Supported democracy: reinventing direct democracy, AI and voting twice

Direct democracy has so far been a form of governance used in very few countries – most notably in Switzerland and California. Recent developments in generative AI – be it in the form of ChatGPT-type powered language models or as image-generating tools – suggest that entirely new ways of organising democracy will materialise very soon. Direct democracy of the Swiss or Californian type could be modified substantially.

Hans Gersbach and César Martinelli call such a democracy a ‘supported democracy’. This updated model of democracy might become an attractive one for the rest of the world as it may overcome the usual objections to direct democracy: namely that citizens lack the necessary incentives to acquire information to enable them to vote on complex policy proposals and that their voting decisions may be driven by short-term or even purely emotional considerations.

An AI assistant can learn citizens’ preferences and values

Gersbach and Martinelli suggest two basic ideas to update the model of direct democracy. During the first phase, all citizens would have access to the same type of AI assistant (henceforth called a ‘digital citizen’ or ‘DC’), which is trained to do two things. Firstly, the DC learns the citizen’s preferences and values through a Q&A process, which aims to create a digital citizen twin of its user. This digital twin replicates the citizen’s values and preferences as accurately as possible.

Secondly, large databases can be used to train the DC to acquire knowledge that is relevant for deciding on policy proposals. Such a trained DC could then suggest how the citizen it replicates would/should vote on particular policy proposals (initiatives, referenda or any law that is to be voted on in parliament) whenever such proposals clash with the status quo and citizens are therefore faced with a binary decision.

By ‘voting twice’ citizens retain their decision-making rights

During the second phase, citizens can vote using their DCs. This could be done in one round of voting on a binary decision. But, as citizens may not trust their DCs, or as the DCs may even be manipulated, Gersbach and Martinelli suggest starting with the following voting procedure (called ‘voting twice’). In the first round of voting on a given binary issue, the DCs of all citizens vote and the aggregation of these votes delivers an outcome that is made public. In the second round of voting, citizens vote themselves. In this second – human – voting round, either citizens follow their DC’s recommendation, they do not follow it or they abstain. Citizens thus retain their decision-making rights if they prefer not to mimic the DC’s decision. The outcome that is implemented is the voting outcome of the second (human) stage.

Like the one-person-one-vote principle, of course, equal access to a DC and standardisation of DC quality are crucial. These principles would need to be governed by rules adopted as part of election laws that are enforced by the state. Moreover, many variations of this process are possible. For instance, one might allow citizens not to let their DC participate in the first round. They would also be allowed not to use their DC at all, so they only vote in the second round.

Digital citizens can become experts

Supported democracy enables citizens to vote with a level of information approximating that of someone who has spent a large amount of time on a particular issue. DCs could even be experts – if trained sufficiently on a certain issue – so citizens following their DC’s decision would be sure to decide as if they were experts themselves. The next generation of AI will perform even better in terms of informed decision-making, thus enhancing the range and quality of democracy. A sophisticated DC can consider different perspectives such as the long-term view versus the short-term outlook.

Concerns that a DC might make mistakes, learn in a biased way or even be manipulated are addressed by our procedures. Voting twice enables citizens to see how DCs decide in the first round of voting. They retain the ability to inform themselves and to vote according to their preference in the second round if they have any doubts about DCs’ decisions.

The hope of more long-termism in collective decision-making

Overall, Gersbach and Martinelli expect that outcomes in supported democracy would allow citizens to vote on issues as if they were well-informed. The DC might even make them look at some points more closely, such as comparing the short- and long-run consequences of a decision. This would certainly encourage more long-termism in collective decision-making. As, ultimately, DC-supported decision-making will foster better-informed decisions, supported democracy will also enhance the performance and reputation of direct democracy.

The key elements of the decision-making process in a democracy are secrecy and equal voting rights, on the one hand, and comprehensive and accurate vote-counting on the other. This must be guaranteed to ensure that collective decisions reflect the electorate’s true preferences and that citizens have trust in democracy and support it. Generative AI provides democracies with a toolkit that allows for innovation.

Of course, supported democracy may not be a priority for Switzerland, which has a long tradition and experience of direct democracy. Nonetheless, supported democracy might become a more attractive model for many countries than today’s direct democracy.

However, a word of caution is in order. Every new procedure proposed for democracy (cf. Hans Gersbach, New Forms of Democracy, Social Choice and Welfare, 2024, forthcoming) must pass the test of in-depth scrutiny. It requires citizens to be trained, and drawbacks or unexpected collateral effects may arise during the implementation process. Only further research will reveal whether supported democracy can keep the promises outlined in this brief column.