Large Language Models (LLMs) are now so good at creating texts that they give the impression of actually “understanding” language. One of the things that makes them so powerful is the so-called modeling of semantic similarities. Iryna Gurevych, Professor at the Department of Computer Science at the Technical University of Darmstadt and Academic Fellow at the Konrad Zuse School of Excellence in Learning and Intelligent Systems (ELIZA), is working on this. In this interview, she explains what AI can already do, where there is still a need for adaptation and why ethical issues must be taken into account. 

Professor Gurevych, you head the Ubiquitous Knowledge Processing Lab (UKP) at TU Darmstadt. What fascinates you about automatic language processing and what key issues are currently driving your research?

Even at a young age, I was fascinated by the idea that a machine could understand language like a human and that this ability could be deliberately programmed. I also found it interesting to see, how machines support the education sector. For example, teachers could be spared some of their workload by solving tasks on the computer and having the solutions checked automatically. The underlying machine intelligence still fascinates me today, even if the methods of automatic language processing have changed completely.

In what way?

While in the past we had the idea of representing all language rules in a machine-readable way and summarizing the processing by rules as in a formal language, today's large language models (LLMs) are trained on very large amounts of data and are extremely powerful. We investigate how LLMs work and try to explain them. We are also researching how these models can be designed in such a way that they are secure and compatible with social values. In the medium term, LLMs will become a ubiquitous tool and everyday companion for all of us.

A central aspect of your work is the modeling of semantic similarities between words, sentences and documents. Why is this topic so fundamental to the understanding of natural language by machines?

Semantic similarity means how similar two linguistic units are in their meaning - even if they are formulated differently. Before a language model responds, a semantic search is first done to look for texts that match the content of the question. This provides the LLM with additional important context in order to answer users' queries precisely and with up-to-date information. This is also known as retrieval augmented generation. Semantic similarity also plays a central role in summarizing texts: with the help of such methods, texts can be semantically clustered so that paraphrases are combined in a cluster. In this way, the machine can structure the content even better in order to create a final overall summary.  

In one of your studies, you highlighted the limitations of AI language models such as ChatGPT, particularly with regard to their ability to think independently and complexly. Their tendency towards so-called “hallucinations” remains a serious problem. What consequences does this have for the future development and use of such models?

We cannot yet place too much trust in the results of LLMs and as users, we should routinely check their results. Of course, we are also doing research to make the outputs of LLMs more trustworthy. Nevertheless, there is a certain tension between creativity and reliability. After all, in the context of some tasks, we want the language models to be creative, i.e. to invent new things deliberately. The good news is that the behaviour of large language models can already be explained to a certain extent—we know this from our research. If we understand how the models learn, we can also control their behaviour and make them more reliable. 

You emphasize the importance of responsible use of AI technologies. What ethical challenges do you currently see in AI research, and how should researchers and society respond to them?

Ethics is an important topic for AI. At the moment, the training of AI is still far too little designed according to ethical aspects—everything that is available on the internet is simply harvested. In general, the use of LLMs should be closely monitored, as they are still a long way from being safe. When children communicate with large language models on their mobile devices, the effects on a child's psyche are neither understood nor researched. This requires great social attention and a close exchange between research, technology providers and society.

As a member of the Zuse School ELIZA, you are involved in the training of future AI experts. What goals does ELIZA pursue and how does the Graduate School contribute to the international visibility of German AI research?

ELIZA creates a great framework for training the next generation of AI experts in an international environment. Through research grants, support for events and international exchange, we create access to expertise in other European countries. This is urgently needed in order to keep up internationally. And it will pay off: through top-class publications in leading specialist forums and through funded individuals who will take up key positions in AI in the future.

Your work has a strong interdisciplinary focus and includes collaborations with international partners. What role do such collaborations play in your research, and how do students and junior researchers benefit from them?

Modern AI research is highly complex and is developing at a rapid pace. It is therefore no longer possible to be an expert on AI as a lone fighter. To be successful, you need alliances that make it possible to tackle major research challenges in larger and more diverse teams. It requires a close exchange and a fresh view from the outside. This is exactly what we achieve with our international partners, who complement us in areas in which we are not experts. This applies in particular to the application areas of AI, such as psychology, ethics or medicine, but also other sub-areas of AI itself, such as robotics or machine vision.

Which current developments and trends in AI research do you consider to be particularly promising, and in which areas do you see the greatest potential for social applications? 

AI will most certainly continue to find its way into the world of work and education, which will fundamentally change many professions. A lot will also change in my job as a professor. Many activities will be outsourced to my personal AI assistant in future, such as creating slides for lectures or assessing students' written submissions. Nevertheless, I don't think my job is at risk. Because all of this will require close collaboration between humans and machines, possibly in a larger team, and will help us to achieve a better quality result that will still always bear the personal signature of an expert.

Interview: Klaus Lüber (5 June 2025)