21 January 2025
Researcher Relay #1
Thomas Hildebrandt advocates for reliable AI in the public sector – It’s time to end “probability guessing”
For the past 15 years, Thomas Hildebrandt has been researching the use of AI in the public sector. Time and again, these projects have ended up on the state’s IT graveyard. After encountering numerous challenges, he now senses a shift towards models that may be less glamourous than language models but are far more reliable and energy-efficient. This is the first article in the Researcher Relay series.
It is easy to see how AI could benefit hospitals or municipalities. Automated processes and more time for citizen-facing tasks are just some of the potential advantages in the public sector’s AI utopia that has been sought after for decades.
Early efforts began with profiling citizens though data registries, followed by chatbots powered by language models. Yes, success has been limited, and many of these initiatives have ended up on the public sector’s graveyard of failed IT projects.
With 15 years of experience researching AI in the public sector, Thomas Hildebrandt, Professor at the Department of Computer Science at the University of Copenhagen, is one of the most prominent voices in the field. According to him, language models like ChatGPT have gained popularity due to their ability to produce responses that seem convincing and trustworthy. However, these models are not designed to operate in a context where answers must always be correct.
“These models don’t generate answers based on true logic. They are merely text generators that predict the most likely next sentence—not based on facts, but on statistical probabilities. This doesn’t work in a public context, where we must be able to explain why a citizen is being denied welfare benefits or why children are being forcibly removed from their homes,” he explains.
Hildebrandt emphasizes that public decision-making demands transparency. Citizens need to be able to trust that AI systems are not simply guessing but are in fact complying with the law. Moreover, these processes must be documentable.
“We must be able to explain how the system arrived at its decision. This is a fundamental requirement that we must demand in any society governed by law.”
From predictive models to hybrid solutions
The history of AI in the public sector is full of failed attempts to integrate hyped technology. Initially, there was a belief that linking data from various registries could predict everything from long-term unemployment to child benefits.
“The result was that caseworkers lost trust in the technology, and users began to complain about discrimination,” Thomas Hildebrandt notes.
A few years ago, language models emerged as a new ready-made solution promising significant changes. Public employees were trained to use prompts, and retrieval-augmented generation (RAG) solutions were developed to train models using databases of legal texts. But even these solutions have proven problematic.
“These systems are too complex to maintain, and what happens when a software update occurs? There are just too many unknown factors,” Hildebrandt warns.
In addition to their imprecision, language models are extremely energy-consuming.
“It takes billions of calculations to run a language model, making them extremely energy-inefficient. A rule-based chatbot uses 1,000 times less energy. It may be less exciting, but it is far more reliable and traceable,” he says.
Hybrid AI may be the answer
For Thomas Hildebrandt and his research team, the future lies in hybrid AI, which combines language models with rule-based systems. This approach allows citizens to interact with the language model in a familiar way, while the answers are derived from a rule-based system that encodes legal provisions.
“It’s about using the best of both worlds. Rule-based systems offer reliability and traceability, while language models provide flexibility and user interaction,” he explains, using a construction scenario as an example:
“If your application to build a carport is rejected, the system must be able to explain the specific rules behind the decision. This is the kind of AI we need—not probability guessing.”
As a result, his team has received support for new projects focused on developing rule-based systems.
“It’s encouraging that we are starting to see signs from the public sector that this is the direction they want to pursue. That’s why I’m optimistic about the future of AI in the public sector,” he concludes.
See the video here:
Researcher relay
The interview with Thomas Hildebrandt is part of the Researcher Relay series, a collaboration where researchers from Danish universities pass the baton to one another.
Thomas Hildebrandt has selected Naja Holten Møller, Associate Professor at DIKU, to take over the relay. She works with AI in the healthcare sector on a daily basis.
