AI can help us find common ground

Could this combat the growing societal divides?
25 October 2024

Interview with 

Christopher Summerfield, University of Oxford

HANDSHAKE

A handshake.

Share

A new study has found that an artificial intelligence tool can help people with different views find common ground. It works by more effectively summarising the collective opinion of the group than humans can. The work was carried out by Google DeepMind and the findings have just been published in the journal Science. Christopher Summerfield at the University of Oxford was involved in the research and he’s been telling me all about it…

Chris Summerfield - The idea behind the project came from a concept which is well known in political science, that our democratic processes can be facilitated if you sample representative groups of people from across the country and get them to, and get them to thrash out their point of view on a political issue. We wondered whether artificial intelligence could be used to make that process simpler. If you have a group of people who want to decide something and you get them around the table, the first limitation is that a table is only so big and so you can only fit so many people around the table. And the second limitation is that when you're having a discussion face-to-face, only one person can talk at a time. And we wondered whether you could overcome those limitations by getting an AI to listen to all the different viewpoints and to generate an output which summarised their collective opinion.

Chris Smith - How do you actually train it to do that in the first place, Chris? Because, for instance, if I wanted to do the equivalent thing medically, I wanted to ask an AI engine, has this person got cancer by looking at a brain scan or something? I'd have to have trained it on what a normal brain looks like so it can spot an abnormal one. So how can you ask it to synthesise a collective opinion when all it's got is a snapshot of a few people's perspectives at that meeting?

Chris Summerfield - Great question. So we did in fact train it. We got together small groups of people and we got them to write their private opinions about issues that related to UK public policy. So things like, should we lower the voting age to 16? And we got the machine to generate an opinion, which you can do because language models, you can ask them things and they will respond. And we then got people to rate whether they agreed or disagreed with the thing that the language model said. By doing that over and over again with lots and lots of participants, we were able to train the model to make a really good guess about whether people would be willing to endorse a statement which was generated by the model.

Chris Smith - If you've got a model which is trying to take everybody's viewpoints into account, don't you just end up with it smearing into the middle and everyone's kind of vaguely happy, but no one's unhappy, but no one's delighted either.

Chris Summerfield - I think that's a super question and there's nothing in the model which will give more weight to one of the participants in the discussion than any other. It's very explicitly trained to give equal weight to everyone who participates. The selection is really beyond the control of the machine itself. But once you have a group which is composed of a bunch of people with different views, what the model more or less, not quite guarantees, but what we trained it to try and do is to produce the opinion which is going to kind of form the best representation of the distribution of views that you get in that group. And so what that means is rather than just, for example, finding compromise is it seems to learn to write a statement which reflects obviously the majority view, because in a democratic process the majority should be a guide for what the solution is, but it also really strongly represents the minority view. So it gives the people reading the statements, we think, the sense that even if their view was in the minority, that their voices feel heard.

Chris Smith - Did you try poisoning the well as it were just to see what would happen if you injected some really extreme viewpoints in there and see if they were captured nevertheless. Or synthesise some responses that were so way off the normal distribution of normality. Did it end up chucking those in and did you get any bizarre hallucinations out of this? Because that's been the other thing that people have worried about with AI systems, isn't it? It just sort of confabulate things and we end up with things appearing that we don't really think are grounded.

Chris Summerfield - So we did do those kinds of tests, although we didn't do them with sort of politically toxic inputs. Rather we made sort of silly messages which had nothing to do with the question at hand. And what we found was that to a great extent, the model was prone to ignore contributions to the debate, which were completely irrelevant to the argument. Because it's been trained to be clear and concise, it knows to basically ignore that information. As for factuality, we didn't control the factuality of the responses. And that would be a very interesting innovation because when citizens juries happen, one thing that the mediators do is they sort of fact check, rather like the moderator did in the recent debate between Trump and Harris.

Chris Smith - How do you see this being deployed then? Do you think that this is effectively something that could be used as an adjunct to the current process? Or could we go the whole hog and just say, right, we are going to have a citizen's jury and we're going to feed all the perspectives en masse into a system like this and what comes out? That's what we're going to go with.

Chris Summerfield - The computer-mediated process that we've built is never going to replace face-to-face deliberation. Face-to-face deliberation obviously has many, many benefits which our tool would never be able to recreate. Such as, when you discuss with someone, you very often build up feelings of kind of empathy or mutual understanding that have nothing to do with the specific outcome of the statement that you end up generating. But what we can do is we can run deliberation at scale at a very commercial end of the spectrum. You could use it for things like market research, but perhaps the most interesting application is in thinking about the political process itself, the ability to aggregate or combine the views of lots and lots of people provides an opportunity for us to surface information to political leaders about what everyone thinks about everything else. And that, I think, could be extremely valuable.

Comments

Add a comment