AI in the public sector is not a new concept. The technology has slowly made its way into municipal offices and hospital corridors. However, the enthusiasm has been lacking. It is timely, therefore, that the government has now established a task force to identify the both the potential and the barriers to AI adoption in Denmark’s public sector.

At Denmark’s national research center for digital technology, DIREC, we have supported a number of research and innovation projects over the past four years, where researchers have collaborated with public organizations to explore how AI can solve challenges and create value for Danish society.

We are facing a multitude of opportunities with AI. However, to truly reap these benefits, we need stronger collaboration between public sector entities, innovative companies, and researchers from Danish universities.

Millions in potential savings

The research projects funded by DIREC demonstrate that significant benefits can be gained from applying digital technology and AI in the public sector.

In one project, researchers across Denmark are collaborating with several hospitals to analyze CT scans of kidney cancer patients using AI. This partnership has shown that treatment times can be shortened by 2-4 weeks, unnecessary biopsies can be avoided, and approximately 15-25 million kroner can be saved annually in the healthcare system.

In another project, researchers are working with Danish utility companies to explore how AI can help prevent flooding during heavy rainfall by efficiently distributing precipitation and wastewater across lakes and water systems.

What all these projects have in common is that AI technology is not something we can simply take off the shelf. Researchers, businesses, and the public sector organizations must work together to develop specific solutions to various challenges. Here are three examples of these challenges:

1. AI must be explainable
Large AI models consist of millions, sometimes billions, of parameters, and their outputs depend on complex combinations of these factors. Over the last decade, attempts have been made to use AI to predict outcomes such as child removals and long-term unemployment. However, these attempts have failed because caseworkers have been unable to understand the models’ logic.

The public sector does not need a chatbot that simply spits out answers. We need technology that can engage with both citizens and public employees. Therefore, it is essential to develop a layer of technology that can explain how models arrive at their conclusions. This means that decisions, such as whether to approve a building permit, should not only result in a yes or no answer but also provide the reasoning behind the decision. Explainable AI (XAI) is crucial for AI to become a valuable partner that supports both citizens and public employees.

2. AI must comply with the law
Large language models do not always provide correct answers. They often have a tendency to hallucinate. This is rarely an issue when generating something like a party song using ChatGPT. But if AI is used as a chatbot in communication between citizens and the public sector, we need to ensure that the model’s responses are correct and compliant with applicable laws.

There is a strong need to build a layer of technology around language models that can provide this assurance. Such a rule-based approach is not currently available in large language models. These models should be seen more as a user-friendly interfaces to a more advanced AI system, one that functions like an experienced caseworker, always arriving at the correct conclusion.

3. AI must protect privacy
For AI to be effective, it needs to be trained on large datasets. However, much of the data used by the public sector is sensitive, which presents a significant challenge in a society that values privacy.

Several researchers at Danish universities are working on solutions to this issue. One potential solution is “federated learning,” where the algorithm is trained locally but “visits” servers at places like hospitals and sends the results back ensuring that data never leaves secure environments. Another approach is “edge-based computing,” which compresses large AI models into smaller versions that can run on local computers, thus avoiding the need to process sensitive data in the cloud.

Create digital dream teams across sectors

These three challenges are just a few examples of what many researchers at Danish universities are dedicating their time to solving. Across these institutions lies deep expertise in what AI technology can achieve and what additional technologies are required to make AI successful in the public sector.

The most important task for the newly established AI task force will be to tap into the vast knowledge available at universities, companies, and public sector organizations. We need to bring together experts in digitization and AI from across Denmark to form digital dream teams.

Only by creating these digital dream teams can we ensure that AI becomes a success in the public sector.

This article was published on altinget.dk on August 26, 2024