Researchers to examine AI explanations
AI can provide answers that seem convincing, but its explanations are not always reliable. In a new research project at Jönköping University (JU), researchers will investigate how logical reasoning can be used to check and improve AI explanations.

Today’s AI systems still struggle to show how they arrive at their answers. When we ask an AI system to explain its reasoning, the explanation may be incomplete or even incorrect. One reason for this is that many AI systems rely on patterns learned from large amounts of data rather than on logical reasoning.
The researchers at JU will study what is known as abductive reasoning. Put simply, this means starting from something we observe and trying to find a plausible explanation. For example, if someone is carrying an umbrella, one possible explanation is that it is raining. But if new information shows that it is not raining, the explanation needs to be revised.
“We want to develop methods that can check whether an AI system’s explanation is actually correct. If the explanation contains minor errors, the aim is also to make it possible to correct them,” says Johannes Schmidt, Associate Professor and Senior Lecturer in Computer Science at the School of Engineering at Jönköping University and project leader for the new project.
Laying the foundation for more reliable AI
The researchers will approach the problem in two ways. The first is to develop a method that can examine an AI system’s explanation and detect and correct errors. The second is to gain a better understanding of what happens inside a neural network when an AI system arrives at its answer.
How AI explanations can be checked and corrected is an area where many questions remain unanswered and where much more research is needed.. The researchers therefore aim to develop a theoretical foundation and new methods that, in the long term, could contribute to AI systems whose answers and explanations are easier to scrutinise and assess.
“AI systems are rapidly becoming more advanced, but their ability to provide reliable explanations is not developing at the same pace. We therefore need to build a solid theoretical foundation for AI systems that not only provide answers, but can also explain and justify them in a way that can be verified,” says Johannes Schmidt.
The project is led by Johannes Schmidt, Associate Professor and Senior Lecturer in Computer Science at the School of Engineering at Jönköping University, and has received SEK 4.3 million in funding from the Swedish Research Council.
The project is planned to run for four years and will also include a doctoral student position. The research will be conducted at the School of Engineering at Jönköping University in collaboration with researchers from Linköping University.