By Thor Blond, Sarah Steinitz and Maja Ejrnæs, Queen Mary’s Center, University of Copenhagen
ChatGPT was launched in late 2022. By 2026, generative artificial intelligence is barely old enough to start school. Yet today, the best models can solve programming tasks on their own that would take a human expert a full working day. And technology has begun to leave its mark on the labour market: According to a new analysis from Denmark’s Central Bank, Danish companies adopting AI subsequently hire fewer new employees than comparable companies.
Politically, the ambitions are as great as the pace is high. Denmark should become a world leader in the use of artificial intelligence, the technology must be rolled out in Danish municipalities on a large scale, and it is supposed to free up at least 30,000 full-time equivalents in the public sector.
At the same time, the questions are piling up. Does artificial intelligence prevent students from learning? What happens to trust when a chatbot replaces the employee at the entrance to civil service? Who is responsible when AI has contributed to a decision, and what does this mean for citizens’ legal rights?
None of these questions can be answered by the technology itself. It requires knowledge of both machines and people, and it requires that the relevant knowledge reaches those who make decisions. Consequently, it is not only important how skilled the large language models become, but crucial whether the state and the government have the capacity to understand and act on what these models set in motion. Here research has a part to play.
Link between development of research and policy
Together with Professor David Budtz Pedersen from Aalborg University, at Queen Mary’s Center we have interviewed 24 civil servants across ministries, including in the Ministry of Digitalisation, asking them how research is used in policy development. In Denmark, it is not research we are lacking. We have strong research environments in both technology and the areas that artificial intelligence intervenes in: Learning, working life, administration, the rule of law and democracy. The problem lies in the connection between research and political decision-making processes. In our survey, three elements in this link are repeated.
Relationships. When an area develops as quickly as AI, an urgent – and often diffuse – need for knowledge translation arises. The need is not necessarily met by ordering a research review or waiting for a new research project to produce results. So who can you call for quick professional sparring? The civil servants emphasize personal relationships and ongoing contact as elements that bring knowledge into policy development. But the contact is unevenly distributed: Some ministries have fixed agreements on research assistance (so-called public sector services), while others only has sporadic contact with the research community.
Translation. Research and policy follow different logics. As one official says: “I think it can sometimes be annoying that [the researchers] – instead of reducing – they actually increase complexity. (…) But in the world we are in – the political world – you are not looking to increase complexity, but to constantly try to reduce it, to be able to explain and come up with solutions.” Artificial intelligence is precisely a field where research-based knowledge can make the picture of the field more complicated. Hence, someone needs to translate research-based knowledge into concrete professional assessments that fit into the current context – and on which political decisions can be made.
Time and professionalism. A more complex scientific landscape calls for a civil service that can handle it. But the pace of the legislative work has been increased, and according to the report Magtudredningen 2.0, 69 % of civil servants say that speed affects the quality of legislation. At the same time, Magtudredningen shows that it is common to spend less than two years in the same position, and that generalists take up more space in civil service at the expense of specialists. As one official emphasizes in our interview survey, you may find yourself at a loss when as a generalist you need to assess subject-specific knowledge: “How do I know that this [study] is better than that one? And is it? Or are they two equally good studies that just show two different things?”.
AI doesn’t solve the problem for us
It is tempting to think that artificial intelligence itself can close the gap between available knowledge and political decisions. AI can help find and synthesize large amounts of knowledge and at high speed. But the official’s question above still stands: If you cannot assess which of the two studies you should trust, you cannot assess whether the large language model’s answers hold up.
In a recent piece in Nature, experts point out that experienced science advisers assess the credibility of evidence based on a number of parameters, including professional insight, researchers’ reputations, and colleagues’ assessments. A language model’s answers can change with small changes in how you ask questions and which parameters you set for its work. At the same time, there are no clear standards and procedures for verifying and evaluating the answers that the models provide. How the technology best contributes to science advice in a credible and useful way is still an open question. And while generative AI can go a long way, technology cannot replace human judgment, political priorities, or democratic accountability. Data from our survey points in the same direction: Knowledge only becomes useful when someone with professional judgement can bring it into a concrete political context in a responsible way.
The challenge is not only Danish. OECD describes how governments around the world struggle to build expertise and capacity to handle and use AI. The European Commission, Finland and the United Kingdom are experimenting with new forms of scientific advice, strategic foresight and rapid mobilisation of knowledge. In Denmark, the Ministry of Digitalisation has, for example, used the research reserve to launch initiatives such as the research centre CAISA (The National Centre for Artificial Intelligence in Society), and has, moreover, established the Centre for Social Media, Tech and Democracy, which has a PhD student employed. These are good steps, but individual initiatives alone will not do it. As an official from the ministry says in our investigation: “It’s a cultural change. And it just takes a long time.”
Build capacity as technology rolls out
Time is not abundant in a busy civil service. And time pressure is not reduced by development in artificial intelligence that charges ahead at a pace that can be difficult to keep up with. If the ambition of 30,000 freed-up full-time equivalents in the public sector is to transferred into better welfare and not just faster case processing, the investment in technology must be followed by an investment in the connection between research and policy. When time and specialized subject knowledge are scarce in the ministries, access to knowledge should not depend on which contacts individual civil servants have in their address book – or how they asked an AI in their prompt. We should think about strengthening the structures of science advice: More permanent connections between ministries and research environments – and people who are tasked with translating between the two worlds. Otherwise, we risk a public sector that works faster and with fewer resources but without making smarter and better decisions.
The ADD guest blog posts provide insight into the controversies and dilemmas we face in the digital age, along with ideas on how to strengthen digital democracy. The guest blog presents projects, activities, experiences and perspectives from ADD’s network of researchers and professionals.
Photo: Philip Davali/Ritzau Scanpix