Interview with Luciano Floridi, the John K. Castle Professor in the Practice of Cognitive Science and Founding Director of the Digital Ethics Center at Yale University, and Professor of Sociology of Cultural and Communication Processes at the University of Bologna.

By Barbara Rosario

Professor Floridi, you are widely regarded as one of the world’s leading contemporary philosophers, the founder of the philosophy of information and one of the foremost thinkers on the digital revolution.

I’d like to begin our discussion with your latest book, which won the Premio Strega Saggistica Internazionale (congratulations!): La differenza fondamentale. Artificial Agency: una nuova filosofia dell’intelligenza artificiale. What is this fundamental difference? Could you outline the book’s central arguments?

The difference is between acting and understanding, and the whole book turns on it. Our tradition has taken it for granted that acting well and on purpose requires some understanding of what one is doing. Behind effective behaviour, there has to be an agent who grasps the point of the action, if only a sheepdog rounding up sheep. AI has broken that assumption. Machines have always done what they do mindlessly, and what is new is the reach, since the tasks handed to them are the ones that until yesterday needed a mind. Systems now translate, diagnose, check legal documents, do advanced maths, write a sonnet, and draft a holiday budget without understanding anything whatsoever. I call this Agency without Intelligence, this is how I read AI.

Three arguments follow. First, the success of AI is not the long-announced marriage between machines and intelligence but the divorce between the capacity to complete a task successfully and the intelligence that completing it would demand of us. My phone plays chess incomparably better than I do and has the understanding of my grandmother’s coffee pot. Second, the divorce does not shift responsibility at all; if anything, it increases the only one available, which is ours. A system that acts without understanding cannot answer for what it does, and the answering stays with the people who designed, bought, and deployed it, however hard it becomes to say which of them. The task can be delegated. The responsibility cannot, and mind that I do not say “may not”, although this follows logically. Finally, the divorce teaches us something about ourselves. We are the ones who decide whether an output is worth producing at all, and why, when, how, and at what cost. We should see ourselves more as designers and less as executors. 

There is a fourth point I care about. We are digitising the world to fit AI, rather than making AI intelligent enough to fit the world. That is a choice, and it is ours. 

Warnings of a looming “jobs apocalypse” seem to have intensified recently. Your book also examines how AI is reshaping the world of work. What is your perspective on these changes? How do you assess the nature and potential consequences of these changes?

Warnings of that kind have a long history and an abysmal record, so I would look at the data before the headlines. Take the case everyone cites. Amazon has cut about 30,000 corporate roles in two rounds, and the first announcement put AI at the centre of its explanation. However, the company itself has since said that AI was not the reason behind the vast majority of those cuts. Reuters reported something duller: overheads and the correction of pandemic over-hiring. Those cuts are roughly 10 per cent of Amazon’s office staff and under 2 per cent of the people it employs worldwide, most of whom work in warehouses. Both figures circulate, and the choice between them usually tells you what the writer wanted to prove. 

The real impact lies in the kind of work rather than its quantity, since the evidence I find credible indicates that the skills required are changing faster across every sector, not only in computing. A student graduating today faces a labour market that differs from the one four years ago, and the present risk of unemployment stems largely from a mismatch between what is asked for and what is offered, not from a lack of work. It follows that training should focus on fundamentals rather than chasing the latest programming language or “prompt engineering”. A university that tries to keep pace with private platforms will always be late, given that its business is to build the capacity to understand, adapt, and learn. 

There is a longer horizon that interests me more, and Keynes recognised it. We have worked hard to avoid having to work, and now that the prospect is within reach for more people and for longer stretches of their lives, we are unprepared for it. Technological unemployment becomes the occupation of free time only if we manage to detach it from the loss of income and social standing, which makes it a question of distribution and tax design, and therefore a political decision. 

You discuss the risks posed by the digital revolution and AI. If I understand your arguments correctly, you believe that, rather than fearing a rogue superintelligence, we should focus on tangible challenges such as algorithmic bias, the responsibility gap, and AI’s environmental impact.

You are right. Let me explain why the superintelligence framing is so attractive. Two narratives dominate the public discussion, appear opposed, yet work as allies: enthusiasm promises a panacea, and catastrophism announces an existential risk. Both take for granted that AI is an intelligence on its way to surpassing us, even if we are not running the same race. They differ only in whether it is sold as salvation or as apocalypse. I argued this in February 2020, in a column for la Repubblica’s weekly, and have not changed my mind. The reader will forgive me a loud ‘I told you so’, which is never a pleasant noise. Those who then announced the imminent end of work owe us an apology, and so do those who promised a panacea. Predictions that carry dates can be checked, and the cost of these was to draw attention away from the real risks, which are concrete and dull. I sat on the European Commission’s High-Level Expert Group, whose guidelines became the ethical frame of the AI Act. What took time there was bias, accountability, and the governance of data. Nobody needed a rogue superintelligence to fill the agenda. 

It helps to separate the old risks that AI amplifies from the new ones it introduces. AI did not create privacy, surveillance, discrimination, and inequality issues, though it makes them worse at speed and at scale. Two of the risks are new. Responsibility is now spread among designers, vendors, deployers, and users so intricately that it becomes hard to grasp. Persuasion has become opaque, since these systems shape decisions and behaviour without the person noticing. To this, I would add the environmental cost, which is real and growing. Training one of these models consumes enormous amounts of energy, and running it for billions of queries every day takes very large quantities of water for cooling. Estimates vary, and none of them is small. Whether the net balance is negative remains open, since AI is also one of the more promising instruments for the energy transition, from optimising electricity grids to researching new battery materials. What is not open is our duty to measure it, and the duty of those who build these systems to publish what it costs. On the label of a refrigerator we find the consumption and the energy class. A model is sold with nothing for a buyer to compare it to. Model cards exist, and they are not an energy label.

You identify traceability and governance as key means of addressing these risks. Could you elaborate on how they might work in practice?

Both work only when built into the process, rather than added at the end. Traceability begins with a practical criterion, reversibility, which I think of as a graded scale. The harder an action is to undo, the more verification must sit in front of it, moving from sampling to case-by-case confirmation and, at the far end, to a prohibition on delegating altogether. A message to a customer and a bank transfer belong on different steps of that scale. Once the scale is fixed, the record has to follow the action: which system produced what, on which data, under whose authority, and who could have stopped it. Without that record, responsibility is not lost but merely untraceable, which, for the person harmed, comes to the same thing. Then governance has to keep a human in command. ‘Human in the loop’ has become an alibi, since it is an operating position that often reduces to signing off on decisions one no longer understands, whereas being in command means choosing the ends, setting the thresholds, retaining the authority to refuse, and answering for the results. 

A parliament is a good place to see what this looks like. In 2024, Anna Ascani, Vice-President of the Camera dei Deputati, kindly invited me to chair its technical and scientific commission, which selected projects to apply generative AI to parliamentary work from about thirty proposals put forward by Italian universities. Three prototypes emerged: NORMA, which analyses legislative production; MSE, which assists with drafting amendments; and DEPUCHAT, which opens what deputies do to the citizens who elected them. Each was assessed for function, architecture, and security before anyone used it, and the decision on what to adopt remained with the Camera. That is the whole of what I mean by human in command, and it is plainer than the phrase sounds. The rest is institutional. Compliance secures a floor and never a ceiling, in that a checklist tells you whether a system meets a minimum standard and says nothing about whether the system should exist or operate, or whether the trade-offs it imposes are acceptable. We therefore need hard ethics, which becomes law and regulation, and soft ethics, which asks what we ought to do once the law has set the boundaries. Governments should fund the second and not only mandate the first. External oversight is unavoidable here, because no firm can be the judge of the trade-offs it profits from. That is a structural point, not a comment on anyone’s character. I have seen this from the inside. In 2019, Google invited me to join its AI ethics council, which it announced at the end of March and dissolved the following week, before it had met once. Some developers now publish a document setting out how their systems are meant to behave, which is worth having. Reading a policy tells you what was promised. It does not tell you whether the promise was or will be kept. 

Your approach is rooted in constructionist philosophy: philosophy should not lose itself in abstract or sterile analysis. Still, it should instead design and develop conceptual models that help us understand and actively shape the present. From this perspective, what concrete recommendations would you offer to governments, institutions, and society at large to ensure that AI is developed and adopted responsibly and serves the public interest?

Constructionism holds that we know what we can build, and that a concept earns its keep by what can be designed with it. Philosophy that comments on the world after the fact has done half its job. The other half is what I call conceptual design, which commits one to actionable recommendations. 

The first recommendation is for governments and public institutions. They buy a great deal of technology, and the terms on which they do so are the fastest instrument of regulation available, since a purchasing rule can be rewritten in a season and a statute takes years. Require the supplier to declare the energy cost, the error rate in the domain of use, and the provenance of the training data. Those numbers will exist, and that is where the work starts, as a declaration is worth only as much as the audit that can test it. In any other sector, the asymmetry between what the seller knows and what the buyer is told would be unthinkable. 

The second concerns the risk of buying a system without redesigning the process it is meant to support. This is the most common and costly mistake I see, and it is made everywhere. I have watched a board resolve to adopt AI and then send people away to find a process to feed it. The dishwasher never learned to wash up as we do. Instead, we changed the plates, the detergents, and the kitchen around it, enveloping the machine in an environment tailored to its measure. AI works where the process has been rebuilt around it, whereas in a process that already limped, it may exacerbate errors. 

Third, education. It pays off the last, and best. Teach what remains true when the tools change, fund the independent bodies whose business is to keep standards alive as technology evolves, and resist the short courses designed for the immediate term, which sell opportunities that expire quickly. 

One warning holds them together. With social media, we optimised for attention and achieved polarisation. We are now optimising for approval and building systems that tell us we are right, which is the opposite of what anyone who has to make a decision needs. 

You are Italian, a naturalised British citizen and currently a professor at Yale, USA. How do you view Italy’s role in the digital and AI revolution, and what are its main challenges and distinguishing features compared with other countries?

Italy now leads the European average in 5G and fibre to the home, which is worth noting because the opposite is repeated out of habit. What the country still lacks is a culture of informed risk-taking and innovation, as well as a stock of digital skills. The former is more challenging because the latter gap can close quickly: the more digital technologies are used, the faster they spread. 

One of Italy’s distinguishing assets is its rich semantic capital, by which I mean the capacity to create, curate, and transmit meaning from the information one holds. With AI, information has become cheap and instantly available, so value shifts to what we can make of it. Italy holds an extraordinary reserve of it: a humanistic tradition, an artistic and archival inheritance, craft, food, design, and a habit of interpretation that no statistical engine possesses. The condition is that the country should stop being embarrassed by its own resources and stop imitating models that do not belong to it. Competing with California on scale is a losing game. Competing on meaning is not. That is a claim I have had to test. This year, when I was President of the Fondazione Leonardo, Micaela Lovecchio and I published a report on the state of AI in Italy, L’Italia nell’era dell’IA. The map shows a country with more research strength than its reputation abroad suggests, spread across universities and institutes rather than a handful of companies, and weakest exactly where everyone suspects, in the passage from research to firms that grow. We presented it at the Camera dei Deputati. The press read it as a search for an Italian third way. But what it documents is narrower and more useful: where the strengths actually are, who is doing the work, and whose responsibility it is to improve them. 

The main challenge is not only Italian. Europe is good at regulation but much weaker at unifying industrial policy, and a federal vision is missing. We were late to the commercial Web, let a lead in mobile telephony slip away, and AI is the chance not to waste another crisis. That requires deciding together which technologies to adopt, from whom, and to which standards, since otherwise talk of law, governance, and trust is decorative. Italy has legislated ahead of the rest of Europe. Law 132, in force since October 2025, is the first national framework designed to sit alongside the AI Act, and implementing decrees began to follow in June 2026. Moving early is a genuine distinction, but it settles nothing on its own, since a law can set terms but cannot create an industry. The specifically Italian risk is a bad habit: we take our innovators, Adriano Olivetti above all, put them on an altar, and light candles. Celebrating Olivetti by building a museum for him is the exact opposite of what he did. History is indispensable as the second dialectical moment, the thing one has to confront. Beginning there is how a culture gets stuck. 

I understand that your research focuses primarily on digital ethics, the ethics of AI, the philosophy of information and technology, epistemology, the philosophy of logic, and the history and philosophy of scepticism. These are wide-ranging and interconnected fields, but could you highlight two or three ideas from your research that you consider particularly important for understanding the challenges we face today?

Three, each leading to the next. 

The infosphere is where I would start, along with the onlife condition it entails. We no longer use digital technologies; we live within them, in an environment where the distinction between online and offline no longer holds as it once did. Once this environmental view is accepted, the questions change. We stop asking how to use a tool and start asking how to design a habitat in which people can live well, a political and architectural question before it is a technical one. 

Semantic capital follows from this. Information used to be scarce and expensive to obtain, and the advantage lay in possessing it. The Web began to remove that scarcity, and AI has finished the job. What remains scarce is the judgement that gives information meaning, places it in a context that holds, and decides what matters and why. A system applied to finance can retrieve every objective fact about a company, whereas deciding whether to invest in it draws on culture, experience, tolerance for risk, and a clear sense of what ought to happen, all of which are forms of semantic capital. That is where value now sits, and it is also what education is for. 

The eclipse of the analogue is the last concept I would press on anyone thinking about geopolitics, because it has a twist. The more the digital absorbs our lives, the less we look at the analogue, and the more power there is in controlling the analogue to control the digital: the satellites, the cables, the electricity, the rare earths, and the sites where data centres are built. Sovereignty is being decided there, in concrete and in copper, while the debate focuses on models and benchmarks. 

Looking back on your career, how did your journey from Italy to the United States begin? Tell us a little about your early years, your studies, and the experiences that shaped your path.

I was born in Rome in 1964 and studied philosophy at La Sapienza, where logic, mathematics, and economics interested me more than the curriculum on offer. The decisive moment came with a book. The Italian translation of Susan Haack’s Philosophy of Logics, Filosofia delle logiche, was clear and sharp as a diamond, and philosophically honest in a way that felt almost transgressive to an undergraduate, so I wrote to her, with an envelope and a stamp – those were the days – and she replied. 

In 1988, with a scholarship from La Sapienza, I went to Warwick as a visiting student to work with her, and I stayed, taking my Master and then my PhD under her supervision. She taught me to think philosophically by example, with meticulous comments on drafts, week after week, tutorial after tutorial. She died in March this year. I owe more of whatever philosophical capacity I possess to those Warwick years than I can easily say. I remember days without end spent between the campus cafeteria and the laboratories, sleeping little, writing my thesis, and running logic programs on the Olivetti M24S available to every student at any hour. Beautiful machines, indestructible and reliable, which I had never seen as a philosophy student in Rome. They ran DOS. They were networked, though the Web did not yet exist. 

The turn to the digital came from something small. From Italy, using one of the first modems, I downloaded files from the Warwick site and realised this would change how people live. When I said so at home, to some professors, I was told the internet was a fad. So I went to do research at Oxford, where I did my postdoc for four wonderful years. My Oxford career began there, more from that closed door than from any plan. Much later came Bologna, where I hold a part-time chair in the sociology of culture and communication in the Department of Legal Studies, and later still Yale, where I founded the Digital Ethics Center. Leaving Oxford was difficult, but Yale could not be dissuaded. And it deserves the credit for that. They were looking for the Director of the newly funded Digital Ethics Center. The Center exists because Yale was prepared to back an ambitious idea and give it the room to become an institution. In the end, it was the classic “offer I could not refuse”. 

If you could speak to your younger self at the beginning of your career, what advice would you give – and what lessons would you share with young Italians entering the field today?

To my younger self, I would say at least three things. Learn English much earlier. Learn more maths and computer science earlier. Realise earlier that you have left home. These are all things I had to do with much more effort later. 

I do not know whether I have any advice for others. But if forced, perhaps I could mention the following. 

Caute ambizioni, cautious ambitions. The example is mountaineering, even if I speak as an amateur. Aim for the highest summit you can imagine, and climb it with the caution of someone who takes one step at a time and knows the risks from having been up there. Above all, do not listen to anyone who tries to curtail your ambitions. The list of people who told me I could not do something or would fail at something else is endless. Luckily, I did not listen to them. 

Stop reading and start thinking: take pen and paper, clear the blinding light of the great masters from your eyes, stay in the half-light for a while, form a few ideas of your own, and only then confront the tradition with them. Modern philosophy begins with Descartes, a military engineer who learned mathematics in the army, and contemporary philosophy with Wittgenstein, another engineer who came to it from aeronautics. Both arrived from the side, with their own ideas, and met the tradition afterwards. 

Then read Aristotle, not the Aristotelians; Kant, not the Kantians; Wittgenstein, not the Wittgensteinians. Philosophy regularly turns into scholasticism, in the conceptual rather than the medieval sense, when philosophical problems become philosophers’ problems, and those interest nobody. The way out is the same in every generation, and it arrives when somebody pulls the philosopher by the sleeve and says, you know a great deal; would you look at what is pressing and difficult for us now? 

Build fundamentals rather than skills with an expiry date, since the tools will change several times over a working life and what helps you learn the next one will not. And take Italy seriously as a place to do the work. The conditions have changed: the infrastructure is in place, the universities are better than their reputation abroad suggests, and the culture is a resource no competitor can buy or copy. Go abroad and take everything the best places have to teach you. Then come home and build something with it. You can live in more than one place, hold more than one passport, and enjoy more than one culture.