Democracy Technologies: Can you briefly explain the format of the recent Citizens Convention on the End of Life in France?
Hélène Landemore: It’s the second citizens’ assembly to be conducted in France. The first was on climate in 2019/2020. This one convened a larger number of randomly selected citizens, 185, for a total of nine meetings. I was on the governance committee, so I saw it up close.
The assembly was convened to make a policy recommendation about whether or not we should liberalise the law on assisted dying, which is currently not allowed in France. All of this in a context where more than 80% of the population, when polled, said they are in favour of giving people the option. Both Belgium and Switzerland are doing it routinely now. But the medical body in France is very much opposed to it.
DT: What were the outcomes of the assembly?
Landemore: The assembly delivered recommendations that included liberalising the law to open up the the possibility of assisted suicide and even euthanasia, under conditions. Macron recently reconvened the convention for one more weekend, because a law was proposed by the government and has now been presented to Parliament. He wanted to show that he took the input of the citizens seriously, and that their deliberations were not without purpose or impact. He wanted to create an intermediary accountability step, allowing him to explain how the law was inspired by the convention’s recommendations and why it stopped short of some of them.
For example, it’s interesting that the proposed law does not use the term “assisted suicide”, but only “assisted dying”. So the government didn’t embrace the kind of brutal honesty of the citizens. They went for a more euphemistic term. And they did not follow the recommendations on the question of access to assisted dying or euthanasia for minors
DT: The citizens’ convention on climate was criticised for its lack of follow-up. Did they take a different approach to expectation management this time?
Landemore: I think they realised they overshot in the first convention, they promised too much. Macron promised to send the recommendations of the first convention directly to a referendum, regulation or the parliamentary debate. It’s not constitutional to make such a promise. So, of course, that promise could not be kept.
In my view, if anything, they overshot in the other direction this time. On the first weekend of the convention, we had a bunch of people saying the same thing over and over again: You are only one voice in a chorus of the people that Macron’s going to listen to. You don’t have the power to decide. It’s not legitimate for you to decide. And they said it so much that in the end, it felt like a democratic regression. From my point of view, it felt like: Wow, this convention will have very little influence.
But in the end, maybe because our expectations were so low, I think it had an influence. We can criticise the government as much as we want – but the reality is that it’s terra incognita in some respects. We’re learning as we’re doing, and a lot of mistakes are bound to be made. So I’m not too judgmental on that front. I think that as long as we continue with the next one, eventually pair conventions with referenda, and we calibrate expectations properly after a while, it should be fine.
DT: David, with make.org you recently launched Panoramic, an AI-tool that seeks to make the content of citizens’ assemblies more accessible. Can you describe what it does?
David Mas: The aim with Panoramic was to benefit from all the advances of generative AI. One of them is that it lowers the barriers to access complex information. Citizens’ assemblies generate a lot of content: you have experts come in, you have debates between the citizens. You have a report written by the citizens. The assemblies are transparent and open by design. Nonetheless, they generate so much content that it can be hard to follow.
Generative AI is very good at summarising data in language that is easy to comprehend for citizens. So we thought this is a good fit. We can take all the content, feed it to the AI, and then citizens can ask questions and have access to a quick summary of what has been said in the debates.
This is great, because what’s interesting in citizen assemblies is not just the outcome, but also the process. Because they are roughly representative, you can expect that any question you have was asked in the debates and answered either by deliberation between citizens or by an expert.
This way, you can extend the experience of a citizens’ assembly to citizens outside the process. The idea is that you can access debates and get all your answers at your own pace. This is a very important step. And it is empowering for citizens, because nobody edits the answers for you. You can ask a question you want from any angle, and AI will provide you with answers based on the debates.
DT: Data and information are never 100% neutral. It matters how it is presented and made accessible. What are the advantages and disadvantages of presenting a citizens’ assembly in this way?
Mas: The advantage is that nobody watches the three-hour videos of a plenary session of the assembly. So I think that the real benefit is that suddenly, people can follow the debates, whereas most people didn’t before.
Of course, no technology is 100% accurate, including generative AI. However, the architecture we use minimises the risk of approximation or hallucination from the AI. This is because the results are based on a semantic search of the content, which the AI then summarises.
DT: The tool uses a retrieval augmented generation model. Can you explain what that is?
Mas: So very simply, it has two components. The first component is a semantic search engine. The idea is that you can search for a relevant subpart of a debate based on semantics, with better results than a normal search engine. Then you have a second part where the AI generates the answers based on the results. So it takes something like 30 or 50 search results, and uses those to write the final answer.
We ran some intensive testing to make sure that the AI was answering correctly. Of course, you can always make AI hallucinate if you want. But if you use a tool to ask honest questions, it does not hallucinate. Sometimes it can miss things. So it’s not perfect yet, but it’s already good enough to make these debates accessible, and this benefit outweighed the risk of approximation.
DT: The ideal scenario is for a citizens’ assembly to spark a wider public debate. As you say, watching hours of video is not the most accessible way. Do you think this technology will reach a significantly different group than those already hyper-engaged?
Mas: It’s always challenging to bring people to a platform. Our platforms are designed to make it easy for people to access and work with online campaigns to bring people to the platform. Right now, we are running a campaign on social media to do that. We already ran a test, and 50% of the people we brought to the platform were actually using it and engaging with it. We designed it for easy onboarding. For example, there are suggestion prompts that you can click on to start the conversation.
Landemore: From my perspective, these kinds of tools are really important – not so much for deliberative democracy, but for open government. You want the public to know what’s going on inside these black boxes, right? As long as citizens’ conventions are black boxes, they lack legitimacy. So you need to make sure people can check which experts were invited, what they said to the convention, and what arguments were presented. If only so that the more motivated people like social activists can put pressure on the organisers and the government to do an even better job.
So it’s very important for transparency and accountability, even if it’s not a tool for mass participation or mass deliberation yet.
That said, I think this tool has a lot of potential if and when we pair citizens assemblies with referenda. Because then millions of people are going to go vote, and they might want to have information that they can trust. At that point, you need a tool that allows you easy access to the content.
Nonetheless, I’m still a bit frustrated with the tool, as it is currently mostly regurgitating what the experts said, because they’re the ones who were filmed. I’m not sure if the discussions among citizens themselves are included, because they took place in small working groups that were not recorded, as far as I can tell. So what the AI actually gives you is the experts’ opinions, presented in front of the citizens, not the citizens’ opinions.
So I would ask you, David, how do you plan on solving that?
Mas: That’s a very good question. We started with the materials available to us – the videos of the plenary sessions, and a PDF of the final report. One source of information that is lacking is what happened in the working groups. But there is an issue here, because usually, citizens in the working groups don’t want to be quoted. You need to reserve a safe space for them.
What we could do is to record everything, transcribe it, summarise it, and put the summary of the debate on the platform, but keep everything anonymous. Then you would have all the arguments that were exchanged during the debate.
Landemore: What about anonymising the participants and having the actual raw exchange of views?
Mas: We could do that too. But usually when you have the full text, anonymisation is difficult. Participants refer to each other by names, like ‘Michel, I disagree with you’ and sometimes the precise way you express yourself can identify you. That’s why we are thinking of using summaries instead.
DT: When generating the summaries, do you think there’s a potential for bias in what the AI picks up? Is there a chance it will skip over people whose arguments are not very concise?
Mas: That’s something we are trying to fight against in a research program with CNRS (Centre national de la recherche scientifique) and Sciences Po. Researchers are assessing if the summary is subject to bias or hallucinations, and trying to fix that. We know that human beings are sensitive to formulation. The hope is that AI might actually be less sensitive to formulation. Because in any setting, people who express themselves well always have an advantage.
AI has the potential to break down this barrier. If you are online, you can use AI to help you express yourself more clearly. When you contribute online, it can suggest rephrasing, or help you to make your argument stronger by asking questions.
Read part two of the interview here.
