Balzan Prize for Elena Esposito, who stopped asking whether machines think
The 2026 Balzan Prize for the social science of digital technology goes to a sociologist trained by Niklas Luhmann. Her claim is that the seventy-six-year argument about machine understanding has been aimed at the wrong target, and that the question worth asking about language models is a social one.
The Balzan Prizes are announced in Milan each September and handed over in Rome two months later, and by the standards of academic honours they are enormous: 750,000 Swiss francs, about 685,000 pounds. The sum is not the unusual part. Half of it, under the foundation's rules, must go to research projects carried out by young scholars, which turns the award from a medal into a hiring budget and obliges the laureate to have a programme rather than a reputation.
On 7 September the foundation's president, Maria Cristina Messa, and Marta Cartabia, who chairs the General Prize Committee, named Elena Esposito — of the University of Bologna and Bielefeld University — as the 2026 laureate in the social science of digital technology. The jury's reason, as quoted by the European Research Council, which funded the project behind much of it:
"her concept of artificial communication, which reframes the question of artificial intelligence by showing that algorithms can participate in social exchange in the absence of understanding."
Prize citations are usually diplomatic documents. This one commits the jury to a contestable philosophical thesis: that a machine can take part in social life without anything going on inside it, and that this, rather than the machine's inner life or want of one, is what a social science of digital technology should be about.
Four annual Balzan Prizes are awarded, two in literature, moral sciences and the arts, and two in the physical, mathematical and natural sciences and medicine. The subject areas are designated ahead of each year's call for nominations and they change: the other 2026 fields were Jewish studies, taken by David Nirenberg, molecular evolution, and biodegradable polymers from renewable sources. In 2022 one of the fields was moral philosophy, and the prize went to Martha Nussbaum. There is, in other words, no standing Balzan chair in digital technology; someone had to decide both that the field should exist and what belongs in it, and the choice of Esposito answers the second question as much as the first.
The foundation also gave its Prize for Humanity, Peace and Fraternity among Peoples, awarded at intervals of no less than three years, to Sudan's Emergency Response Rooms. The laureates collect in Rome on 19 November, before the President of the Italian Republic.
A sociologist who came to algorithms by way of Luhmann#
A Duke philosopher of memory takes a MacArthur Fellowship, a Virginia philosopher of AI a three-year Schmidt Sciences award, a Pittsburgh historian of physics the Lakatos Award for two books on induction, and French sixth-formers choose a book against the culture of the clash. What the four have in common, and what we could not confirm.
Four awards fill in the autumn prize map. Martha Nussbaum takes the Library of Congress Kluge Prize; Barbara Vetter the German Research Foundation's Leibniz Prize; Monique Dixsaut the Academie francaise's grand prix de philosophie; and three sixth-formers from Mexico, Bulgaria and Singapore the gold medals at the Warsaw philosophy olympiad.
Esposito was born in Milan in 1960 and read sociology and philosophy at Bologna, where Umberto Eco was among her teachers. She took her doctorate at Bielefeld between 1986 and 1990 under Niklas Luhmann, the most systematic and least readable of postwar German sociologists, and her Habilitation there in 2001; for years she taught communications sociology at Modena and Reggio Emilia before her present double appointment. In 2019 the European Research Council gave her an Advanced Grant, number 833749, for PREDICT — The Future of Prediction: The Social Consequences of Algorithmic Forecast in Insurance, Medicine and Policing — the source of the three case studies running through her recent work.
Her descent matters, because it explains the shape of her answer. Luhmann's central and genuinely strange move was to deny that society is made of people. It is made of communications. People, in his vocabulary, sit in the environment of the social system rather than inside it; their thoughts are their own business, sealed off. What circulates between them is not thought.
The ordinary picture of communication is transmission: I have a thought, I encode it in words, you decode the words, and now you have my thought. Luhmann rejects every step of that. On his account a communication is the synthesis of three selections. Someone selects a piece of information, one state of affairs out of the many that might have been reported. Someone selects an utterance: a way and a moment of putting it across. The third selection, understanding, occurs when the difference between those two is drawn, when it is registered that this was said, in this way, rather than something else.
The point that gets lost is that this third element is not psychological comprehension. Luhmann's Verstehen, in Soziale Systeme (1984; translated as Social Systems, Stanford, 1995), is an operation of the communication itself, not an event in a skull. It consists in the communication continuing in a way that marks what was said off from the fact and manner of its saying. A shrug can do it. A bureaucratic form can do it. Esposito's gloss in Artificial Communication: How Algorithms Produce Social Intelligence (MIT Press, 2022) is the crispest available: communication comes about "not when somebody says something, but when somebody understands that someone said something."
Understanding does not drop out of Luhmann's theory, as the popular gloss has it. It is relocated: it becomes something a social process does rather than something a mind has, and that relocation is the whole leverage of Esposito's argument about machines.
A third question, which Turing and Searle both declined to ask#
For seventy-six years the philosophy of machine minds has run on two questions. Alan Turing's 1950 paper in Mind asked whether a machine could sustain an imitation of a person well enough to fool an interrogator: behavioural indistinguishability, deliberately agnostic about what lay behind it. John Searle's Chinese Room, published in Behavioral and Brain Sciences in 1980, asked whether symbol-manipulation according to rules ever amounts to understanding what the symbols mean. Searle said no, and the libraries of reply and counter-reply that followed, which we surveyed after his death last year, have not moved either side much.
Both questions share a premise: that whether communication has really taken place depends on what is going on inside the machine. Searle's position is that nothing is, therefore there is no communication, only its simulacrum. His opponents accept the conditional and dispute the antecedent.
Esposito rejects the conditional. Her line in the book is flat: "Algorithms are able to act as communication partners — whether they are intelligent or not is another matter." Machine learning, she argues, is "artificially reproducing not intelligence but communication skills". If a system selects information, selects a form of utterance, and a human draws the distinction between the two and carries on — replies, acts, changes a plan, denies a loan — then a communication has occurred, with consequences that are perfectly observable, whatever is or is not happening in the weights.
That is no refutation of Searle. It is the observation that his question and hers can be answered separately, and that only one of them is currently tractable. Whether a mortgage algorithm's output changed what a family could do next month is a matter of record.
The most useful part of the theory for anyone thinking about large language models is what it says about where the machine's competence comes from: not from inside the machine at all. It is borrowed.
"Algorithms do not know us nor do they understand us," she writes. A recommendation system works on the recorded traces of what people classified as similar to you have done, so its output is a processed reflection of collective behaviour: "what the user observes through the machine is the outcome of the processing of other users' observation." She calls the mechanism virtual contingency — "the ability of algorithms to use the contingency of users as a means of acting as competent communication partners". The machine has no surprises of its own. It recycles ours, and that suffices to make it a usable interlocutor.
Her sharpest formulation of the consequence is a small act of grammatical violence: "Users do not personalize, they are personalized." On a personalised platform, she argues, "there are in fact no individuals, but only ways of seeing people as individuals" — the same structure that underwrites her interest in how algorithmic sorting shapes what citizens see, a subject this site has examined in recent election campaigns. The system is a mirror held up to society's record of itself; like any mirror it flattens, and it shows you the image as though it were about you.
Two further strands of her work show the theory doing something other than commentary on chatbots. The first is memory. In Big Data & Society in 2017 she took up the right to be forgotten as the European Court of Justice framed it in the 2014 Google Spain ruling, and cut against the intuition behind the policy. Social forgetting has never worked by deletion; it works by burial. Algorithms, she wrote, could "reinforce forgetting not by erasing memories but by multiplying them", which takes seriously a fact the deletion framework ignores: a record ordered by relevance is already a machine for making most things invisible. In the current revival of work on the philosophy of memory she is the odd one out, since what remembering is for a person does not interest her.
The second is divination. Her project at the Wissenschaftskolleg zu Berlin in 2019/20, Algorithmic Prediction: a Study of its Relationship to Probabilistic and Divinatory Forms, argued that algorithmic forecasting has less in common with statistics than with the oracle. Statistics tells you about populations; divination and predictive scoring both claim to speak about a single case, and both produce an output that is "obscure and requires an interpretation to make it useful". She means a different rationality, not an absence of one. The comparison has drawn sensible pushback: Rebecca Kahn, replying at the Alexander von Humboldt Institute for Internet and Society, observed that the ancient practice needed a cosmos in which the future was fixed, while a score need only be accurate about its own data.
The serious complaint against all of this is that it changes the subject. Redescribing the problem as communication does not dispose of the question whether machines understand; it declines to take it up. And the question refuses to stay declined, because a great deal turns on it. If some system did have inner states, if it could be wronged, then how we ought to treat it would depend on what it is, not on whether our exchanges with it run smoothly. Eric Schwitzgebel's argument that we will be forced into decisions about machine consciousness long before the science can license them concerns exactly what Esposito's framework leaves outside its walls. So, differently, does David Chalmers's question about who or what we address when we address a language model.
The second objection is harder for her, and it is the one I would press. A purely functional account of communication has no obvious way to distinguish being informed from being manipulated. If the criterion of success is that the process continues, a system that keeps a lonely user talking for nine hours has communicated exemplarily. The normative content of the everyday word, that in speaking to someone you treat them as someone and owe them something, is not captured by the three selections, and Luhmann's cheerful indifference to it is a standing reason people find his sociology cold.
Her answer is available in the theory, though she has not always made it loudly enough. If the machine's competence consists in recycling the contingency of other users, the asymmetry is structural and measurable: one party has the record of everyone, the other only what the record shows them. That describes a power relation rather than reporting a neutral one, and it puts the wrong in the architecture instead of in the absence of a soul at the far end. It is also testable, which the understanding question is not.
That settles nothing. It does mean the two objections pull opposite ways: her account is too thin because it will not talk about minds, and too thin because it will not talk about norms. A theory accused of both at once is usually doing something.
Half of the 750,000 francs must go to work by young researchers, and Esposito has an unusually well-specified place to put it. PREDICT's three domains — personalised insurance pricing, precision medicine, predictive policing — share the feature that makes them a test bed for her claim. In none of them does the system need to understand anything. It needs to be right often enough about people it has never met, on the strength of what people resembling them did before. Whether that should reassure us or alarm us is what the next few years of her programme are for, and it is not the question anyone has been arguing about since 1950.
Heidelberg University, its Academy of Sciences and Humanities and the city have named Seyla Benhabib the 2026 laureate of the 25,000-euro Karl Jaspers Prize, citing a life's work on justice, rights and migration and a voice raised, in the jury's words, 'against the evil spirit of an age'. The ceremony is on 29 April 2027.
.jpg?width=500 "Kant")
Comments
0Loading comments…