"Restoration is possible, but not by nostalgia" – Luciano Floridi on AI, information space, and education

The University of Bergen has appointed Professor Luciano Floridi of Yale University as an honorary doctor, acknowledging his contributions to research, society, and culture. Flordi is internationally known for his work in information ethics and the philosophy of technology, and has spent several decades studying how digital technologies and artificial intelligence shape society.

In conversation with Infomedia, the honorary doctor discusses the erosion of the public information space, responsible AI, and what education should look like in an age of artificial intelligence.

Your academic formation was in logic and epistemology, yet you went on to become one of the defining voices in the philosophy of information and AI ethics. You once reflected on how your early intellectual interests were shaped, in part, by engagement with religious thought and questions about how meaning and messages are communicated by God to humanity, if at all. Could you elaborate on that formative period? What questions were you wrestling with, and how did they ultimately orient you towards digital ethics?

– My early intellectual formation began with a question that, looked at from one angle, is theological, and from another, semantic: how can a message travel from sender to receiver such that the receiver grasps what the sender meant? In the Catholic tradition I grew up in (I’m no longer a believer but I’m agnostic, which means I do not believe in the non-existence of God), this took the form of revelation, and asked what it meant for God to communicate with humanity and for humanity to understand. This got me interested in Shannon and information theory. And in epistemology, because in Athens, before Christ, it was already the problem the sceptics worried at: if any sign can be misread, what licenses the claim that I have understood what was sent? The question survived in me even as the answers I once accepted gave way to others.

Logic and epistemology gave me the tools to ask the question more precisely. My doctoral work in philosophy of logic and epistemology taught me that the problem of understanding is not solved by accumulating evidence; it is reframed each time the medium changes. When the medium became digital, the philosophical questions multiplied rather than dissolved. Old philosophers’ problems became new philosophical problems again. Who counts as the sender when a system produces text without an author, and what does it mean to grasp content that no one meant? Digital ethics was not a turn for me; it was the same question in a new register. I had begun by asking how meaning crosses a gap. I ended by asking who is responsible when meaning is produced across a gap no one has crossed.

In 1996 you warned that the internet would corrode the public information environment – a prediction that has proved accurate in ways few anticipated at the time. You were also among the first to analyse what large language models are, and crucially what they are not, well before their mainstream adoption. We are now facing an information environment in which the volume, velocity and plausibility of AI-generated content is outpacing our collective ability to make sense of it. Was there anything about how this unfolded that surprised you – and do you believe the conditions for a trustworthy public information environment can still be restored?

– My 1996 paper – ‘Brave.Net.World: the Internet as a Disinformation Superhighway?’ – carried a title I would not have chosen had I expected, thirty years later, still to be answering questions about it. The argument was not prophetic; it was, in retrospect, analytic. Combine zero marginal cost of publication with no gatekeeping and no liability, and the system will be flooded with low-quality content. It was already the logical conclusion thirty years ago. The only real question was how long the system would take to find that equilibrium.

What I did not foresee was that the same dynamic would scale through machine generation, and that the dominant business model of digital media would actively reward the flood rather than resist it. I expected noise; I did not expect noise to become the product at an industrial scale.

Restoration is possible, but not by nostalgia. The information environment we lost was sustained by editors, libel laws, advertising mediated through trusted brands, and a relatively narrow technical bottleneck on production. None of those will come back. What can be built is a different equilibrium: provenance infrastructure that makes the source of content recoverable, liability regimes that reward platforms for accuracy rather than engagement, an education system that teaches the reader’s craft as deliberately as it teaches the writer’s, and above all a fundamental shift from provenance (who or what made it) to responsibility (who answers for it?). The trustworthy public sphere of the twentieth century is not coming back. A trustworthy public sphere of the twenty-first century has to be designed, the design work is overdue, and it can use the same technology that today is part of the problem, to make it part of the solution.

The commercialisation of large language models has transformed operations across sectors, including warfare and national security. When Anthropic recently refused the Pentagon’s demand for unrestricted access to Claude for autonomous weapons and mass surveillance, it lost the contract – OpenAI accepted the same terms and secured it. This episode laid bare a central dilemma in AI: how to balance commercial opportunity and national security imperatives against ethical principles. Having advised organisations like Microsoft, IBM and Meta on responsible AI, how do you read what happened – and what does it tell us about whether ethical commitments in AI can survive contact with power and money?

– What happened to Anthropic is what always happens when a firm tries to convert a stated principle into a binding commitment that costs money. Anthropic declined two specific use cases – autonomous lethal systems and mass domestic surveillance – and was branded a ‘supply chain risk’ for the trouble. The contract went to companies whose own red lines may, on paper, look similar, but whose deployment architecture and reliance on existing law place the substantive guardrails elsewhere. Whether those substitute guardrails will hold is something we will know in a few years, not very soon. The episode confirms a point I have been making for some time. Ethics, when it is left as a matter of voluntary corporate self-restraint, has the half-life of the market conditions that allowed it. When the buyer is the state, and the state is willing to pay a premium for unrestricted use, voluntary restraint is selected against. This is not a moral failure of any one company; it is what happens to ethics that has not been turned into or supported by law.

The lesson is therefore not ‘choose the more ethical company’ but ‘build the structures that make ethics survive a change of market conditions’. Procurement rules, audit requirements, statutory red lines, competition and anti-trust, independent oversight: the unglamorous instruments of governance. Without them, the question of whose conscience holds is reduced to the question of whose investors will tolerate a lost contract. We have just been shown the answer.

You played a central role in advising the European Union on the AI Act and the framework for Trustworthy AI – arguably the most ambitious effort to date to ground AI governance in ethical principles and give them the force of law. What does it actually look like to translate research and ethics into binding policy – and what lessons do you take from that experience?

– The work I did with the High-Level Expert Group on Artificial Intelligence, which drafted the Ethics Guidelines for Trustworthy AI in 2018–19, was a remarkable, sustained encounter with what it actually takes to move from a published argument to a legal text that will bind hundreds of millions of people. The honest answer is that the process is long, partial, and humbling.

It is long because the legitimate stakeholders are many, and each one has both standing and a veto. It is partial because the law that emerges is always less than the ethics that inspired it; the AI Act is a compromise, not a perfect recipe, and that is the right relation between the two. It is humbling because a philosopher used to refining a paragraph until it is exactly right has to accept that legislative text is refined under different constraints – enforceability, jurisdictional fit, the patience of the trilogue process, constraining interests, technical requirements, legal coherence, political costs, and so forth – and that the elegant formulation may have to give way to the one that survives a Council working party at 2 a.m. and works.

What I took from that experience were three points. First, ethics that refuses to enter the legislative process forfeits its claim to be taken seriously; complaining from the sidelines is a luxury position. We cannot afford it, especially today. Second, the translation work – finding the legal concept that can carry the philosophical claim – is itself a substantive contribution, not a downgrade. Finally, the work continues after enactment, because implementation is where the ethics either lives or dies. I am also very interested in the latter phase.

AI is transforming what it means to learn and to teach. As students increasingly turn to AI to complete assignments and homework, traditional models of assessment are under serious pressure. Do you think they are still fit for purpose – and what concrete changes should educators be making to how they design assignments, measure learning, and structure teaching in the age of AI?

– Traditional models of assessment were already under strain before AI. They were holding only because students and teachers respected the convention. Now that they are not required to, the structure that depended on the convention is failing. The take-home essay, the unsupervised assignment, the term paper produced over a fortnight, and the related forms of assessment: these were never measuring what we claimed they were measuring. They were measuring willingness to comply with a procedure. AI has merely revealed the gap.

Three concrete changes are overdue. First, separate sharply between assessment of process and assessment of product. A finished essay can be produced by a machine. A defence of the essay’s argument under questioning cannot be produced by a machine, at least not in the room where you meet the student. Oral examination, viva-style assessment, and in-class supervised writing should return as the spine of any qualification that claims to certify thinking. That is time-consuming, I know. Years of one-to-one tutorials in Oxford taught me that. But it is also the best we can do, and digital technology can also help us. Second, treat AI as part of the curriculum rather than its enemy. Students will use these tools throughout their working lives, so will the teachers, and the university owes both training in how to use them well, including the disciplined recognition of where the tools fail. Finally, the purpose of education must be restated. It was never to certify that the student had done the work, but that the student had become someone capable of the work. The certificate is not a receipt. It is a claim about the person to whom it is given.

This is a precious moment for redesigning education from first principles. We will not get a better one. The institutions that act now will set the template that the others will copy for the following years.

You are being awarded an honorary doctorate in recognition of your extensive work that has shaped both academic thought and public policy across the world. What does this distinction mean to you?

– An honorary doctorate from the University of Bergen means a great deal to me. Bergen is a remarkable place: a coastal city with a long memory, a university that has done patient and highly distinguished work on media and information studies for decades, and a department whose colleagues I have followed and learned from. To be welcomed by them, on their terms, in their tradition, is amazing. It is like being told that the work I have been doing for thirty years – across philosophy of information, AI ethics, and digital governance – has not been worthless. There is nothing like having your peers say that they appreciate your efforts. I do not take that lightly. I receive it as both a gift and a charge: a gift to be grateful for, and a charge to keep doing the work for as long as the work is worth doing. Fingers crossed, I have another couple of decades at least.