This is how artificial and human intelligence work together to strengthen democracy
image: Tara Winstead via Pexels
How should the municipality of Eindhoven organise the reception of asylum seekers? What is the best electoral system for the Dutch Parliament? Governments regularly ask citizens for their opinions. Computer scientist Michiel van der Meer has improved the method used for citizen assemblies. He will defend his PhD thesis on 26 March.
Since 2006, governments have consulted thousands of people on topics such as COVID-19 measures or the safety of provincial roads. In his PhD research, computer scientist Michiel van der Meer demonstrates how people and artificial intelligence (AI) can work together to get more out of these citizen assemblies.
‘AI helps with scaling up: it can process far more text and arguments than people can.’ However, people are better at understanding context, Van der Meer explains. ‘By combining human and AI efforts, we can process more opinions while also uncovering the diversity of arguments presented.’
Do not leave summarisation entirely to AI
The National Climate Citizens’ Assembly is currently underway, in which 175 Dutch citizens are formulating advice for the government. Van der Meer explains: ‘During such an assembly, large groups of people have lengthy discussions at the table. Afterwards, a summary is created that everyone must approve.’ These days, AI can assist with this process. ‘You feed it the entire transcript, and it generates a summary.’
Van der Meer regrets that no human involvement is required for the summarisation process. ‘That stage is where a lot of social and interpersonal dynamics still take place. People gain a better understanding of each other.’ That is why he developed hybrid analysis methods.
AI can filter the opinions of tens of thousands of participants…
Van der Meer conducted his research within the Hybrid Intelligence research programme, a collaboration of seven universities. While he is earning his PhD in Leiden, he also received supervision from TU Delft and VU Amsterdam. His research made use of the online public consultations run by Populytics. This company, a spin-off from TU Delft, allows thousands of people to step into the shoes of policymakers via an online platform. Participants can write down their arguments in their own words, which researchers then analyse.
Municipalities are often the clients for these consultations, but in 2021, as many as 30,000 Dutch citizens shared their views via Populytics on the abolition of COVID-19 measures. Van der Meer says, ‘Populytics was unable to analyse all 30,000 opinions. That is exactly the kind of task where AI can help.’
…But only if it is trained to detect all those opinions
Current language models such as ChatGPT struggle with this task. They are primarily trained to find an average—the most likely option. ‘As a result, they often overlook the rarer arguments. These are crucial, as people with uncommon arguments frequently feel unheard.’ In an ideal democracy, and certainly in a public consultation, these arguments must be included. ‘You want every participant to recognise themselves in the final report.’
In an ideal democracy, the rare arguments are also included
Humans must place opinions in the right context
Van der Meer designed a hybrid system in which AI and humans collaborate to extract diverse opinions from large datasets. ‘The language model in my system is different from ChatGPT, for instance—it is specifically trained to detect opinions.’ In Populytics’ COVID-19 consultation, the language model identified opinions that were interesting enough for human review. A group of people then classified all the opinions into categories. Van der Meer says, ‘I found ten clusters of arguments that were missing from the final report. But my analysis also overlooked five! So, there is still room for improvement.’
People must also go door-to-door
Does this mean we will soon be able to organise large-scale citizen assemblies effortlessly, ensuring that everyone feels heard? ‘Yes, but with a caveat,’ says Van der Meer’s PhD supervisor, Catholijn Jonker, Professor of Interactive Intelligence at Delft and Leiden. ‘If you want inclusivity, you must also engage people who would not normally participate—such as those who lack digital skills, do not speak the language, or distrust the government.’ This means that people with strong social skills must also go door-to-door. So, the process cannot be quick, easy, or cheap. ‘Automation should never be a goal in itself—you must combine the right forms of intelligence to carry out the task as effectively as possible.’
Michiel van der Meer will defend his PhD thesis, Opinion Diversity through Hybrid Intelligence, on 26 March.