288 master’s students, 130 doctoral students, over 400 scientific publications, 22 awards, three start-ups and 45 industry partners – these are the impressive achievements of the DAAD-sponsored Konrad Zuse Schools after three years of funding. At the third annual meeting of the graduate schools in late October, which was held at the Zuse School ELIZA in Darmstadt, one of the questions discussed was the extent to which human thinking is already embedded in AI systems and how this can benefit humans rather than replacing them.

In the hall outside TU Darmstadt’s Intelligent Autonomous Systems (IAS) Lab a group of people are watching as a machine does its best not to slip on the smooth floor. The four-legged robot is being controlled by the kind of console we are all familiar with from computer games. The faster the joystick is moved, the more hectically the machine tries to keep its balance. It taught itself how to walk – by a process known as reinforcement learning. “It learns how best to move by trial and error and by a system of rewards and punishments,” explains IAS doctoral student Nico Bohlinger. This is made possible by the rapid advances in AI research.

 Research and ethics must go hand in hand

But how can pure mathematics lead to behaviour that took millennia for evolution to master? And how close has AI already come to what we call thinking? This was one of the key questions addressed at the third annual meeting of the Konrad Zuse Schools of Excellence in AI, which was held at the Zuse School ELIZA at TU Darmstadt at the end of October. In his opening speech, Professor Matthias Oechsner, vice president for research at TU Darmstadt, already warned that “AI is not a neutral tool,” explaining that it replicates social prejudices and must not be allowed to serve economic interests alone. He stressed that deep tech requires deep thinking – and appealed for research and ethics to go hand in hand.

The Zuse Schools have accepted this challenge, according to Dr Michael Harms, Deputy Secretary General of the DAAD. He describes them as places where excellence meets responsibility and where interdisciplinary and intercultural dialogue is cultivated. At their heart are people who develop AI, he says – young talents from around the world. “The future of AI will be shaped by your questions, your visions and your values,” emphasised Harms, adding that they all were playing their part in designing a future based around “AI made in Europe”, a future in which intelligent systems reflect democratic and European ethical values.

What can AI learn from humans?

Iryna Gurevych (professor of Ubiquitous Knowledge Processing at TU Darmstadt’s Department of Computer Science), Letiția Pârcălăbescu (AI researcher with Aleph Alpha), Marc Toussaint (professor of Intelligent Systems at TU Berlin) and Volker Tresp (professor of Database Systems and Data Mining at LMU Munich) discussed how close AI has come to thinking like a human. First, however, they needed to clarify what we actually mean when we talk about “thinking”. According to Toussaint, thinking in the world of AI generally means weighing up different alternatives, whereas human thinking also involves imagination and intuition. At the same time, intuition – in the sense of being able to draw conclusions rapidly and spontaneously – is essentially something that AI systems are better at than humans. “What we describe as our gut feeling is based to some extent on data we have gathered throughout our lives, though also on data that we as living beings have collected during the course of evolution,” explained Toussaint. “Machine learning draws on this data more directly and translates it much more quickly into decision-making rules.”

However good machines may be at intuitive decision-making processes, they are often fairly poor when it comes to what we call common sense. Letiția Pârcălăbescu offered an explanation as to why this is the case: generative AI systems are trained on texts, yet there are many things that we take so much for granted that we do not document them in text form. “It would make no sense whatsoever for us to write down that grass is green. Consequently, such statements are massively underrepresented in the training data. And that’s precisely why the systems are so bad with regard to such knowledge,” said Pârcălăbescu, a researcher who works for the German AI company Aleph Alpha.

Being dependent on datasets that are often deficient is what makes AI prone to drawing incorrect conclusions. As Iryna Gurevych stressed, however, humans aren’t much better in this respect, as they also live in filter bubbles of their own experiences and biographies. Yet together they could compensate for each other’s mistakes. Gurevych urged that AI should be used as a thinking partner that offers new perspectives to critically engage with – in the spirit of true human-AI collaboration.  

AI in climate research

Professor Markus Reichstein from the Max Planck Institute for Biogeochemistry in Jena is regarded as a thought leader at the interface between climate and AI. He highlighted in his keynote how machine learning can help us better understand the complex interactions on our planet.

Satellite images, sensor networks and citizen science yield huge quantities of data. AI can analyse it and reveal patterns that have previously remained invisible. Reichstein’s team is working with the Red Cross for example on early warning systems that predict droughts or floods and can direct humanitarian aid in a more targeted manner. He believes it is essential to follow a hybrid approach – machine learning combined with physical knowledge. This gives rise to models that are both explainable and robust. Data- and hypothesis-driven research shouldn’t be polar opposites. Reichstein believes that AI isn’t an end in itself but a tool for uncovering interconnections – thereby helping humans to safeguard the future of the planet.

Multimodal AI in medicine

Combining many different data sources could also revolutionise medical diagnostics. Valentine Idakwo, a doctoral student at the Zuse School relAI at LMU Munich, is studying how multimodal machine learning models could help detect heart valve diseases at an early stage. While doctors draw on many sources of information – laboratory results, ultrasound, smartwatch data – many AI systems have so far used just one. A multimodal approach involving different types of data could be the solution, allowing more precise diagnoses to be made earlier.

Using AI to predict thoughts

How do our brains react when we watch a film – and how can this activity be predicted with the aid of artificial intelligence? This is precisely what Deniz Kucukahmetler, a doctoral student at Leipzig University’s Zuse School SECAI, has been exploring. Her team developed a multimodal deep learning model that can predict how the brain will react to audiovisual stimuli, and has concluded that AI these days can predict cerebral processes amazingly well but works quite differently to the brain itself. In the long term, such models could help us better understand neural dynamics – for example to diagnose epilepsy at an early stage or study emotional responses.

When AI speaks like a human

Florian Kandra, a master’s student at the Zuse School ELIZA at Saarland University, is investigating whether large language models behave similarly to humans when engaging in dialogue – that is, whether they also adapt their style to that of their conversation partner. In his experiment he instructed two AI agents to chat to one another. What he found was that the sentence structure of the two systems became more and more alike the longer the conversation continued. According to Kandra, this ability of AI to adapt its language use could make future human-machine communication more natural – and show that AI doesn’t imitate only content but also social patterns.

AI in industry

One major strength of the Zuse Schools network is their contact to industry. Dr Claudius Gläser from Bosch Research revealed just how huge the demand for excellent research is in industry. In driverless vehicles, AI is increasingly replacing rule-based systems. Modern neural networks process all sensor data – supplied by everything from cameras to lidar instruments – simultaneously and then immediately take driver decisions.

The sector is also experimenting with large language and multimodal models capable of interpreting specific traffic situations. Gläser envisages two main fields of research for the future: simulation and scaling. Digital twins and reinforcement learning allow vehicles to drive millions of virtual kilometres and hone their skills in the process. He says this is crucial if they are to overcome the legal and safety challenges that still persist. “Perhaps most cars will be based on an AI model in future – but only if we can understand, test and trust it.”

Ultimately, all the participants agreed: given that artificial intelligence is already having such a massive impact on all areas of our lives, we humans must keep pace with its development. Training young researchers is one vital key to achieving this. “The future of AI will not be written only in code but also in the questions, visions and values of young researchers,” is how DAAD Deputy Secretary General Michael Harms summed it up.

Klaus Lüber (12 November 2025)