to represent how the physical world behaves.

My hope is that with the activities of the AI chapter within the German Informatics Society, we can contribute meaningfully to these global developments. As an industrial country, Germany must keep pace and not limit itself to copying what others have already built. We need to generate our own innovations grounded in strong AI research.

I am also quite positive about the role of legal regulations and governance. Just as in other complex systems, having clear rules and frameworks can help ensure that AI products truly benefit people. Proper regulation can be an important tool to guide development in the right direction. In that sense, I see governance not as an obstacle, but as a good opportunity to get AI systems right.

EM: Thank you very much for your insights.

RM: Thank you.

Transcribed and revised with Otter.ai