We seem to be caught in a loop that AI will either save or ruin democracy.
At one extreme, politics scholar David Altman worries that artificial intelligence is leading the world toward an “AI-driven political system in which approval is manufactured, public debate is hollowed out and citizens have no meaningful voice.”
On the other hand, former Google CEO Eric Schmidt and his head of research, Andrew Sorota, dream that “we must instead use AI to reinvigorate democracy, making it more responsive, more deliberative and more worthy of public trust.”
As upcoming elections in the United States, Canada and Brazil fuel speculation about AI’s impact, I have focused on a deeper question: why is the future of democracy so tied to the future of digital technologies?
How we regard political systems
My book SimPolitics offers a vital history of how we came to think about political systems like computer systems.
I call this “a computational imaginary of politics” that influences how we imagine politics to work, and why we often allow technology experts to serve as experts in politics.
The most tangible example is the current belief that AI can manipulate voters. While controlled experiments show chatbots can change minds, coverage quickly leaps to grand predictions about political transformation.
What matters more than the chatbot is a person’s willingness to discuss politics with a chatbot, something research participants chose to do. In reality, most people ignore efforts to persuade altogether.
Our ability to speculate about politics with technologies like chatbots comes from what what I call the aforementioned computational imaginary of politics.
My book explores how easy it’s become to imagine political change through technology. Through interviews and in-depth archival research, I trace the experiments, prototypes and lost stories to find the beginnings of this imaginary and help develop better critical skills to question its legacy today.
Politics and the involvement of computers has a long history in the United States. Even since the first UNIVAC computer, political scientists, programmers, politicians and sometimes even peace activists turned to computers to predict the next election or how voters would react to a new policy. Some tried to find a path towards world peace or to guess the next moves of the Soviet Union.
My book offers a way to critically reflect on the implications of the computational imaginary, posing questions we should ask whenever we hear about new applications of computers in politics:
- How do simulations mimic human behaviour specifically?
- How do simulations stay true to the worlds they model?
These two questions matter today, not only to understand AI’s impact, but to avoid being conned into believing the political version of AI snake oil — in other words, that AI can solve the world’s most pressing issues.
Mimicking human behaviour?
To answer the first question, it’s important to note that today, AI is widely construed as a stand-in to human behaviour. New companies like Aaru promise to use AI as proxies to voters, like political parties that poll public opinion without talking to human beings. By using polling data or custom AI models, these firms promise to replicate public opinion.
That’s not an innovation. My book traces many different efforts to use data, mathematics and cybernetics to model the minds of American voters. What I found isn’t that these efforts are misguided — voters are neither so mystical as to elude being predictable nor so forgone that political free will is in danger of being automated. Instead, I found that human agency is different from computer agency — and that difference is important.
Computers tend to move faster, are more dynamic and easier to understand than people — that was the appeal of modelling voters with computers in the first place. But there’s a risk in believing that a computer behaves more like how we would like voters to behave than how voters actually behave.
Advertising depends on the belief that information can nudge behaviour. Some computer models support this fantasy, creating versions of voters as if they were consumers shopping for candidates like they shop for soap.
The difference is vital at the moment because via AI, we are letting a computer model we cannot know stand in for a voter, creating a gap between how AI works and how we assume voters behave. What the first questions asks is to not assume that computer agency is equivalent to political agency. In fact, understanding the difference is necessary to interpret the results.
SimPolitics is not just about American political campaigns, however. In my research, I found computer models scaled up to simulate international politics and the realpolitik of the Cold War while drilling down deep into the minds of voters.
Strikingly, the same early computer model promised to simulate both voter behaviour and the behaviour of nations with the same formulas. Rational choice theory — a school of thought broadly claiming that utility maximization or self-interest best explains behaviour — explicitly claimed that its modelling could explain how a nation behaved as well as an everyday voter. And since utility maximization was a calculation, rational choice helped popularize the idea of voters and nations behaving like computers.
This increase in scale led me to my second question: How do simulations stay true to the worlds they model?
Being true to the worlds we simulate
As the scale of models got bigger and more sophisticated, a different challenge emerged: how to ensure computer systems stayed true to the political systems being modelled, or what is called addressability.
The answer may sound simple at first — better data and better computers. But greater computation capacity meant that models could simply behave more autonomously, risking being more disconnected from the world. Modellers, instead, had to strike a balance, making sure they did not get lost in the virtual worlds they created.
Read more:
Is your AI chatbot manipulating you? Subtly reshaping your opinions?
When looking at the results of any new AI experiment, my second question looks to understand how models still correspond to the real world. In short, simulations have limits. These limits are political, not computational. There is no technical fix to ensure computer models are in sync with the worlds they simulate.
SimPolitics is different from many of the current crop of books on technology in politics because it’s not about trying to fix American politics. Rather, it’s looking at the consequences of the computational imaginary for Americans and for the rest of the world.
The book warns that AI can perpetuate political fantasy, through how it works and how it warps political thinking. The book offers concrete questions to gauge AI’s impacts, a remedy to chatter over whether AI will save or damn democracy. If we let idle speculation drive our political futures, however, we lose the opportunity to imagine better worlds with machines — but not thorough machines alone.
This article draws on research supported by the Fonds de Recherche du Québec and the Social Sciences and Humanities Research Council of Canada.
