I have followed Luciano Floridi's work with interest for years, long before artificial intelligence became a boardroom topic. What has always struck me is that he rarely asks the question everyone else is asking. While the public conversation about AI swings between enthusiasm and alarm about what the machines can do, Floridi keeps turning the question around: what kind of world are we building for the machines to work in, and what happens to meaning when information becomes abundant?
Two of his recent interventions, one methodological and one philosophical, answer those questions in ways that fit together more tightly than they first appear. This article is my attempt to read them as a single argument.
Luciano Floridi is an Italian philosopher, born in Rome in 1964, and one of the founders of the philosophy of information as a distinct field. After two decades at Oxford, where he was Professor of Philosophy and Ethics of Information at the Oxford Internet Institute and directed its Digital Ethics Lab, he moved in 2023 to Yale University as Founding Director of the Digital Ethics Center. He also holds a chair in Sociology of Culture and Communication at the University of Bologna. Beyond academia he has sat on advisory bodies for the European Commission, including the High-Level Expert Group on Artificial Intelligence, and has shaped digital-ethics policy in Italy and the United Kingdom.
His intellectual project is laid out in a tetralogy published by Oxford University Press: The Philosophy of Information (2011), The Ethics of Information (2013), The Logic of Information (2019) and The Ethics of Artificial Intelligence (2023). Along the way he coined vocabulary that has entered everyday use: the infosphere, the informational environment we now inhabit; onlife, the condition of living with no clear boundary between online and offline; and the Fourth Revolution, the idea that after Copernicus, Darwin and Freud, Turing has displaced humanity once more, this time from the centre of the informational universe. A recurring thread in all of it is the claim that digital technologies do not simply give us new tools; they re-engineer the environment in which we act, and the ethics and design of that environment deserve at least as much attention as the tools themselves.
That thread is exactly what connects the two ideas discussed below. Both are about the environment around the algorithm rather than the algorithm itself: one about its physical and organizational structure, the other about its layer of meaning.
Why the model is rarely the problem
Roughly four out of five enterprise AI initiatives fail to deliver what they promised. The explanations usually point at the model: too weak, too generic, trained on the wrong data. Floridi's work suggests instead that the model is rarely the culprit, and that failure happens in two distinct places, one after the other.
Think of an AI model as water. Water is powerful and, these days, abundant and cheap. Whether it irrigates a field or floods a basement has nothing to do with the water itself. It depends on the canals that carry it and on the soil it reaches.
Phase one: the canals
Floridi has argued for over a decade that successful automation works by enveloping: rather than building machines clever enough to cope with the messy human world, we reshape a portion of the world into an environment the machine can handle. A dishwasher does not understand dishes; it operates inside a box designed so that it does not need to. Autonomous vehicles work because roads, sensors and traffic rules were built around them; warehouse robots work because the shelving was designed for robots, not for people. The intelligence is as much in the envelope as in the agent.
With Bruce Benson at FTI Consulting, Floridi has turned this idea into an implementation methodology, the Floridi Curves, whose central claim is that most AI failures are failures of the envelope rather than of the algorithm. Enterprise AI needs its envelope too, and almost nobody builds it first. Organizations buy the model, then discover that the data is dirty, the processes are undocumented, the decision criteria were never written down, and nobody owns the action that should follow from the output.
This is the canal problem. Without channels, the water goes everywhere and nourishes nothing. The questions that belong here are practical: what inputs will the system actually see, and are they fit to be seen? Who checks whether the output is correct, and does that person have the competence to check it? Who acts on the answer, and by what rule? Answering them well gets an organization from a failed project to a running one. It does not yet get them anything worth having.
Phase two: the soil
The second idea comes from Floridi's work on what he calls semantic capital, which I will call here meaning infrastructure. His definition is simple and demanding at the same time: semantic capital is anything that increases our capacity to give meaning to the world and make sense of it. The key word is capacity. Floridi is not describing a collection of things we own but a power those things confer on us. A dictionary of ancient Greek on my desk is not semantic capital for me if I cannot read Greek; it is for the scholar who uses it to understand an inscription. The capital lies in the interpretive power a content gives to someone, not in the content itself. That is why he borrows a word from economics: like financial capital, it accumulates, can be invested to generate more, depreciates when it is not curated, and makes people and organizations unequal according to how much of it they hold.
The definition also widens what counts as content. We tend to think of content as something codified and transmissible: a book, a film, a dataset. Floridi includes everything that is never written down: personal experience, family stories, the jargon of a trade, the way a community reads certain gestures. A nurse with twenty years on the ward sees things in a patient that no manual describes. All of this is semantic capital, because it increases that person's ability to understand a situation. The consequence for organizations is immediate: most of their semantic capital is tacit, distributed across people, and does not coincide with their archives.
The most important move, though, is the one that separates semantic capital from information. A content, Floridi insists, is never just a well-formed package of data with syntax and structure. As long as it remains that, it is information: correct, perhaps, but inert. It becomes semantic capital only when it meets someone who perceives it, interprets it and folds it into their own story. Meaning is not a property of the content but of a relationship. The same family photograph is a JPEG to a compression algorithm, a historical document to an archivist, and a piece of identity to the grandson who recognizes his grandmother as a young woman. The file is identical; the meaning generated could not be more different.
From this follows the consequence for machines, which Floridi draws without hesitation. If semantic capital arises only when content is integrated into a lived story, then a system that lives nothing cannot have it, however sophisticated its processing. A language model can recombine and produce fluent text about loss, belonging or fatigue, but it has never lost anything, never tired, never recognized itself in a narrative. It manipulates the syntax of meaning without possessing its semantics. Floridi's image is the pearl: it forms because the oyster has been wounded and responds by coating the irritation, layer upon layer. Meaning forms the same way, around whatever has caused us friction, and machines have no grain of sand.
This is what makes the second phase so easy to skip. Information, Floridi argued at Orbits in Milan last November, has become a commodity. A model can produce in seconds what used to take years of research, so the output itself no longer differentiates anyone. What is scarce is the human and organizational capacity to make that output mean something specific, here, for us. Perfectly channelled water poured onto dead ground grows nothing. An AI system can return impeccable data to an organization that has no shared understanding of what the data refers to, why it matters, or what it should change. The result looks like a map with no legend: every coordinate is correct, and none of it tells you where to go.
Consider a dashboard that reports a churn rate. The model that produced it may be flawless. But the number is only usable if someone in the room knows how "customer" was defined when the metric was first built, which segments were deliberately excluded, why a spike two years ago was ignored, and what level of churn this particular company has historically been willing to tolerate. None of that is in the data. All of it is in the meaning infrastructure, and most of it lives in people's heads rather than in any system. And because generative AI produces confident output whether or not that layer exists, its absence is no longer visible as a gap. It becomes visible only later, as a decision that was technically well-informed and organizationally wrong.
Two phases, one failure
The canal problem and the soil problem are usually treated as separate conversations, engineering on one side and culture on the other. They are better understood as consecutive phases of a single task: making machine output usable by humans. Skip the first and the project collapses visibly. Skip the second and it succeeds quietly into irrelevance, producing the same commodity information as every competitor and leaving people to guess what it means.
The practical lesson reverses the usual order. Before asking which model to buy, ask whether the canals exist. And before trusting any data the system returns, ask whether there is soil for it to land on: a shared, explicit, curated layer of meaning that turns correct output into a decision somebody can stand behind. Water is cheap. Canals can be engineered. Soil has to be cultivated.
References
Floridi, L. (2014). The Fourth Revolution: How the Infosphere is Reshaping Human Reality. Oxford University Press.
Floridi, L. (2018). Semantic Capital: Its Nature, Value, and Curation. Philosophy & Technology, 31(4), 481–497. https://doi.org/10.1007/s13347-018-0335-1
Floridi, L. (2019). What the Near Future of Artificial Intelligence Could Be. Philosophy & Technology, 32(1), 1–15. https://doi.org/10.1007/s13347-019-00345-y
Floridi, L. (2023). AI as Agency Without Intelligence: On ChatGPT, Large Language Models, and Other Generative Models. Philosophy & Technology, 36, 15. https://doi.org/10.1007/s13347-023-00621-y
Benson, B., with Floridi, L. (2026, January 23). Floridi Curves: De-Risking AI. FTI Consulting. https://www.fticonsulting.com/insights/articles/floridi-curves-approach-taming-ai-problems-industry
Floridi, L. (2025, November 19–20). Keynote at Orbits – Dialogues with Intelligence, second edition, Palazzo del Ghiaccio, Milan.