“We Should Have a European AI Agency” Interview with Holger Hoos at IJCAI 2026

Interview: Elena Müller, GI

The Return of Symbolic AI

Elena Müller: Not long ago, rule-based AI systems that follow explicit logical rules, the way older expert systems worked, were considered a dead end. Now there’s renewed interest in hybrid AI systems that combine today’s data-driven learning models with those older rule-based approaches. Is AI coming full circle, or are we entering a completely new phase?

Holger Hoos: It’s a little of both. My original interest in AI was actually on the symbolic side, on logic and reasoning, because I think what brought us to where we are today is a clever combination of mathematics and engineering. Neither field works as well as it does without rigor: in mathematics, proofs need to be correct or they aren’t proofs at all; in engineering, we need confidence that a bridge won’t collapse under a specified load.

In AI, we’ve discovered that a lot is possible with slightly less rigorous approaches. They’re still built on rigorous mathematics, but we often can’t prove under which conditions they perform well and under which conditions things go wrong. It’s encouraging to see that rigor is being put back into the picture. I expected this would happen sooner or later, because in the real world, we need AI systems with provable properties, or at least systems that have been stress-tested sufficiently thoroughly that we can rely on them.

EM: Would you say large language models can ever replace symbolic reasoning?

HH: No, I don’t think they can, though they might emulate it to some extent. I want to be careful here: I don’t believe large language models process information the way the human brain does, though some people do. I think that’s far-fetched, even if there are certain similarities.

Humans can do rigorous reasoning, but as a computer science professor, I know how hard it is for students, even very talented ones, to acquire that ability. Human intelligence isn’t naturally built for it; we have to train ourselves hard to reason correctly, especially when it comes to deeper reasoning. I think large language models are a little like that: we can train them to do certain kinds of reasoning, but it’s inefficient, and it still comes without proper guarantees. So it makes complete sense not to rely on them alone for this.

Where AI Reasoning Still Falls Short

EM: Where would you say today’s AI reasoning systems still fall short?

HH: We have to be careful what we mean by reasoning. With large reasoning models, the shortfall is that they don’t guarantee the reasoning is correct. Yes, they can find correct mathematical proofs, but they can also make serious errors. A correctly implemented theorem prover simply doesn’t do that: it’s correct by construction.

It’s worth remembering that computers were invented partly because human reasoning has weaknesses, and many of those same weaknesses now show up in generative AI models. We don’t see them in proper AI reasoning methods like automated deduction or AI planning, but those methods are hard to work with. Only experts can really use them, and that’s a real disadvantage: needing an AI expert just to use an AI system is a very high barrier to entry. What we ultimately want are AI systems that are more accessible and still capable of correct, deep, and efficient reasoning or planning when it’s needed.

EM: What kind of reasoning is still missing? AI systems are getting better at step-by-step logical reasoning; current challenges might be spotting general patterns or drawing strict logical conclusions. What’s your perspective?

HH: The things that work relatively well typically still don’t come with guarantees. There are ways to trick generative models that are decent at what I’d call shallow reasoning into making mistakes you could never induce in a tool like a SAT solver, because they’re built differently.

A simple example: there are still cases where a cutting-edge model can’t correctly count how many times a certain letter occurs in a word. Five-year-olds can do this, and we’re talking about supposedly superhuman AI failing at it. The reason is that nobody should count using a complicated deep neural network; we already know how to count efficiently. Our brains aren’t built for that, but we can build machinery that counts, sorts, and reasons very efficiently.

Should we throw all of that away just to say we trained a system from scratch? That makes no sense; it’s like trying to produce a genius without books or cultural transmission, purely through interaction with the world. That isn’t how we cultivate human intelligence, and I don’t think it’s how we should approach AI either. The next question we ought to ask is whether we even want AI systems that fully replace our intellectual capabilities, or whether it’s better to build systems with specialized, well-understood capabilities that optimally support us in solving problems.

Toward AI Systems With Genuine World Knowledge

EM: Many researchers argue that future AI systems will need an internal understanding of how the physical world actually works, causes and effects, physics, objects, rather than just predicting the next likely word. Do you think building this kind of internal world model is the next step, and how should it look?

HH: AI systems that overcome today’s limitations need a better understanding of the world, not just physics, but chemistry, biology, and everything we’ve discovered and documented. It helps to look at the history of science. Fields like physics, biology, chemistry, even the social sciences, rely heavily on mathematics, because it’s enormously useful to have compact, understandable models of how parts of the world work.

Even if an AI system could in principle derive all of this from scratch, from observations and data alone, it would be horrifically wasteful. We don’t teach humans that way either; we give them what’s sometimes called textbook knowledge. I think there should be ways of giving AI systems that kind of knowledge, rather than only letting them digest natural language and run a statistical generative process on top, the way current generative AI does. Being able to explicitly introduce knowledge of the world and then use it rigorously would be an important step forward.

EM: Do you have suggestions for how this could be done, how AI could interact more with the real world?

HH: I don’t have the ultimate answer to that, and even a good answer may well win you a Turing Award. But I suspect the way forward is to combine statistical techniques, the kind behind current generative AI, with rigorous symbolic techniques. The real question is how best to do that. Many approaches are being explored, and they all seem reasonable. What I don’t think will work is simply scaling up what we already have and hoping these properties emerge; even if they did, it would be far too wasteful.

There are also many areas – medicine is one, particularly the diagnosis and treatment of rare diseases – where by definition we have very little data. A purely data-driven approach has basically no hope of working well there. Something different is needed there.

AI Augmenting Humans, Not Replacing Them

EM: What’s your own opinion on that: should AI aim to replace humans across the board?

HH: I think it would be a complete disaster if we built AI systems capable of replacing humans in any job. Our societies wouldn’t know how to deal with the economic repercussions. Economic and political power would concentrate dramatically around whoever can provide and scale that technology, and I think it would be the end of democracy. It would also be the end of a world where people work not just to get tasks done, but because work gives them a certain satisfaction.

There are plenty of tasks nobody enjoys, and automating those is a good idea. Given demographics, in Germany for instance, we can see a future where work needs to be more efficient, because there will be fewer people to do it, and AI-based automation is a good approach for dealing with parts of this problem. But believing the end goal should be a machine that replaces humans at everything, including creative work and emotional connection really means to embrace a future in which we are superfluous. This would deprive future generations of the most meaningful things life has to offer.

EM: That’s a strong statement. What would your own vision for AI be, then?

HH: My vision is for AI to stop focusing on systems that replace humans across the board, and instead to look at specific situations where people need and benefit from better support. Imagine a future where hospitals have to work with far fewer doctors and, especially, nurses. What can AI do to help in that particular situation? That’s a much easier and more useful pursuit than building AI meant to fully replace a nurse or a doctor.

Global AI Trends and the Role of Middle Powers

EM: Given those two visions, where is the global trend actually heading?

HH: As far as those currently leading development are concerned – mostly large American, multinational but American-controlled companies – the trend is toward AI that replaces human capabilities outright. The people running these companies are quite open about it, and I think we should reject that.

Europe has an important role to play here, together with what the Canadian prime minister Mark Carney, at Davos earlier this year, called other “middle powers.” We share a common concern: being left behind and becoming dependent on technology that isn’t the kind we want or that serves us well.

It’s a bit controversial to say, but I think China has a role to play too. It’s interesting that China is currently giving some American companies a run for their money, showing that with fewer resources, results can be just as good. I believe there are plenty of people in the United States, in China, and everywhere else who prefer a vision of AI that supports people rather than replacing them. We should connect those people and interests, and make a stand to push AI in a direction that actually benefits us.

The Role of IJCAI and ECAI

EM: Looking back at the IJCAI-ECAI conference that just happened: how did it go for you, and where do these conversations fit into IJCAI’s or ECAI’s role?

HH: The IJCAI and ECAI communities have always leaned toward a less extreme vision of AI than replacing humans outright. The researchers most focused on pursuing that vision are often absent from these conferences altogether, because they work for companies that don’t share their results, or you find them at venues like NeurIPS, ICML, and ICLR. Not everyone there pushes in that direction, but you find more of that viewpoint there.

That’s one advantage of IJCAI-ECAI: a more balanced view of AI, and broader coverage. Those other conferences are prime venues, but very focused on machine learning. Machine learning is hugely important, but there’s more to AI than that. Here, we get the broader picture, including a focus on connecting different AI techniques, more than at those other conferences. In parts of the machine learning community, people sometimes spend too little time understanding the full breadth of AI, and end up reinventing wheels, sometimes worse ones than those that already exist. Venues like IJCAI and ECAI because they give a broader view of the field, particularly for young researchers, and create an environment where combining different areas of AI is normal rather than unusual.

A Proposal for European AI Governance

EM: Is there anything else you’d like to share, any thought you’d like to add before we close?

HH: Thank you for the opportunity. A question people often ask at this point is what the middle powers, including Europe, can actually do to succeed with their vision, to have global economic success and impact.

I have a clear and simple idea. We need to stop pretending that politicians and bureaucrats can, in a fine-grained way, develop and oversee AI research programs. Honestly, that’s a nonsensical idea. If they tried it in physics, they’d be laughed off the stage; our colleagues in physics know perfectly well it wouldn’t work.

Right now, the European Commission, the governments of Germany and other EU member states, the British government, and others are trying to do this from within ministries or directorates-general. It simply doesn’t work. What would work is doing it the way physics and space research do it: build an expert-driven organization, like CERN or ESA. Give it a mandate and resources, hold it accountable for good work, but let scientists and experts run it.

It works wonderfully for ESA: Europe has capabilities so important that NASA relies on them constantly. The James Webb Telescope, arguably the most advanced telescope in space, was launched by a European rocket, because we have a great European Space Agency. We should have a great European AI agency the same way. As with ESA and CERN, we could let non-European countries join too. If we did this, I’m certain that within five years we’d have wonderful results, and within ten years we’d be where we should be: able to compete globally with the United States and China in AI, and to build the AI systems our societies actually want and need.

EM: Thank you so much for this conversation.

HH: You’re very welcome.

Transcribed and revised with Otter.ai