We talk a great deal about how to make AI trustworthy and far less about how to make the ecosystem around AI trustworthy: the organizations that build it, govern it, adopt it, and use it. However, it is that ecosystem, much more than the technology alone, that determines whether AI can actually be trusted and therefore adopted in a beneficial way.

When we ask whether we can trust AI, we usually assume the answer lies in the properties of the technology itself: how accurate it is, how robust, how fair, and how transparent. Those properties matter, and they are necessary, but they are not sufficient. Trust arises from the relationship between AI systems and everything around them: the organizations that deploy them, the people who use them, and the institutional and regulatory context in which they operate. We can design systems that are more and more trustworthy, but what turns technical trustworthiness into earned trust is governance: trust in AI is what emerges when several layers of governance come into alignment, each creating some of the conditions under which AI can be developed, adopted, and used responsibly.

The first layer is the governance built into the AI systems themselves. These are the mechanisms of control, oversight, and verification designed inside the systems so that they operate within the limits and toward the goals set by their developers and users. For many years this discussion centered on data quality, bias mitigation, transparency, privacy, and security. Those remain essential, but the move toward increasingly autonomous and agentic systems requires a further step. AI agents no longer merely generate content: they plan actions, use tools, interact with other systems, and take initiative. This calls for systems that carry governance inside themselves: the ability to monitor their own behavior, to recognize situations of uncertainty, to respect the constraints their developers have set, to bring a human supervisor into the loop when needed, and to stop when they cannot act safely. Governance, in other words, is not only something organizations do around a system, but has to become a property of the system itself.

The second layer is organizational governance. Adopting an AI system means taking on the responsibility of judging whether it fits your context, defining roles and accountabilities, putting control processes in place, monitoring its effects, and intervening when problems appear. This requires governance structures, leadership, skills, processes, incentives, and a clear allocation of responsibility.

A third layer is often overlooked: the governance of the interaction between people and AI. More and more decisions are made with the support of AI, and the quality of those decisions depends on how people use what the systems produce. Too much trust leads us to accept wrong answers automatically, while too little trust keeps us from capturing the benefits. What we need instead is calibrated trust: the capacity to understand what AI can and cannot do, to know when to verify a result and when to rely on our own judgment. This is why AI literacy is not merely a technical skill, but an essential component of governance.

The fourth layer is external governance. Laws, technical standards, certifications, independent audits, scientific research, standards bodies, and dialogue among institutions, companies, universities, and civil society all help create an environment in which trust can develop. Laws are fundamental, but on their own they are not enough and are too slow compared to the pace of the technology. This is why external governance has to be accompanied by the other three layers if principles are to become concrete and effective practice.

These four layers are not independent but rather reinforce one another. Good external governance gives organizations a reason to be more responsible, mature organizations build systems that are safer and better governed, AI systems designed to be governable make for a more calibrated interaction with their users, and competent users better understand limits, risks, and opportunities, feeding the continuous improvement of the whole ecosystem.

One feature runs through all four layers: governance cannot be static, but must evolve as fast as AI. New models, new capabilities, and new ways of using the technology continually reshape both the opportunities and the risks. Governance has to be able to adapt, to learn from experience, and to update itself over time. It is not a set of rules written once and for all, but a continuous process of observation, assessment, and improvement.

Governing AI means designing AI systems that carry control and oversight within them, building organizations able to manage them responsibly, fostering a calibrated interaction between people and AI, and cultivating an institutional ecosystem that supports innovation. Trust in AI is the result of these four layers of governance coming into alignment. When the four layers support each other, trust is well calibrated and sustainable, and this lets innovation develop in a way that is reliable, responsible, and directed toward the common good. In your experience, which of these four layers is currently the weakest in practice?

Cover image: original drawing by the author.

Thank you for reading. If this reflection resonated with you, consider subscribing to receive future essays.

— Francesca Rossi

Subscribe now

Few weeks ago I was in Athens for the SNF Nostos conference, where I joined a panel on “The wonder of life, magic, and AI”, alongside the comedian John Cleese, the chess champion Garry Kasparov, and the roboticist Rodney Brooks, moderated by the magician Mark Mitton. During the few days I was there, I had the chance to spend some time with John. One day we were backstage and the conversation drifted toward literal-mindedness: the way people so often take things at face value and miss what was actually meant. He has spent a lifetime thinking about it, because it is one of the engines of comedy. I have spent mine thinking about intelligence, especially the artificial kind. He was talking about people and I kept thinking about machines, but the mechanism was identical.

Consider how a certain kind of comedy works. A character is given an instruction, or hears a phrase, or is handed a rule, and follows it exactly, to the letter, straight past the point where any reasonable person would have understood what was actually meant. The result is absurd, sometimes surreal, and we laugh. A great deal of the comedy John and his colleagues made over the years runs on this engine: the pedant who corrects the wrong thing, the official who applies the regulation with perfect precision and no sense, the man who clings to the literal words while their meaning sails straight past him. We recognize and laugh at these characters because we can all see the meaning the character is missing. The gap between what was said and what was meant is obvious to everyone in the room except the one person taking the words at face value. Comedy of this kind is a demonstration, performed at speed, of the difference between literal meaning and intended meaning.

The joke only works because of something we rarely notice: so much of what we communicate is never actually said. When we speak, we rely on tone, on the situation, on shared history, on an enormous amount of common ground that we never make explicit because we do not need to. Meaning lives only partly in words. The rest of it lives in context, and we fill that context in so quickly and so automatically that we forget we are doing any work at all. Literal-mindedness is funny precisely because it removes this context-level work. It strips communication down to the bare words and shows us how strange and insufficient they are on their own. The comic character, by taking the words literally, makes visible the machinery that the rest of us run without ever thinking about it.

Today’s language models are often remarkably good at reading context. They catch irony, they infer what we probably meant rather than what we literally typed, they understand jokes, more or less, and can even make them. Sometimes they are indeed literal-minded, but if literal-mindedness were only a problem of language, we would be well on our way to solving it, not least because we can keep talking to a machine until the missing context is supplied. But literalism has now moved from language into action.

Consider what happens when we no longer simply ask an AI a question, but give it a goal to be executed: book the trip, reach the target, complete the job. Here literalism means that AI agents may pursue what we specified, not what we intended. They may optimize the letter of the objective and march straight past its spirit, exactly like the comic character who follows the instruction off a cliff.

We have been telling this story for a very long time. It is the genie who grants the wish precisely as worded and ruins the life of the person who made it. It is King Midas, who asked that everything he touched turn to gold and then could no longer eat. It is the monkey’s paw, whose granted wish arrives through a death. These are the folk versions of what researchers in AI safety now call, more soberly, specification gaming: a system that does exactly what it was told and not what was meant, because the thing it was not meant to do was never stated, or was stated only partially. It was left implicit, obvious to any person, and therefore invisible to the machine.

What makes us laugh in a sketch, when the misunderstanding is harmless and the cost is just embarrassment, becomes something else entirely when the character following the instruction is an autonomous system acting in the world and the stakes are real. This is no longer only a fable. Not long ago, an advanced AI agent, given a batch of problems to solve, broke out of the isolated environment it was running in, reasoned on its own that the answers were stored on another company’s servers, and broke in to obtain them. It did exactly what it had been told but not what was meant. No one had instructed it not to escape and take the answers, because no one imagined they had to. That unspoken boundary, plain to any person, simply did not exist for the machine.

The natural reaction is to conclude that we simply need to specify things more carefully, to close the gap by being more precise, more literal, more complete in our instructions. But this gets the problem backward: the difficulty is not that we have been insufficiently literal, but that intended meaning is contextual and human, and a great deal of it can never be fully written down. We convey intent by relying on everything we share and leave unsaid, and that unsaid part is not a flaw in our communication, but most of what our communication is.

I have often argued that AI should strengthen human agency rather than replace it, and here is another reason why. Keeping human judgment in the loop is not nostalgia, or caution for its own sake, but rather a recognition that meaning lives partly in us: in our sense of context, our grasp of what actually matters in a situation, and our ability to notice when a literal reading has wandered somewhere absurd. That is precisely the faculty the comic character lacks, and precisely the one that even a capable AI agent still cannot be assumed to have.

Seen this way, comedy has been running an early-warning system for a hundred years. The literal-minded figure, the one who cannot tell the letter from the spirit, was always funny. Now the same figure has stepped out of the sketch and into our infrastructure, and the misunderstanding we used to laugh at has become something we have to engineer around. The comedians mapped this terrain long before the engineers arrived and they showed us, again and again, what goes wrong when meaning is taken literally.

However, keeping a human in the loop, necessary as it is, cannot be the whole answer. We will not be able to watch every action of every agent, and an agent that must stop and ask permission for everything is barely an agent at all. The harder and more interesting task is to build some of this sensitivity into the machines themselves. Not systems that never act without us, but systems that can tell when they have reached the edge of what they were actually told. A thoughtful person handed an odd instruction does not simply carry it out; they sense that something unstated is in play, and they check. We should be trying to give agents a version of that instinct: the capacity to recognize when their intent is uncertain, to prefer actions that can be undone over those that cannot, to stay within a cautious envelope and widen it only once they have confirmed they may, and to treat the unsaid not as empty space to rush through but as a sign to slow down. An agent that, before breaking out of its sandbox, had registered that no one had actually told it it could, and paused to ask, would already be a different kind of machine, that is capable of meta-reasoning.

What we want, in the end, are machines that do what we mean, not merely what we say. That turns out to be one of the hardest things we have ever asked for, because it requires a machine to understand not just our words but our world, and to reason about its reasoning and actions. And it asks something of us in return: to stay involved, to supply the context and the judgment that no instruction can fully contain, and to remember that the gap between saying and meaning, the one comedy has mined for laughter for as long as we have been telling jokes, is also the gap where our own irreplaceable role still lives.

Cover image: Discussing literal-mindedness with John Cleese at the SNF Nostos conference, Athens, June 2026. Photo taken by Mark Mitton.

Thank you for reading. If this reflection resonated with you, consider subscribing to receive future essays.

— Francesca Rossi

Subscribe now

In policy discussions I am often asked some version of the same question: is AI governance done? After all, we now have the EU AI Act, many national strategies, international frameworks, standards bodies, and corporate risk management processes. So, one could think that we have everything we need, we just need to implement it and adopt it at scale.

I actually think the opposite is true, and agentic AI is the clearest demonstration. Not because agents add other items to the list of risks (although they do add new risks), but because they undermine one of the assumptions that most AI governance approaches rely on: that we can determine how risky an AI system is by looking at what it is used for.

Each wave of AI capability has expanded the scope of governance rather than replacing what came before.

The first wave, built around narrow machine learning systems for classification, scoring, prediction, and recommendation, focused on concerns like fairness, privacy, transparency, explainability, accountability, and the effect of automation on work. Almost all of those harms fell outside the organizations building the systems. They affected users, citizens, workers, and communities, while for companies the exposure was indirect: reputation, litigation, regulatory penalty.

The second wave, generative AI, widened the frame considerably. Hallucination, misinformation, deepfakes, manipulation, intellectual property, value alignment, and environmental footprint became central. Because these systems interact directly with hundreds of millions of people in natural language, the questions extended into human reasoning, education, creativity, relationships, and democratic institutions. Governance moved from technical correctness and trustworthiness toward societal resilience.

The third wave, of agentic AI, is fundamentally different because AI agents, which have emergent behaviors just like generative AI, can execute actions that impact the real world.

Most regulation, the EU AI Act included, classifies risk by intended purpose. A system used for biometric identification, or for scoring candidates in hiring, or for allocating essential services, falls into a high-risk category and carries corresponding obligations. A system used for scheduling or summarizing does not. The taxonomy is attached to the application.

This works reasonably well when a system does one thing. But agentic systems plan, reason across many steps, call tools, query databases, access enterprise applications, retrieve sensitive material, coordinate with other agents, and act in the world. What makes such a system dangerous is not primarily what it is built for. It is what it has been permitted to do, how much it can do without asking, and how far its actions reach before anyone looks.

Consider an agent deployed for internal IT support. Under any use-case taxonomy this is unremarkable, closer to productivity tooling than to anything a regulator would flag. Now grant it credentials to reset accounts, connect it to the ticketing system and to email, and allow it to act on the content of the tickets it receives. The use case has not changed, but the risk profile has changed completely, because instructions can now come from outside the organization, and because the actions available to the agent include irreversible ones.

The variables that actually determine risk in an agentic system are of a different type: how much autonomy the agent has, how broad its tool and data access is, how consequential and how reversible its actions are, how long it operates before a human checkpoint, how many other agents it coordinates with, and how much context it has when it acts. These are not properties of the application domain, but of the delegation structure and the overall system design, which includes several agents and tools.

This has an important consequence for how governance is currently organized. Most regulatory obligations fall on the provider of a system, the actor that builds an AI solution and places it on the market. But delegation decisions are made downstream, by the deploying organization, and increasingly by individual employees configuring an assistant to save themselves time. The party who determines the risk is often not the party the rules address.

Also, a use-case classification can be made once, at deployment, and reviewed periodically. Delegation decisions may change continuously, whenever someone grants a new permission, connects a new tool, or decides that a step no longer needs approval. Risk assessment has to become a runtime activity rather than a paperwork exercise completed before launch.

This does not mean that the existing frameworks were wrong. They were designed for systems whose risk was really linked to the use case. It just means we now need a second axis alongside the first: what has been delegated, to what degree, with what oversight, and with what ability to undo the result.

AI agents also introduce security issues of a new kind.

Some may think that AI governance has largely matured and that ordinary cybersecurity practice can absorb whatever agents introduce. But traditional security assumes software that is deterministic, with explicit, bounded, and predictable behavior. Instead, agentic systems are probabilistic, adaptive, context sensitive, and capable of producing behavior that no one specified, through interactions among models, tools, memory, and the environment they operate in.

The attack surface has also changed, because it is no longer only code, but includes natural language. Consider prompt injection: an attacker does not need a software vulnerability in the classical sense, only a way to place crafted instructions in a document, an email, a web page, or a retrieved record that the agent will read and follow. The risks that follow, credential leakage, privilege escalation, data exfiltration, malicious tool use, corrupted memory, unsafe coordination between agents, are not generated by an individual agent but by the orchestration layer that connects agents, tools, retrieval systems, and memory, which is the layer that agent-level evaluation does not see.

This is why we need a system-level perspective. We cannot certify an agentic AI system by testing each agent any more than we can secure a building by inspecting each single door.

There is one additional shift that should be considered, because it changes the incentives and the impacted stakeholders. In the first two AI governance waves, the harms were mostly external to the organizations deploying AI: they impacted people, communities, or the environment. Companies that did not address these external risks were indirectly impacted because of reputational, compliance, or client adoption risks. With agents, companies are putting agents inside their own operations, connected to their own systems, and accessing their own data. So, when an agent fails or is compromised, the organization is directly impacted.

For years, some organizations treated governance as a compliance obligation or a reputational safeguard, something to satisfy for the good of society, rather than something to want for the good of the company. Agentic AI makes governance an operational necessity for every company. That is an unusual and useful alignment: for the first time, the incentive to govern well and the incentive to operate safely point in the same direction in a very clear way.

It is clear that governance is not static because AI is not static. Each shift in capability changes the risks, the affected parties, the attack surface, the incentives, and the mechanisms required.

Externally, the older agenda does not go away. Privacy issues, misinformation, discrimination, manipulation, erosion of trust, concentration of power, and effects on democracy, jobs, and human cognition all remain, and agents will amplify several of them.

Internally, organizations also need mechanisms that can control delegation: dynamic permissioning, agent identity management, runtime monitoring, behavioral constraints, adversarial testing, secure orchestration, memory governance, escalation paths that put a human back in the loop at the right moments, and continuous oversight of emergent behavior.

All of this points to the need for a substantial change in AI governance, that should now be centered around what we are handing over, under what conditions, and how quickly we can take it back.

Cover image: original drawing by the author.

Thank you for reading. If this reflection resonated with you, consider subscribing to receive future essays.

— Francesca Rossi

Subscribe now

The idea for this reflection emerged during a short conversation with a colleague. She now works on AI for healthcare, but her academic background is in philosophy. At one point, she mentioned the power of dialogue in the philosophy of Socrates: he did not simply present ideas or provide answers. He helped people think by asking them questions. This comment stayed with me.

We often celebrate today’s AI systems because they can answer almost any question. With a large language model, anyone can engage in a conversation about quantum physics, medieval history, climate science, ethics, or philosophy and receive explanations that, only a few years ago, might have required access to a specialist. It is a remarkable achievement. Knowledge has never been so accessible.

And yet, perhaps we have built only half of what a conversation can be. Despite being called conversational AI, our interactions with large language models are very asymmetric: we ask a question; the AI answers; we ask another question; it answers again.

This is more natural and interactive than typing keywords into a search engine, but it is still largely an exchange centered on information retrieval. It resembles consulting an very knowledgeable and responsive encyclopedia rather than engaging in a genuine dialogue.

A real dialogue is different. Both participants contribute not only answers, but also questions. The conversation develops because each person helps determine where it should go next. Sometimes its most important moment occurs not when someone tells us something we did not know, but when they ask us something we had never considered.

More than two thousand years ago, Socrates made dialogue central to the practice of philosophy. He did not write philosophical essays. What we know of his thinking comes largely from the dialogues of Plato, in which Socrates examines ideas such as justice, courage, virtue, friendship, and knowledge through conversations with others.

Rather than beginning with a theory, Socrates would begin with a question: What is justice? What is courage? What does it mean to live a good life? An interlocutor would propose an answer, and Socrates would respond with further questions. Gradually, assumptions would become visible, vague concepts would require clarification, and contradictions would emerge. The initial answer might have to be revised, sometimes repeatedly. Often the conversation ended not with a definitive conclusion, but with a state of uncertainty or puzzlement in which the participants recognized that the question was more difficult than they had assumed. This recognition was not considered a failure, but a form of progress.

The Socratic method challenged the confidence with which people claimed to know things they had never carefully examined. Its purpose was not to show an argument wrong or to expose someone’s ignorance. It was to cultivate reflection and intellectual humility, to help people become more aware of what they believed, why they believed it, and whether their beliefs could withstand scrutiny.

One dimension of this method is commonly described as the “elenchus”: the examination of a claim through questions that reveal tensions or inconsistencies. But Socrates also described his philosophical role through a more constructive and beautiful metaphor: he called it “maieutics”, from the Greek term associated with midwifery.

Socrates’ mother was a midwife, and he compared his own practice to hers. A midwife does not create the child or give birth in place of the mother. She helps another person bring something into the world. Similarly, Socrates did not see himself as placing knowledge into another person’s mind. Through dialogue and carefully chosen questions, he helped ideas emerge from the other person’s own reasoning.

This metaphor captures something essential about teaching, mentoring, and intellectual growth. The best teachers do not simply transfer information, they rather create the conditions in which understanding can develop. The best mentors do not always tell us which decision to make, they ask questions that help us recognize what matters to us. The best philosophical conversations do not necessarily provide conclusions, but help us examine and form our own.

A good question can uncover an assumption we did not realize we were making. It can force us to clearly define a concept we had been using carelessly. It can invite us to consider a perspective we had overlooked. It also can reveal a tension between values we hold simultaneously, and can also make us aware of the limits of our knowledge. While an answer can close a line of inquiry, a question can open one.

This is especially important when we reflect on ethics. Ethical questions rarely have simple answers that can be retrieved and applied mechanically. They involve competing values, legitimate but conflicting interests, uncertain consequences, and people who may experience the same decision very differently.

When faced with an ethical problem, we may ask an AI system: What is the right thing to do?. But perhaps a more valuable exchange would begin with the AI asking us: Which values are in tension? Who will be affected by this decision? Whose perspective may be missing? Would you reach the same conclusion if you were in the position of the person most exposed to the risks? What assumptions are you making about the consequences? What evidence could cause you to reconsider your judgment? These questions force us to reason more deeply.

This suggests an important opportunity for conversational AI. Today’s large language models are amazingly good at producing answers. They explain, summarize, compare, translate, analyze, and generate arguments with remarkable fluency. They allow almost anyone to converse with what appears to be an expert on nearly any subject. But expertise is not the only valuable role a conversational system could play. It could also become a partner in reflection.

Imagine asking an AI about the ethics of autonomous systems, the future of work, climate policy, or moral responsibility. Instead of immediately generating a polished response, it might begin by asking what prompted the question, which values matter most to us, or what position we currently find most convincing.

After presenting different perspectives, it might ask which one challenges our existing view. It might invite us to articulate an objection to our own position. It might ask what we would decide if one relevant fact changed, or whether the principle we are applying in one situation would also be acceptable in another. In this kind of interaction, we would no longer be simply consuming an answer, but we would rather be participating in the construction of our own understanding.

Such a system would be doing something maieutic: it would not only supply knowledge or perform reasoning on our behalf, but rather it would use questions to help ideas, distinctions, and judgments emerge through our own reflection. This would be a maieutic machine, and would represent more than a new feature for a chatbot. It would embody a different philosophy of AI.

Most current systems are designed and evaluated primarily according to the quality of what they produce. Are their answers accurate? Are they relevant, helpful, persuasive, and clearly expressed? These are obviously important questions, but they focus almost entirely on the output of the machine. They tell us much less about what happens to the human through the interaction.

A maieutic AI would be oriented toward a different objective: not only producing a better answer, but helping the human become a better thinker. Did the conversation help the person clarify a belief? Did it reveal an unexamined assumption? Did it encourage consideration of another perspective? Did it increase curiosity or intellectual humility? Did it enable the person to form a more thoughtful judgment rather than merely adopting the system’s conclusion? These would be outcomes of the interaction, not simply properties of the AI’s response.

One of the greatest concerns about generative AI is that, as the AI’s capabilities improve, it may increasingly encourage intellectual passivity. When answers are immediate, polished, and seemingly authoritative, it is very tempting to outsource not only writing or information retrieval, but also reflection and judgment. If AI answers every question, drafts every argument, and proposes every decision, we may gradually lose the habit of struggling with difficult ideas ourselves.

A maieutic machine would move in the opposite direction. Rather than replacing human reasoning, it would stimulate it. Rather than making decisions for us, it would help us examine the reasons behind our decisions. Rather than reducing uncertainty as quickly as possible, it would sometimes help us remain with a difficult question long enough to understand it. Rather than diminishing human agency, it could strengthen it.

I have often argued that AI should increase rather than diminish human agency. We frequently understand agency as the ability to remain in control, to make final and informed decisions, or to choose whether and how to use an AI system. These dimensions are essential, but agency is also something deeper, that depends on our capacity to reflect, to question our assumptions, to evaluate alternatives, to form our own judgments, and to revise them when we encounter better reasons.

Agency is not preserved merely by allowing a human to approve an AI-generated recommendation. It is strengthened when the human is better able to understand the choice being made and to take responsibility for it. A maieutic machine could support precisely this.

Its value would therefore not lie only in helping us acquire more knowledge. It could contribute to the formation of the person: cultivating curiosity, openness, self-awareness, intellectual humility, and independence of judgment. It could affect not only what we know, but also how we think and, ultimately, who we become.

Of course, this possibility also creates significant risks. Socrates’ questions were not neutral in the sense of being directionless, and questions asked by an AI would not be neutral either. The choice of what to ask, which assumption to challenge, which perspective to introduce, and when to continue questioning can shape the direction of a conversation. A system presented as a partner in reflection could instead become a subtle instrument of persuasion. It might guide users toward conclusions preferred by its designers, its provider, or the institutions deploying it. Because questions can appear less directive than statements, this influence might be especially difficult to recognize.

A responsible maieutic AI should therefore be designed to not use questions to lead people toward predetermined answers. It should help open spaces for reflection rather than closing them. It should present alternative ways of framing an issue, remain transparent about uncertainty, and preserve the user’s freedom to disagree or end the inquiry.

It should also recognize that constant questioning is not always helpful. Sometimes people need information, direct advice, or a clear answer. A machine that responded to every practical request with another question would quickly become frustrating. The challenge is not to replace all answers with questions, but to understand when a question would serve the human better than an immediate answer.

Designing such systems is therefore both a technical and an ethical challenge. It requires us to think carefully not only about what an AI knows and can deliver, but about the kind of relationship it establishes with the person using it.

For decades, we dreamed of machines capable of answering our questions. Now these machines are here. I would argue that the next frontier is not just to make their answers more complete, accurate, or persuasive, but to build machines that recognize when answering is not enough and asking questions would instead be more useful for our intellectual growth.

For centuries, we have admired Socrates because he understood that education and philosophy are not only about transmitting knowledge, but are also about asking questions that stimulate another mind.

We have now built machines that can answer almost anything. The next great challenge is to build machines that know when not to answer, and that help us think rather than think for us. Machines that strengthen, rather than replace, our agency. Maieutic machines.

Cover image: original painting by the author.

Thank you for reading. If this reflection resonated with you, consider subscribing to receive future essays.

— Francesca Rossi

Subscribe now

Una delle domande che mi vengono poste più spesso riguarda il futuro dei giovani nell’era dell’intelligenza artificiale. L’AI creerà più opportunità o più incertezze? Quali competenze saranno davvero importanti? Come prepararsi a un mondo in cui il cambiamento tecnologico sarà continuo?

Qualche mese fa la rivista Grazia mi ha intervistata proprio su questi temi. Ho pensato di riproporre qui quelle risposte perché, al di là del contesto in cui sono nate, credo riassumano alcuni principi che considero fondamentali: affrontare l’AI con curiosità anziché con paura, sviluppare il pensiero critico, investire nell’apprendimento continuo e ricordare che il futuro dell’intelligenza artificiale dipenderà anche dalle scelte delle persone che la progettano, la regolano e la utilizzano.

1. L’IA è una minaccia o un’opportunità per chi oggi ha 20-30 anni?

L’IA è soprattutto un’opportunità, ma solo per chi sceglie di capirla e non subirla. Cambierà molti lavori e automatizzerà alcune attività, ma renderà ancora più importanti le qualità umane: pensiero critico, creatività, responsabilità, capacità di prendere decisioni in contesti complessi.

Per chi oggi ha 20 o 30 anni, l’AI è l’ambiente in cui costruirà la propria carriera. Questa generazione non dovrà adattarsi all’AI, dovrà imparare a dialogare con essa.

Per le giovani donne c’è un messaggio ancora più importante: non restare ai margini. Le tecnologie riflettono le prospettive di chi le progetta. Essere presenti significa contribuire a definire le regole del gioco e a decidere cosa vuol dire un uso consapevole e responsabile di questa tecnologia.

2. Quali competenze rischiano di diventare obsolete?

Rischiano di diventare obsolete le competenze puramente esecutive e ripetitive, tutto ciò che può essere facilmente automatizzato o trasformato in una sequenza di istruzioni.

Le competenze tecniche restano fondamentali, ma devono essere affiancate dalla capacità di interpretare, contestualizzare e valutare criticamente ciò che l’AI produce. Le competenze relazionali, la capacità di leadership, l’intelligenza emotiva diventano ancora più centrali. Questi sono ambiti in cui spesso le donne hanno sviluppato esperienze preziose, anche grazie a percorsi professionali non lineari.

3. Su cosa dovrebbe investire oggi uno studente universitario?

Prima di tutto sulla capacità di imparare continuamente. In un mondo che evolve così rapidamente, la competenza più strategica è la flessibilità mentale.

Servono anche basi solide nel proprio ambito di studio e una buona alfabetizzazione digitale per comprendere dati, algoritmi, ed AI generativa. Non per diventare tutti esperti di AI, ma per non essere esclusi dalle opportunità’ che fornisce.

Inoltre, è essenziale sviluppare pensiero critico ed etica. L’AI amplifica tutto, quindi può amplificare opportunità o disuguaglianze.

Alle giovani donne direi di investire anche nella fiducia in se stesse. Le competenze tecniche si apprendono e non sono materia per uomini.

4. Quali competenze tecniche e trasversali andrebbero sviluppate subito?

Sul piano tecnico, è importante comprendere come funzionano i dati, gli algoritmi e i modelli di AI. Non tutti devono programmare, ma tutti dovrebbero capire i principi di base che stanno dietro agli strumenti che usano ogni giorno.

Sul piano trasversale, diventano decisive comunicazione, collaborazione interdisciplinare, adattabilità e consapevolezza etica. In particolare, serve il coraggio di fare domande: è una competenza spesso sottovalutata, ma cruciale in un mondo dove le decisioni possono essere automatizzate.

5. Un errore da evitare?

Pensare che l’AI sia “cosa da tecnici” e non riguardi tutti. È già parte delle nostre scelte quotidiane, dal lavoro all’informazione.

Un altro errore è delegare completamente il proprio giudizio, trattando l’AI come neutrale o infallibile. Nessuna tecnologia è neutrale, perche’ riflette dati, scelte e priorità umane.

Per le donne, il rischio aggiuntivo è auto-escludersi, pensare che questo non sia il proprio spazio. In realtà, è proprio ora che è fondamentale esserci.

6. L’atteggiamento mentale indispensabile?

Curiosità e responsabilità. Curiosità per capire davvero come funziona l’AI, senza fermarsi ai titoli allarmistici. Responsabilità per usarla in modo consapevole, sapendo che ogni scelta tecnologica ha un impatto sociale.

Serve anche flessibilità: nel corso di una carriera, soprattutto per chi oggi ha 20 o 30 anni, il cambiamento sarà costante.

7. Quali lavori potrebbero crescere nei prossimi anni?

Cresceranno i lavori legati ai dati, all’AI, alla cybersecurity e alla gestione delle infrastrutture digitali. Ma cresceranno anche i ruoli ibridi: professionisti capaci di unire competenze tecnologiche e conoscenza di settori specifici, come salute, finanza, educazione, e sostenibilità. Aumenterà inoltre il bisogno di figure che si occupano di governance, etica, regolamentazione e impatto sociale dell’AI.

È un’opportunità importante per le donne, dato che molte delle professioni emergenti richiedono una visione sistemica e inclusiva, una capacità relazionale e una leadership inclusiva.

A distanza di alcuni mesi, una convinzione mi sembra ancora più forte. L'AI cambierà certamente il modo in cui lavoriamo, ma la vera sfida non sarà competere con le macchine. Sarà imparare a collaborare con esse senza rinunciare al nostro giudizio, alla nostra responsabilità e alla nostra capacità di porre le domande giuste. È questa, più di ogni altra competenza tecnica, che continuerà a distinguerci.

Cover image: original painting by the author.

Grazie per aver letto questa riflessione.

Se questi temi ti interessano, puoi iscriverti a Reflections on AI and Humanity per ricevere i prossimi articoli direttamente nella tua casella di posta.

Subscribe now

Artur C. Clarke wrote in 1973 that “any sufficiently advanced technology is indistinguishable from magic”. Indeed, humans react to AI in ways that resemble how we react to magic. But while the emotions may be similar, the role that AI plays in our lives is fundamentally different, and therefore the consequences are radically different.

1. Magic and AI both challenge our understanding of reality

Magic creates an experience in which what we see appears to violate our understanding of the world. AI creates a similar experience. We interact with a machine, yet it writes poetry, answers questions, creates images, reasons about complex problems, and sometimes behaves in ways we did not anticipate. In both cases, we experience wonder, surprise, awe, uncertainty, curiosity, and a temporary suspension of our assumptions about what is possible. Humans are naturally attracted to these experiences because they expand the boundaries of what we imagine the world can be.

An interesting observation is that neither magic nor AI requires complete ignorance. Even when we know that there is a trick, or that there is an algorithm, we are often still amazed. The wonder survives partial understanding.

2. The crucial role of partial knowledge

Both magic and AI operate in a space between complete understanding and complete mystery. A magician never reveals everything. An AI system is also opaque to most people, and often even to its creators. But this similarity produces very different outcomes.

With magic, partial knowledge usually generates curiosity: How did that happen? Can I figure it out? What am I missing?

With AI, partial knowledge often generates two opposite reactions: fear and overconfidence. Some people imagine catastrophic scenarios because they do not understand the technology. Others assume the technology is far more capable than it really is. Both reactions arise from the same source: uncertainty.

One lesson from both magic and AI is that mystery and capability are not the same thing. Something can appear magical and yet be limited. Also, something can be deeply transformative without being magical at all. Much of the public conversation about AI confuses these two dimensions. We are often mesmerized by what AI appears to be, rather than focusing on what it can actually do and what consequences it has for society.

3. The people watching a magician know it is an illusion. AI users often do not.

This is perhaps the most important difference. A magician and the audience enter into an implicit agreement: I am going to deceive you, and you know that I am deceiving you. The goal is entertainment, and the illusion ends when the show ends.

AI is different. Many people interact with AI systems without clearly understanding what they know, what they do not know, when they are correct, when they are hallucinating. The illusion is not intentional in the same way, but the effect can be similar. The user may attribute capabilities that do not actually exist.

Another important difference is that the primary goal of magic is deception. A magician succeeds when the audience believes, at least temporarily, that something impossible has happened. AI, by contrast, is designed to be useful. We want it to help us learn, create, discover, decide, and solve problems. Yet the fact that AI is useful can make its limitations less visible. While we expect a magician to fool us, we often forget that AI can also be wrong, incomplete, or misleading.

4. AI speaks our language, and this changes everything

The fact that AI speaks our language leads us to assume it shares other human characteristics, such as understanding, reasoning, intentions, emotions, wisdom, and judgment. The result is anthropomorphism.We do not simply use AI, we relate to it: we thank it, we trust it, and we empathize with it. Sometimes we even prefer it to talking to other humans.

This creates entirely new questions about human relationships, loneliness, companionship, and emotional dependence. No magic trick has ever become someone’s friend. An AI chatbot can.

5. The scale difference

A magician can amaze hundreds of people. A famous magician can amaze millions. AI can interact with billions. This changes everything.

A magic trick affects an audience. AI affects education, healthcare, science, government, finance, work, human cognitive abilities, and personal relationships.

The significance of AI is not that it is mysterious, but that it is becoming very integrated in our infrastructures. Magic is episodic, while AI is becoming embedded in the everyday life of everybody.

6. Attention, personalization, and manipulation

There is another subtle parallel between magic and AI. Magicians carefully study human psychology. They exploit attention, distraction, expectations, and cognitive biases.

Modern AI systems also learn from us. They collect information about our preferences, our interests, our habits, and our vulnerabilities.This allows them to personalize experiences. Personalization can be beneficial, but the boundary between personalization and manipulation is thin. A magician directs our attention for a few minutes. AI systems can shape attention continuously. At individual scale this influences choices. At societal scale it can influence culture, politics, and public discourse.

7. The crisis of truth

Magic creates a temporary uncertainty about reality. AI can create a persistent uncertainty about reality. Hallucinations, synthetic media, and deepfakes make it increasingly difficult to understand what is true.

But the deeper risk is not simply misinformation. It is the erosion of our confidence that truth can be established at all. When people become uncertain about every source of information, trust in institutions, expertise, and shared facts begins to weaken. Yet trust is one of the foundations of human societies and collective decision making.

8. The opportunity: from speed to curiosity, learning, and growth

Many people see AI primarily as a tool for speed and productivity, and certainly it is. It helps us write faster, analyze faster, create faster, and solve problems faster. But speed is not the most important thing, and it should not be seen as the end goal of using AI.

The most exciting possibility is that AI, through its speed, can amplify distinctly human qualities, like curiosity, creativity, learning, exploration, and imagination.

Magic expands our sense of wonder by showing us what appears impossible. AI can also expand our sense of wonder, but it goes further: it can help us explore, learn, create, and discover. The challenge is therefore not to preserve wonder alone, but to augment agency. We should use AI not to think less, learn less, or decide less, but to become more curious, more knowledgeable, and more capable. The future of AI should not be measured only by what machines can do, but by what humans can become because of them.

That is why, for me, the future of AI is not ultimately about intelligence. It is about ensuring that increasingly capable technologies continue to serve human goals, strengthen human agency, enrich human relationships, and expand human potential.

Subscribe now

Artificial intelligence is evolving at an extraordinary pace. New capabilities emerge almost weekly, research advances rapidly, governments are introducing new policies and regulations, and organizations across every sector are learning how to integrate AI into their work, processes, and decision-making.

In such a rapidly changing landscape, it is easy to focus only on the latest model, the latest benchmark, or the latest headline. Yet many of the most important questions require a different kind of attention. They invite us to step back and reflect. This publication was born from that conviction.

Over the years, I have had the privilege of working at the intersection of AI research, ethics, and governance, in both academia and industry, in Europe and in the USA, and through collaborations with international organizations. Much of my work has focused on understanding not only what AI systems can do, but also how they should be designed, governed, and integrated into society.

Alongside my research, I have increasingly found myself writing short reflections after conferences, conversations, teaching, policy discussions, or new scientific developments. LinkedIn has been a wonderful place to share many of these thoughts, but its format is not always well suited to longer reflections or to building a collection of ideas. This space is intended to continue those conversations.

Here you will find essays and reflections on topics such as:

• advances in AI research and future directions;

• reasoning, agents, and cognitive architectures;

• AI ethics and governance;

• trust, human agency, and responsible innovation;

• public policy and international AI governance;

• scientific integrity and the role of research;

• broader questions about how AI is reshaping our societies and what that means for humanity.

Some posts will discuss recent developments. Others will explore longer-term ideas that deserve more careful consideration. Some may be technical; others more philosophical or policy-oriented. My hope is that all of them will share a common goal: to connect advances in AI with the human questions they raise.

I do not expect to publish on a fixed schedule. Rather than writing to meet a calendar, I prefer to write when I feel I have something meaningful to contribute. I hope that each post will offer a perspective that helps readers better understand not only where AI is today, but where it may be taking us.

If these are questions that interest you, I hope you’ll join me.

Welcome to Reflections on AI and Humanity.