Menna El-Assady from ETH Zurich is working on approaches that promote mutual understanding between AI and humans. The aim is to ensure better interaction.
When people use AI for their work today, do they actually understand what they’re doing?
Menna El-Assady: This is definitely a problematic area. Currently, many people trust that the answers given by AI are correct – without knowing how these answers come about in the first place. We therefore need to develop an understanding of what data and what basic assumptions are contained in a model. Only then can we interpret the answers we receive correctly. The research area that deals with explaining the mechanisms of AI models to users is called explainable artificial intelligence (XAI). So we’re trying to understand how AI works and develop a kind of instruction manual for working with AI on this basis.
Prof. Dr. Mennatallah El-Assady is the holder of the endowed professorship funded by the Dieter Schwarz Foundation at the Department of Computer Science at ETH Zurich. There, she heads the Interactive Visualization and Intelligence Augmentation Lab (IVIA Lab), which is also funded by the Dieter Schwarz Foundation. Born in Egypt, she works at the intersection between data analysis, visualization, computational linguistics and explainable artificial intelligence, and her research focuses on the co-adaptive development of humans and AI.
How does this instruction manual help?
El-Assady: The main thing is that we use AI for tasks for which it is actually suitable. Its strength lies in analyzing and evaluating immense amounts of data. I can thus use AI as a filter that processes and weights information. So before I flood my brain with information on a particular subject, and perhaps overwhelm it, I can let the AI sort this information. However, we must not neglect critical thinking when it comes to making informed decisions. It is important that we maintain a healthy degree of skepticism towards AI output. This requires that we understand something about the area in which we are consulting AI.
The more convincing the results, the more frequently AI is used. How do you ensure that these results are relevant to the respective task?
El-Assady: This is definitely very challenging, because people in different professions also have different requirements for artificial intelligence. We are currently developing AI agents that are trained to help with specific work tasks, such as collecting, structuring and summarizing analyses. Crucially, the AI then makes suggestions rather than pronouncements.
Because AI doesn’t yet really understand how humans process information?
El-Assady: Exactly. That’s what we’re trying to address. Some language models, for example, don’t understand how people make decisions. And they have no grasp of logic. Whether I ask where Bern is or where Bern is not, the most likely answer is: “In Switzerland.” This lack of understanding presents us with a huge task that will occupy us for years to come. But we can now detect when models give logically inconsistent answers. Communicating these to users in a meaningful way is then the task of the systems that we develop.
Wouldn’t it be better if AI could follow human logic?
El-Assady: It’s a kind of mutual learning. In the next step, we want to teach AI models how humans would make decisions. Through human feedback, they will learn whether they have correctly understood and implemented tasks and questions. Meanwhile, we are consulting with experts from various disciplines to establish how best to formulate specific questions for AI in order to obtain valid results. We are working to enable companies – or individual people – to adapt AI models to their needs. A key question is what is expected of the AI: comprehensive answers or rather suggestions and ideas.
When does AI provide comprehensive answers, and when does it offer suggestions?
El-Assady: In our guidance models, we distinguish between orienting, directing and prescribing. With orienting, the AI offers various options without stating what is right or what is wrong. Directing uses rankings based on probability, while with prescribing, the AI provides exactly one answer.
You are placing people at the center. What could be the next steps for better collaboration?
El-Assady: For example, we’ve built a prototype to explore how we can improve interaction and trust. A central element here is interactive explainability: if you click on a system-generated proposal, you will learn how the AI came to this answer. A second important aspect is dealing with uncertainty: we’re trying to figure out what happens when the AI doesn’t find enough information to provide a conclusive answer. Instead of speculating, the AI directs the question back to the questioner. Technically, all of this is possible.
Freedom and individual responsibility, entrepreneurial spirit and open-mindedness: ETH Zurich stands on a bedrock of true Swiss values. The university for science and technology dates back to the year 1855, when the founders of modern-day Switzerland created it as a center of innovation and knowledge. At ETH Zurich, students discover an ideal environment for independent thinking, and researchers, a climate that inspires top performance. From 2026, the first ETH professorships at the Heilbronn Education Campus will begin their work to develop solutions to the challenges surrounding digital transformation and to train the next generation of critical and creative minds.
Funded by the Dieter Schwarz Foundation, the ETH Zurich is gradually establishing 15 professorships at the Heilbronn Education Campus, which are dedicated to “responsible digital transformation”. The main research areas are data science, artificial intelligence (AI) and cybersecurity. With a duration of 30 years, the funding agreement ensures long-term collaboration and continuity in scientific work.
Such models should be particularly exciting for companies. Are you witnessing a corresponding level of interest?
El-Assady: I hope so, because the interaction will increase. In the future, work teams – whether in management positions, production or creative professions – will increasingly rely on interactions with AI-based systems.
No teachers, no lectures, just project work: the 42 Heilbronn coding school is training the programmers of tomorrow. Two students and their managing director share their experiences.
read more
Researchers at the MOLIT Institute for Personalized Medicine are developing methods to analyze medical data and improve individualized cancer therapies.
read more
To harness the potential of high-performance computers, Hartwig Anzt from the TUM Campus Heilbronn develops and optimizes methods for solving mathematical equations. His goal: more efficient software.
read more