“In the War of Narratives, Our Ultimate Defense Is Discerning Truth From Lies”: An Interview With Mariarosaria Taddeo

Share
Share
10
17.11.2026
Disinformation threatens democracy, and the only way through it is to ground public debate in objective facts, so that the decisions we reach are ones everyone can live with. This is the view of Mariarosaria Taddeo, professor of digital ethics and defense technologies at the University of Oxford’s Oxford Internet Institute (OII). Taddeo is one of the leading voices on artificial intelligence regulation and so-called autonomous weapons. She is convinced that the self-regulation pledge recently announced by major U.S. tech companies is nothing more than a smokescreen. In this conversation with Páginas Digital, she addresses the challenges posed by AI and explains why, in her view, AI will never become sentient. Photo: University of Oxford

The digital world is pouring fuel on the fire of fake news and disinformation. We’re living through a full-blown “war of narratives,” and reality itself seems to have vanished. Is there something within us that could possibly withstand such power?

Yes, but less and less. We have to rely on our autonomy—on our ability to discern truth from lies. But the reality is that we live in a world where the goal is often to flood the zone. And it doesn’t matter whether the flood consists of accurate information, disinformation, or plain misinformation—it is, false content.

We hear a lot about “cognitive warfare.” In Italy, it’s a hot topic right now, with elections coming next year and fears that Russia, or certain political actors, might deliberately wage that kind of war. But this isn’t just about spreading lies. It’s a revealing example, because I believe the real goal is to create what I call “epistemic vertigo”: a state in which neither individuals nor communities can tell what is true from what is false. It doesn’t matter whether people believe vaccines are harmful. What matters is that there is no longer a shared set of facts from which to begin the conversation.

This is the problem for democracy, because objective facts are the starting point for public debate and, therefore, for political decisions. So no, I don’t think individuals can hold the line on their own. This is troubling for individuals and democratic societies alike. We need much tougher rules—or, rather, rules that are actually enforced—to protect everyone’s rights and autonomy.

So can we still develop the ability to think critically?

Absolutely. Taking the time to go back to original sources, thinking things through, and digesting information is essential. But that time has to come from discipline and habit. This is basic stuff, really. A popular-science article might take 15 minutes, sometimes half an hour. A biology article might take an hour if I want to really understand it—and that’s popular science, not a research paper. For starters, having an hour to spare is a luxury these days because, unfortunately, the worse the economy gets, the harder we have to work. So if you want to carve out that time, it has to become a habit.

The problem is that we’ve gotten used to intermediaries. A journalist reports a fact; an influencer reads the story and records a video; someone comments on the video; I read the comment—and by then, I’m miles away from the original source. We’ve gone from a “war of narratives” (I really like that expression) to one long game of broken telephone. There is where it becomes a problem.

You’ve stressed the need to deliberate. What do you mean by “deliberating”?

“Deliberating” is a big word. Habermas, who died recently, argued that liberal democracies deliberate under the principle of universal agreement. Opposing sides have to come to the table and reach a solution everyone agrees with: a compromise in which both sides give up roughly the same amount. Deliberation requires expertise, information, and participation. By deliberative democracy, I don’t mean direct democracy. That’s not the point. We have our institutions, whether we like them or not.

We need social dialogue, as Habermas phrase it

Of course. It’s the public debate we’ve always had. Public debate becoming polarized, fragmented, and finally disappearing is a current development. In Italy, Emma Bonino died recently. She was a driving force behind public debates on abortion and on legalizing cannabis. Sometimes, as with abortion, a law gets passed; sometimes nothing happens because the debate doesn’t gain enough traction. But the debate was there. There were forums; there were newspapers. Little by little, that debate has faded, and in my opinion, democracy has taken a massive hit. When public discussion dries up, politics starts to seem pointless, and at some point those in power check out. Then, they use the machinery of democracy simply to consolidate themselves and reinforce their own power.

Could AI’s sheer growth in scale ever tip over into self-awareness?

No, because artificial intelligence manipulates symbols without giving them meaning. If I say “red,” no images form in an AI’s “mind” the way they do in ours—images tied not just to an object, but to sensations, memories, and perceptions. AI lacks that capacity, not because it lacks sensors, but because it hasn’t spent roughly two billion years evolving the way our brains have, or an ant’s brain, or a whale’s. Evolution gave living brains the ability to connect a symbol with its meaning. That’s why, when a mom or dad says “cup” to a child, the child eventually recognizes the cup. Without the “hardware” that lets the child make those connections, Mom and Dad could say “cup” until they ran out of breath. AI is that child without the right hardware to attach meaning to symbols on its own. We can teach it, but only one association at a time; it does not evolve in that sense. And if it can’t understand the world around it, it can’t understand itself either. I’m sure of that.

Now, if someone wants to tell me I’m wrong, fine. But the burden of proof is on them, and until they find it, I’m right.

Going back to the previous question: What does it take for a person to determine if AI content carries semantic value or not? 

Well, look: training and education, for sure. We can’t keep using it as naïvely and mindlessly as we have so far. But it also takes self-care, because the real damage AI can—and will—do to individuals is to wear down their cognitive capacity: our ability to discover, and the joy of understanding. I don’t mean the bookworm who loves to study and never comes up for air. I mean the joy of figuring things out—how a refrigerator works, how the two pieces of a vase fit together. Give that up, and we lose something important. In Europe, we have Kant, who taught us Sapere aude: “Dare to know; have the courage to think for yourself, and stop leaning on your teacher.” It’s no coincidence that Kant also gave us the concept of dignity, which is the foundation of Europe.

If we lose the joy and the courage of understanding—of making sense of the world and acting on it, which is uniquely human—then, simply put, we risk ending up like the humans in Pixar’s WALL-E. They’ve depleted planet Earth; now, they live on a spaceship, gliding around in hover chairs; they’ve become obese because they never move; and they depend on AI to keep the drinks coming, igniting the never-ending cycle which only makes them more dependent. It’s simply a cartoonish image, but the warning is real: be careful. Our lack of engaging with information—about doing the work of understanding it—threatens the habit of cognitive effort. And cognitive effort is what keeps us capable of making our own decisions. Without understanding, we can’t decide. And without decision-making power, we’re no longer autonomous.

In recent weeks, there has been a lot of discussion about an artificial inteligence apocalypse. Is it a possible risk? 

No. I can say this with total confidence. You know the rattle you dangle in front of a kid while the nurse gives them a shot? Same thing. It’s a smokescreen. These alarms are sounded often, and for no good reason. First, there is no science behind them. AI is still a machine that runs extremely complex statistical calculations on huge amounts of data at high speed. It doesn’t understand what it’s doing. And because it doesn’t understand what it’s doing, it poses no apocalyptic risk: it will not develop the intention to kill us or wipe us off from the face of Earth. Problems will come, including systemic ones, and I’ll get to a few in a minute. But they’ll come because we built or used AI inadequately.

If AI ever does kill us, it will be because some fool, somewhere, got something terribly wrong. And the risks we’ve been hearing about lately come down, in my view, to Anthropic and companies like it needing to sidestep regulation and push self-regulation, so they can write the rules of the game themselves.

So this is a marketing strategy—a way of telling everyone: “It’s going to kill us in 10 years; we have to stop it. Let us be the ones to stop it; we know how.” There is a real risk, and it’s far more troubling than Armageddon. We’re building an extremely powerful technology, one with enormous potential to improve humanity’s future. It isn’t a bad technology; it’s a wonderful one. But it’s now moving very fast and working very well, and we haven’t built guardrails to match. I’m talking about two specific stages: validation and verification. They’re how we keep control of this technology—not only when we use it, but also while we build it. People talk about “alignment,” but that term carries some misleading baggage. What it’s really about is control. And our ability to control AI is shrinking. Why? Because we jumped into a race to build instead of into a race to regulate and govern, after being told for years that regulation slows innovation. That’s why development is now running ahead of us.

What risks I do see: A few days ago, I heard that an alert went off and a U.S. aircraft was on the verge of bombing some Chinese ships. This is one example, but there are others. AI is extraordinarily good at finding vulnerabilities in computer systems, and it finds them in seconds, when humans usually need days. No computer system is secure by default. The old understanding was that if you had a system, an intranet, or a piece of software, sooner or later someone would break in. But that “sooner or later” mattered a great deal: if it meant months, or even days, I could prepare. Take away the “sooner or later,” and I can’t. And if Anthropic’s AI can do it, and so can open-source models—the free Chinese ones anyone can download—then anybody can do it.

No computer system is secure, yet we are an information society, a digital society. We’ve wired computer systems into everything: kids’ electronic school records, hospitals, courts, the defense sector itself. Worse, they run critical infrastructure: water systems, railroads, airports. Now imagine this AI, which we have little control over, falling into the hands not of a major political power but of a mid-sized criminal organization. What happens then? The damage isn’t just economic or physical; it impacts the trust that holds our society together. If we can’t trust that our bank accounts, schools, or hospitals are secure, we have a problem. That’s the risk I see. It isn’t inevitable, but it’s concrete—and systemic.

After this summer’s attack on the Hugging Face platform, OpenAI said the AI had made an autonomous decision. When AI makes an autonomous decision, is it because its developers chose not to set limits on it, or because it’s bound to end up making its own decisions?

Words matter. AI doesn’t “make decisions.” We say that because it’s convenient shorthand, and because AI is built by computer scientists, not philosophers—and computer scientists can be careless with language.

AI makes inferences, which is very different from making decisions. In the Hugging Face case, the outcome was the result of two things: first, the task the AI was given, and how it was designed; second, the fact that OpenAI’s developers had built a fragile testing environment, one that let these agents obtain credentials and connect to the network. Read the postmortem, where the engineers explain how the way they set up the test environment allowed the agents to reach the internet, and it’s obvious that AI didn’t decide anything. It just did the math—step by step. That matters. First, it places responsibility squarely on the people who build this technology. Second, it reminds us of something we already knew: these are machines that learn from their interactions with their environment, and as they learn, their objective is to optimize a function—the goal they’ve been handed.

How they reach that goal can’t be explained step by step, and the more complex the environment the agent works in, the harder that gets. We used to say an algorithm is a recipe. It isn’t anymore. There’s a well-known example in the literature: a team of engineers builds an AI model, gives it a building’s floor plan, and says, “Find the fastest way to count the windows in this building.” The AI’s answer: “Bomb it. Zero windows.” The AI has no idea what “bombing” means. It isn’t evil; it just figures, “Destruction, zero windows. Done.” The engineers never saw it coming, because the more complex the model and the larger the dataset, the better AI gets at finding patterns.

Think of a sharp lawyer: the ink on a new law is barely dry, and he has found the loophole. So yes, the question is how we design these systems, and how we regulate them. And we have to let go of the outdated idea that we can hand AI all the rules up front, and count on them to be followed  all the way through. That’s not going to happen. We need constant, ongoing oversight. These machines learn but don’t understand, so who knows what they’re really learning?

Let’s turn to self-regulation. Why do American companies insist self-regulation is the answer? Do we need it?

No. We’ve seen this movie before, with privacy. The big companies—Facebook, Google, all of them—saw regulation coming and said, “Let’s set our own digital ethics principles,” and so on. Those principles didn’t last. When Trump took office, Zuckerberg announced he was getting rid of the systems that filtered and monitored fake news on his platforms. Self-regulation has turned out to rest on a deeply cynical approach. Do we keep going along with it? We’ve learned that these platforms are designed to hook kids and keep them hooked, and it’s already very serious with social media. Can we really leave it to companies to monitor a technology that is part of our infrastructure, that may pose a security risk, and that by now makes decisions for us in every area of life? Banks don’t approve a mortgage without running the risk through AI. Hospitals will be using it for diagnosis sooner or later. In defense, it’s already being used for life-and-death decisions. I think that’s absurd.

You’ve studied AI and autonomous weapons in depth. What is their role and the extent of their use in places like Gaza, Iran, and Ukraine?

Much further along than anyone expected. The first report on the possible use of autonomous weapons came from the United Nations in 2021. It’s a somewhat shaky report, though, because it’s hard for an outside observer to tell whether a drone is truly autonomous , or partly autonomous and partly remote-controlled. An autonomous weapon is one that makes every call on its own, all the way up to killing a person.

We now know these weapons have been used in Ukraine. We know for a fact that the Russians have used them, because they’ve been recovered, and they behaved in ways that allow us to say they were fully autonomous. Before that, the Ukrainians had said unofficially that they, too, had used fully autonomous weapons and had killed two Russian soldiers with them. And we need to be clear-eyed about this: everyone is developing autonomous weapons. There are no pure villains or heroes here. That was obvious even before Ukraine, if you look at the UN group that has spent 13 years working on lethal autonomous weapons. It wrapped up its work at the end of August. In 13 years, all it managed to do was agree on a definition of lethal autonomous weapons—two years ago. No limits, no standards, no regulations. That’s the problem with AI in defense: every country is adopting it; every NATO member is building AI into its defense systems.

Iran is very advanced in this area, and Israel is in a league of its own, with truly cutting-edge digital skills and capabilities. Nobody is doing the serious diplomatic and political work of figuring out how international humanitarian law applies to these weapons. International humanitarian law exists and is applied, but applying it takes work. It’s a bit like applying biology to medicine: biology alone doesn’t cure anyone until it’s turned into standards, procedures, and so on. Right now, we have nothing.

In late February 2026, a U.S. missile strike on Minab, Iran, destroyed a school, and killed more than 150 girls. Anthropic’s AI system, Claude, was used in that attack. Who is responsible in a case like this?

That’s exactly the problem. Given the total lack of standards we talked about, we haven’t even decided how to assign responsibility when AI is part of the decision chain and something goes wrong. On top of that, AI isn’t built the way an ordinary machine is. It’s built in a highly distributed, iterative way: one team of engineers builds the model, another handles the data, another runs the tests; then there’s a first round of interaction, a review, a second round. Meanwhile, engineers come and go, the data changes, the law changes. So this machine—which we don’t control well and which is a black box, because we can’t lift the hood and see how it reaches its outputs—gets plugged into a decision-making process: picking targets. That process is distributed, too. There’s a chain of command, there are roadmaps, there are deadlines. Something goes wrong, someone makes a mistake, and 600 children die. Whose fault is it? With no rules and a structure this decentralized, AI becomes the perfect setup for blame shifting. That’s very serious in general, and extremely serious in wartime, because if you can’t assign responsibility—moral responsibility above all—you’ve taken down the last guardrail between war and atrocity.

War is always horrific. But international humanitarian law, and the ethics of war, appeal to the conscience of the people who fight. The judges at Nuremberg made that point in the opening pages of their judgment: they had not put the Third Reich on trial in the abstract; they had tried the people who served it, as individuals, because in war, responsibility is personal.

Long before drones decide to kill, and long before Terminator-style robots—which will never happen—we’ll have already turned war into an abomination, because we’ll have stripped away the last safeguard that lets human beings stay in control: see the target as a human being, and hold back. That is deeply troubling.

With Israel’s use of AI in Gaza, the picture gets even darker. In Iran, it may have been a mistake: the maps were out of date. In Gaza, according to reports—not official sources—Israel used a suite of AI systems to identify both physical and human targets. One of those systems had a false-positive rate of 10 percent. Thirty-seven thousand targets were identified, 37,000 were struck, and 3,700 people died for nothing. The decision window, again according to the reports, was 10 seconds: human operators had 10 seconds to sign off on each strike. Ethics aside, these questions cry out for a regulatory answer. Is 10 seconds enough, or should the law require five minutes, or two? That legal vacuum is real. Is a 10 percent false-positive rate acceptable to society? Can we live with it, or is it beyond the pale? As an ethicist, I’ll leave those questions aside, but as a citizen, I’d at least like a legal, regulatory answer—and there isn’t one.

Palantir has been working with U.S. Immigration and Customs Enforcement (ICE) to detain and deport migrants in real time. Has AI built a surveillance system that’s watching all of our data right now?

Well, not necessarily everyone’s. But mass surveillance isn’t some far-off risk, because we all live in the digital world, and in the digital world, we leave a trail. It’s like Hansel and Gretel’s breadcrumbs: everything we do—from using a credit card to buy this newspaper instead of that one, to seeing this movie instead of that one—makes it possible to track us and profile us.

To me, Palantir’s support for ICE is no different from Trump’s use of AI in the early days of the war in Iran. Remember when he went head-to-head with Anthropic and declared, “I am morality,” like a latter-day Louis XIV? Both decisions spring from the same root, because international humanitarian law and democracy spring from the same root. Both were born after World War II from the idea that state power has limits—beyond its borders and within them. When a state crosses the line abroad, sooner or later it crosses it at home, too. International humanitarian law and liberal democracy are twins: when one sneezes, the other catches a cold. I think that matters, because for those of us in the West who are lucky enough never to have lived through war, defense always seems like somebody else’s problem. Whether or not Putin attacks, we need to stay alert, because defense is where the state first flexes its muscles.

Share

Comments 0

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.

Related News