How to get people to agree with each other using AI - an interview with Prof. Chris Summerfield
Exploring work on the "Habermas machine"
Many people worry that the increased use of language models online could lead to a dystopian hellscape, but could AI also work *for* democracy?
A paper in Science last month was widely discussed in journalism and AI circles after detailing an effort to facilitate meaningful discourse, and agreement, using artificial intelligence models.
The tool, called the Habermas Machine after philosopher and social theorist Jürgen Habermas, was seen as better than randomly selected humans at analysing input from a group of participants and generating a statement that the most people agreed with.
Prof. Chris Summerfield, formerly of Google DeepMind and now at University of Oxford, told Overtone that the project was inspired by Hélène Landimore’s book Open Democracy and the problem of creating discourse groups at scale.
“You know, a table is only so big, and if you've got to get everyone around the table, then that's going to intrinsically limit the number of people who can be involved in discussion. At this point, a light went on. Maybe you could actually use a language model to solve that problem,” Summerfield said in an interview.
The group created a structure of multiple AI applications working together, one of which generated statements and the other of which predicted how likely the statement was to be accepted by the different human members of the group, allowing it to select the one that would be most acceptable.
Summerfield said he was surprised by the level to which the Habermas Machine incorporated viewpoints that were not the majority in each group. In some ways it goes against a common understanding of language models as repeating patterns seen in their training data because they occur the most often. A piece from the flurry of attention around AI image models last year spotlighted how they create Native Americans or Soviet soldiers with beaming American selfie smiles, presumably because there was large amounts of training data that was similar and despite the fact that Soviet soldiers and Native Americans don’t share the same cultural smile as the 21st century U.S.
“What it seems to do is write statements which are faithful to the majority, but include prominent elements of dissent. So they weave in stuff of people who would otherwise feel kind of disenfranchised. It weaves their views in so that they feel heard. And this is what you hear again and again when you interview people from citizens assemblies, that one of the reasons that people feel positively about the process is not necessarily because the outcome matched their viewpoint. Very often it doesn't, but people feel like they can contribute,” Summerfield said.
The question then, is what exactly would an artificial intelligence solution like the Habermas Machine be scaling if it was used online in a widespread way?
Much of academic media studies and discourse theory is looking at idealized speech situations, where people are logical and measured in their statements, rather than the real rough and tumble tumult of a social network, for example.
Another wrinkle from the study focused on citizens assemblies is that the artificial intelligence machine wasn’t in charge of supplying facts or reading material to the participants. Two people can be influenced against trusting a source or disregard information based on where it is coming from.
It is something that we at Overtone think a lot about when we want our customers to be able to judge a piece of content not only from whether its source is famous or popular or not, but based on what it contains (Is it opinionated? Is it factual? Is it angry? Is it happy?). Perhaps it is easier to get people to agree if they start from the same place, or you at least know which audience they are coming from and the characteristics of the media they are reading.
It is something that Summerfield wants to continue to research in the hopes of finding better methods for making discourse happen.
“Many, many things can be cast as a mechanism design problem. And I do think people are a lot smarter than we give them credit for. People also are a lot more well meaning than we give them credit for. If you find the right modes of expression, then things can be functional,” he said.
“If you take a look, [at the citizen assemblies] one of the things you will see is that the viewpoints that people express are remarkably sensible. They're remarkably articulate. They're remarkably heartfelt. People just come across really well. It's like the opposite of what you would expect if you just scroll through Twitter for a while.”