Interviews
Marta Cambronero Deliberating When We Disagree: Conflict, Hatred, and the Public Sphere
Marta Cambronero Garbajosa is a journalist, holds a Ph.D. in Information and Knowledge Society from the Open University of Catalonia (UOC), and is a postdoctoral researcher with the Communication Networks & Social Change group. Her research focuses on the digital public sphere, political antagonism, and citizen responses to hate speech—topics she addressed in her doctoral dissertation Communication Against Hate (2025). We spoke with her about the limits of democratic deliberation, the role of emotions and algorithms in shaping public opinion, and the need to build spaces where conflict can be expressed without turning into violence or exclusion. A conversation that also explores what role artificial intelligence can play in a democracy that needs to learn to tolerate disagreement.
The Public Sphere After Habermas
Platoniq: Let’s talk, then, about public discourse and the public sphere—especially the digital one. For a long time, we imagined the public sphere as a space where different positions compete through arguments. Does this image still help you understand a public conversation shaped by platforms, algorithms, metrics, and also personal emotions—whether genuine, recreated, or even simulated?
Marta: Yes. Our initial understanding of the public sphere is the one defined by Habermas in his seminal work. Since then, there has been a great deal of work on the concept, but almost always the most interesting developments have centered precisely on the critiques leveled against that concept.
There’s a preliminary point that I think is important to keep in mind, because it helps clarify the rest of the discussions surrounding the public sphere: understanding Habermas’s method. He doesn’t specify or clearly explain it in his work, which is why it has taken so many years to understand it. You have to read a lot of different authors to see if it really works that way, and even then, there’s no consensus among everyone who studies it.
What he does is employ the method of immanent criticism—typical of the Frankfurt School, which he learned from Adorno—which consists of comparing what a system proclaims with what it actually does. Based on this dialectic, he observes how modernity as a whole has been constructed, on the one hand, through an Enlightenment vision of human potential—of what humans might be capable of achieving—and, on the other hand, through the realization that these possibilities have always been threatened or, in some way, contradict what we subsequently find empirically in reality.
In his first work, Habermas discusses two fundamental transformations of the public sphere that prevent it from functioning as the system itself claims it should.
The first major transformation is the bourgeoisie’s domination over the discourses of the modern public sphere. In other words, it succeeds in passing off its class interests as if they were the interests of society as a whole. At that time, there was a struggle: the labor movement was not only emerging but had become a powerful force. We now know how it played out historically, but at the time, discourses from various classes were in contention, and it was unclear what would happen. The same can be said of the suffragists. The discourse of the male bourgeoisie prevailed.
The second shift is the transition from a decentralized communication system to a centralized and controlled one. We moved from small newspapers appearing here and there, without much stability, and from the importance of literary salons in shaping public opinion… All of that shifted toward a different model: the major mass media outlets emerged. First came the news agencies, and then this entire process intensified with the advent of radio, television, major newspapers, and the entire conglomerate of cultural industries. The second major transformation, therefore, is a media one.
This introduction is important because Habermas argues that for deliberation to exist, certain conditions must be in place. Regardless of whether this normative vision is realized in reality—or whether it is realized only sometimes and perhaps with nuances—the value of his analytical framework is essential for thinking about the public sphere.
Platoniq: I think it’s very clear. Sometimes it’s not at all clear, but you’ve made it super clear.
Marta: So, what’s going on? What’s happening is that there isn’t a specific moment when democracy—or the deliberative conditions that allow for a legitimate democracy to be sustained through public opinion—begins to deteriorate. They never really existed in the first place. What we’ve seen in recent years is that the threshold separating us from those optimal conditions for deliberation is getting farther and farther away. In other words, the possibility of aligning the actual ways public opinion is formed with the normative model is becoming increasingly remote. And so a sense emerges that normative models no longer work.
Add to that the fact that today’s algorithms are much more opaque. It’s not as if the previous way of organizing public information was particularly transparent, but…
Platoniq: At least you used to be able to go somewhere where someone would explain it to you.
Marta: Exactly. That doesn’t happen anymore. We can’t audit the algorithms. It’s not even very clear whether they were specifically designed to produce the effects they ultimately do. We’re still trying to understand why they work this way. What we do know is that, when it comes to content visibility algorithms, they’ve been designed primarily to maximize profit. They’re intended, above all, for the digital advertising system. That’s why we don’t know to what extent all the harm they cause is a byproduct that no one had explicitly anticipated. They’ve been designed with something like this in mind: “I want to know what interests you; I want to know where you’re focusing your attention, because to me, that translates into exchange value for advertisers. I can tell them: this person’s attention, at this moment, is worth this much.” And this is multiplied by millions of people and millions of operations every second.
Platoniq: You get what you want, but uncontrolled side effects emerge.
Marta: Yes, it has side effects. And that’s where a great deal of effort has been focused in recent years: trying to understand whether there is an influence and, if so, what kind. In other words: we want to know whether certain phenomena that concern us socially—for example, increasing polarization, the loss of consensus around certain democratic values, or the erosion of support for human rights and the basic rights of particularly vulnerable populations—are being influenced by this current form of the public sphere, in which these algorithms predominate.
We’re talking about algorithms, although there are obviously other actors and elements involved. And we’re still trying to understand whether there are direct effects between them.
Emotions, Fear, and Opinion Formation
Platoniq: Going back a bit to Habermas: I don’t know if I’m mistaken, but one of your criticisms might be that he offers an overly rational view of the public sphere. How much weight do emotions such as fear, outrage, resentment, or a sense of belonging carry today in the formation of public opinion? Especially, once again, through social media.
Marta: To think that opinion is formed solely through rational means is, once again, a normative view of how public opinion—and even personal opinion—is formed.
If we have favorable conditions for rationality—so that it is not tainted or subject to manipulation—we can agree that a discussion based solely on rational arguments would be a positive thing. However, once again, empirical reality goes beyond normative models. It doesn’t always conform to them. And it’s not a matter of will; sometimes it’s simply that human beings are like that. We are both rational and emotional.
What’s happening is that we’ve focused on checking whether or not we conform to the normative model because, ideally, that would be wonderful. But it simply doesn’t work that way.
When we form an opinion about something, various components interact with one another. An emotion may be contributing to the formation of a rational idea. And this is perhaps a path worth exploring much more deeply today, because sometimes we fall into a dichotomy: “It’s just that emotions now dominate politics.” Well, not now. They’ve always been there. What’s happening is that we’ve focused on checking whether or not we conform to the normative model because, ideally, that would be wonderful. But it simply doesn’t work that way.
Specifically, regarding fear: I’m not an expert on fear either, even though I wrote my thesis on hate speech. What I can say is that it’s one of the most powerful, potent, and also most complex emotions. In reality, all emotions are complex. They don’t stand alone. They’re usually linked to other emotions, and when we talk about negative emotions, they’re typically linked to unmet expectations as well.
Platoniq: Some emotions also seem to have a greater capacity to rally people than others. I’m referring to the level of identification and, above all, thinking about certain far-right strategies: not so much creating emotions that didn’t exist as simulating, provoking, or triggering them.
Marta: I think the left has made use of emotions at times. It has relied heavily, for example, on joy…
Platoniq: I was perhaps referring to the negative ones. I should have phrased it that way.
Marta: There are certain fears, certain opinions, or certain interpretations of what’s happening that cause shame, because we’ve been raised with certain values. We end up, in a way, repressed by that shame.
When certain political factions say, “You don’t have to feel ashamed here”—obviously they don’t phrase it exactly that way—the message is something like: “You don’t have to be ashamed of feeling this way here. Here, you can be openly xenophobic. Here you can put your own people first and forget about everyone else.” That part is indeed constructed, though it’s not necessarily any less authentic than the rallying against the far right that takes place on the left.
When Deliberation Encounters Antagonists
Platoniq: That’s when we move toward antagonism.
Marta: Exactly. The problem arises when society empirically demonstrates to us that the dysfunction is structural, that not everything can be examined solely through a normative deliberative model. And at that point, the theory of political antagonism better explains some of these phenomena.
As long as we remain within a certain democratic consensus, there are problems, of course, because there are always oppressed groups, social movements that want to change things, and so on. But among the dominant actors, everyone stays within a certain boundary. There are certain things that are taken for granted. You don’t have someone in power trying to undermine the system itself.
In that context, you can indeed say, “Let’s try to reach rational agreements through communication.” This would be Habermas’s theory of communicative action: “Let’s present our arguments, negotiate, and seek agreements.” But what happens when powerful actors emerge whose goal is precisely to undermine the very foundations of the system? Well, in that case, it’s harder to know what to do. This is best explained through the theory of political antagonism. How can you maintain that space for democratic communication if you can’t manage to be more than just the other side—the one that wants to destroy it?
Platoniq: That’s where issues of hegemony come in, and perhaps a politics more rooted in conflict and confrontation—which isn’t exactly what deliberation seeks to foster. In our case, we talk a lot about creative conflict: taking positions that may go beyond pure rationality while also working on processes of listening to the other’s perspective.
Since we’re on the topic of Habermas and talking about conflict, how do you view assemblies as a kind of “mini-world,” a small-scale representation of the world? Where is the room to recognize the value of potential antagonism at certain moments during an assembly?
Marta: That’s very interesting. In fact, some of the projects I’m proposing in different settings follow exactly that line of thinking. We also need to cultivate the ability to listen to things we don’t agree with. Saying “it’s no big deal” wouldn’t be right. Of course, something serious can happen if certain discourses end up becoming hegemonic or seizing power. But it’s very important to understand the mechanisms that make those discourses resonate with people who are very close to you—who might be your family members, coworkers, or people you run into while shopping at the supermarket.
The challenge is how to develop methodologies—how to craft, construct, in a positive sense, ways to be present, to listen to that without legitimizing it and without letting it influence you.
The challenge is how to develop methodologies—how to craft, construct, in a positive sense (we were talking earlier about the political manipulation of negative emotions)—ways to be present, to listen to that without legitimizing it and without letting it influence you. That is one of the major challenges. It’s not about a “laissez-faire” approach. It’s about saying, “I’m going to listen to you, and I’m going to challenge you politically.” Because if I don’t listen to you and don’t understand what’s going on there, I can’t respond in a way that might make you change your mind.
What democratic skills do we need?
Platoniq: Returning to the topic of democratic skills: What skills do you think a person—a citizen—needs today to be able to truly participate in the public sphere?
Marta: On the one hand, the ability to discern between what is true and what is not. This involves fact-checking, but it goes beyond that. It’s not just about distinguishing between what is true and what is false, but also between what is important and what is not. And it’s also about being aware of what you’re doing communicatively: who you’re telling what to, through which channels, and with what effects. This is important because it’s what allows you to understand whether you might be unwittingly serving someone else’s tactic or strategy. The mechanisms of communication and manipulation within this antagonistic public sphere have become so convoluted that you, while defending something with every reason in the world and with the best of intentions, may be—without knowing it or intending to—aiding or playing into the strategy of a political adversary. Therefore, this capacity for discernment is essential: where you are, what you’re doing, what you’re not doing, and what others are doing—that would be one aspect. And it may well be the most important one.
Then there’s another one we’ve already talked about: being able to listen to someone who doesn’t think like you without losing your cool, so to speak. Being able to disagree and being able to listen to that disagreement.
Conflict, Hate Speech, and Discursive Violence
Platoniq: I’d like to come back briefly to something you’ve been working on quite a bit, which has to do with this potential confusion surrounding the “construction” of hate speech.
Democracy needs conflict. I’m a big fan of Chantal Mouffe, for example. But we also need to set limits on that conflict. What’s the risk of lumping things like hate speech, discursive violence, political antagonism, protest, or simply uncomfortable positions into the same category?
Marta: Since conflict is inevitable, let’s do something good with it. Of course. As for the concepts you mention, it’s helpful to think about them in light of Judith Butler’s work. In Language, Power, and Identity, she explains—through the analysis of three different cases—how harm is not caused by discourse itself, but rather by the way one actor subordinates another through language. Ultimately, it’s a matter of power and how it’s exercised discursively or symbolically. In that sense, discursive violence and hate speech could be practically synonymous. “Hate speech” is actually a very specific legal category that emerged after World War II to place certain limits on freedom of expression—very specific limits intended to protect certain social categories or particularly vulnerable groups.
That category has been expanding. Or, rather, an increasing number of characteristics have been defined that determine when a person should be protected from potential verbal attacks. For example, in Spanish law, one of those grounds is ideology. It does not say “left-wing ideology.” It says ideology. In other words, a person whom someone calls a “fascist” could file a hate speech complaint if they claim the remark was made with the intent to insult them. That is the formal reality of the “hate speech” category.
I believe that when we say we are against hate speech, we are often referring precisely to discursive violence in the sense that Butler gives it: we criticize the use of words not for dialogue, mutual understanding, or the building of consensus, but to undermine all of that and to display power, domination, and political antagonism. And in that sense, all the concepts you’ve mentioned are related.
Platoniq: On that note: How can we create spaces where disagreement can be expressed freely without leading to harassment, exclusion, or silencing—both online and in person?
Marta: In a context like that of deliberative assemblies or the kinds of spaces you create, for example. These are ideal spaces for practicing the kind of listening we need to distinguish between what is a political position and what is part of that more performative dimension—that aggressive and exclusionary packaging. It’s one thing to hold a political position—such as, “I think indigenous or local people’s rights must be safeguarded before welcoming a new population,” a view with which you may strongly disagree for very good reasons—and quite another to dehumanize the excluded group in an extreme way. And it’s not that dehumanization is any less serious simply because it’s implicit. In fact, it’s more dangerous because different audiences can digest it more easily. What I’m saying is that we need controlled spaces from which we can examine that difference in detail and, from there, work on political and discursive proposals that can undermine the ones being put forward by the far right.
Embedded AI and Affected Communities
Platoniq: I’d like to start with a brief introduction before inviting you to the session on the experiment we’re working on. We’re developing an AI prototype—we’ve called it GRID—that operates with multiple agents possessing situated knowledge. Each agent has its own personality, but also a set of training materials specific to that agent. In addition, we ask specific organizations—whether we’re talking about migration, homelessness, etc.—to participate in that training. In other words, we conduct the training with an organization by physically meeting with them. It’s a fairly analog process.
We also have a personality called “the philosopher,” which my colleague Cristian is in charge of. We have—I don’t know—fifty years’ worth of fanzines and unpublished books, things we don’t want to simply hand over to Anthropic for them to buy and dissect. These are very ad hoc materials. Therefore, none of these agents contains “all knowledge.” They are very specific corpora, with a very clear bias: to bring to light feminist, anti-racist, intersectional, or minority perspectives. The question is this: how do we do this without turning them into a quota within the corpus we give to that agent? And in this sense, how do we ensure that the AI doesn’t “speak on behalf of” those communities without making the process through which we construct that knowledge visible?
Marta: Ideally, we’d want to involve those communities from the very beginning of the creation process, in the co-design phase. Not just in training, but also in the creation of categories.
Platoniq: In the ontology.
Marta: Yes, exactly. In the ontology. Because there can also be a knowledge gap here. And not just regarding anti-racist or feminist communities, but regarding any citizen. Even among people who work in research, there are those who know more and those who know less about large language models. Not everyone has the same level of knowledge. And I think that’s another area where we can intervene: making it possible to make decisions at all levels of the agent’s development even without all the technical knowledge.
We’d need to create mechanisms to facilitate that. And then, of course, truly take contributions into account and ensure they’ve been properly incorporated, with mechanisms so that participating communities can validate or, if necessary, modify design decisions.
Red Lines for Deliberative AI
Platoniq: In our case, we have one version that is completely separate from any ChatGPT-type LLM and another that does use that kind of engine. The standalone version can run with minimal resources on a local installation. It’s quite a bit slower, of course, and the types of responses are also different.
We’re trying to build this precisely to address a problem: virtually no citizens’ assembly tackles issues that could potentially be truly contentious. And I’m not just talking about immigration or violence. It could be something as seemingly simple as waste management in a city—if no politician wants to use that issue as a topic for deliberation.
What we want to create is that phase that precedes the moment when an institutional mandate with a closed-ended question already exists. Because there are usually a number of “red lines” that no one wants to cross, either politically or in terms of regulations. We want to open up a space beforehand, intended more to set the agenda that will later make it possible to convene a more representative assembly.
Perhaps the most general question here—one that goes well beyond our specific case—would be this: regarding experiments like the Habermas Machine, if the goal is not for AI to deliberate on our behalf, but rather to help us deliberate, disagree, and act more effectively, what red lines would you draw?
Marta: One of the risks I see in the Habermas Machine and other similar experiments is that what is actually deliberation carried out by machines is passed off as genuine deliberation. On the other hand, if it’s done within a framework linked to real citizens’ assemblies—and if it’s at our service, at the service of people—I think that could be interesting.
Then there’s the risk that the tool itself might be steering the conversation too much. This could happen due to poor design, an intentionally harmful design, or simply as a result of an unintended consequence. In the latter case, it would also be a design issue that could likely be remedied by keeping the design open to improvements. I’m not sure if this is exactly a red line, but it’s certainly something that would need to be closely monitored.
Platoniq: In a scenario where people get their information through AI, do you think we’ve overly individualized the problem of misinformation? In other words, are we asking citizens to be more critical without simultaneously demanding changes to the infrastructures that organize information?
Marta: Absolutely. Disinformation is already just another structural element of the digital communication system. It is the digital advertising system itself—in the sense of exposure and visibility—that puts it into circulation. And I say “the system” because there are various elements within it that interact in such a way that they end up spreading it on a massive scale. For example, those who rely heavily on view counts—such as influencers, some politicians, and some journalists—tend to slip in some false or misleading information from time to time to keep the flame of engagement alive. The monetization model doesn’t distinguish between the quality of what’s being said; what matters to it is selling ads to the people viewing the content.
Filters can be put in place, of course, such as those used by content moderators and AI trainers, whom companies rely on to prevent much worse content from reaching us. The problem is that putting filters in place to block misinformation isn’t that easy, because much of the misinformation circulating today isn’t false information whose falsity can be verified; rather, it’s the “cherry-picking” variety—that is, it consists of focusing on the most favorable aspect of a fact while ignoring or concealing the rest. The line separating this technique of information manipulation from the deliberate framing used in political communication is very fine. False positives could seriously undermine freedom of expression and deliberative discourse.
The problem is that media literacy has its limits when we’re dealing with increasingly opaque technologies, such as recommendation algorithms.
That is why it is so important for people to have strong media literacy skills to identify fabricated and self-serving lies. The problem is that media literacy has its limits when we’re dealing with increasingly opaque technologies, such as recommendation algorithms. Furthermore, since disinformation is mixed in with other content, avoiding certain news sources comes at a cost—it means missing out on things that help us stay connected to the world. It’s not so easy to figure out what the solution is; there doesn’t seem to be a single solution. What seems clear is that the engagement model of commercial digital spaces—and the fact that there are people who want to monetize them at any cost—have a lot to do with the problem.
Platoniq: Based on your work on antagonism, what would AI have to do to avoid domesticating political conflict?
Marta: That’s a very difficult question. I assume you’re referring to an AI agent developed for this specific purpose—like the one you’re creating at Platoniq—rather than ChatGPT or some other generic model. In a controlled environment like that of deliberative citizens’ assemblies, I imagine that whoever is there participates without disrupting the most basic dynamics, so here we’d be talking more about agonism than antagonism. I think AI can be very useful there to foster active listening without stifling dissent. AI could help, for example, by synthesizing political positions, highlighting the areas of agreement and the areas of disagreement between them. It could also rephrase people’s positions without betraying the underlying idea, in an effort to foster greater convergence—which is somewhat similar to what the Habermas Machine aims to do.
What would AI have to do to avoid defusing another, more antagonistic type of conflict? Well, I don’t know, to be honest. I wish there were an AI that could defuse this kind of conflict, but it doesn’t work that way. There are decisions that AI simply cannot make—that much is clear. For example, in an emotionally polarized context, do you show the summary of positions to both sides of the debate, to just one side, or only to those people whom the algorithm calculates to be more moderate? Because if you show it to everyone, you might end up with even greater hostility between groups, which in the end is worse than continuing to ignore each other and failing to find common ground. It’s one thing not to tame the conflict, but it’s quite another to exacerbate it. We don’t want that. And how do you determine who the moderate people are? There are certain ethical and political limits regarding these methodologies that we’d have to take into account. If we truly want to redesign things based on democratic criteria, there’s a preliminary discussion we haven’t had yet—namely, whether we want algorithms and AI to measure these kinds of things, to engage in a certain kind of social engineering with us. Perhaps it depends on factors such as who controls that AI. Many people will object to this, but it’s surely necessary if we want to develop a massive public sphere —an alternative to the commercial one— dedicated to fostering democratic understanding. As I say this, I have many doubts about it because it’s a nascent and highly complex issue. Maybe it’s not what we need? Maybe it’s better for AI not to do any of this? Maybe it’s not even possible given the planet’s limitations? What’s certain is that these are questions we’ll need to be able to answer at some point.
