ITU researcher investigates how AI can better communicate uncertainty

ITU researcher investigates how AI can better communicate uncertainty

AI-based language models are increasingly being used in healthcare, where they can provide answers to complex questions in a matter of seconds. However, although their responses often appear clear and convincing, they are not necessarily accurate - and this is why such models need to learn how to communicate their own uncertainty.

Written 26 August, 2026 12:35 by

Imagine that you are feeling unwell and experiencing a number of specific symptoms. Before calling your doctor, however, you turn to your preferred AI platform, enter your symptoms, and receive a response containing a series of percentages indicating the likelihood of different diagnoses. For example, it might estimate a 15% risk of gallstones, a 10% risk of liver damage, and a 75% likelihood that your symptoms are caused by a temporary irritation of the digestive system

From probabilities to language

or people without a medical background, these percentages, and the way they express uncertainty, may appear unnecessarily alarming and, in the worst case, counterproductive. This is the challenge that Associate Professor Christian Hardmeier from the IT University of Copenhagen is investigating together with colleagues from DTU and the University of Edinburgh as part of the project "Conveying Caution & Confidence: Quantification and Communication of Uncertainty in Large Language Models."

"Numbers alone are not enough. A confidence score of 75% may make sense to a statistician, but not to a ten-year-old," explains Christian Hardmeier. "We want to use linguistic strategies, such as qualifying statements with words such as 'perhaps' or 'probably', so that uncertainty is communicated in a way that people can understand. Crucially, those expressions must accurately reflect the model's actual level of confidence."

Christian Hardmeier is therefore working to improve AI systems' ability to communicate the level of confidence they have in their own responses. Ultimately, the goal is to give users a more realistic understanding of when they can trust an answer and when they should remain cautious.

"When we speak to another person, we assume a shared understanding. Reservations, enthusiasm, uncertainty, and professional expertise all play a role. By contrast, the responses generated by language models, however well phrased and confident they may appear, are the result of calculations and statistical processes. Nevertheless, we tend to attribute human characteristics to them."

Uncertainty as part of the model

A central part of the research is focused on making uncertainty an integral part of the model itself. Using Bayesian methods, the researchers seek to quantify how likely it is that a given response is correct. Bayesian statistical methods continuously update probabilities as new information becomes available. However, this presents a significant technical challenge, as modern language models consist of billions of parameters, and existing methods for estimating uncertainty are difficult to scale to that level.

At the same time, it is not enough simply to calculate uncertainty. It must also be actively reflected in the model's responses. Moreover, the accuracy of any response is always dependent on the training the model has received. "If a question concerns something the model has not been trained extensively on, it will still attempt to provide an answer. It does not 'know' that it lacks the necessary expertise, and so it produces a response that appears plausible. However, if a model appears more confident than it actually is, this can lead to misunderstandings and incorrect decisions," says Christian Hardmeier.

Different needs in healthcare

The project focuses in part on healthcare, where the need for clear and accurate communication is particularly important. In collaboration with partners such as the Virtu Research Group in the Capital Region of Denmark, the researchers are investigating how uncertainty needs to be communicated differently to different user groups, including patients and healthcare professionals.

However, Christian Hardmeier does not believe that large language models will fully replace humans. "I do not think they will become capable enough in the near future. That is why we only use these models in conjunction with human care," says Hardmeier. For example, the Virtu Research Group is developing a method for analysing therapeutic conversations that could be used in the training of therapists. Meanwhile, the 1813AI project is exploring the use of a chatbot to collect information while callers are waiting to speak to the 1813 medical helpline, with the aim of improving the efficiency of the consultation itself.

For Christian Hardmeier, research into uncertainty in AI is fundamentally about basic research. "It is relevant to almost all applications of large language models, but it may not be the most critical component in every practical application," he says.

Further information

Jari Kickbusch, phone 7218 5304, email [email protected]