Not a Tsunami

Schools aren’t ready for what AI is bringing. But they have everything they need to cross the transition — if we decide to do it properly.

“Both the AI utopians and the AI doomers are far too optimistic.” — Marc Andreessen, Latent Space podcast, 3 April 2026

“The school of the twenty-first century will not be measured by how much it knows, but by how much it is capable of learning continuously, together.” — UNESCO, Reimagining Our Futures Together, 2021

Monday, 11 May 2026, half past noon. The cafeteria of Liceo STEAM International in Rovereto — a four-year Cambridge International upper-secondary school in northern Italy where I teach English and Literature. A hot tray in front of me, pasta with tomato sauce and a salad. Across from me, at the same long table of pale formica, a colleague — an Italian and history teacher who has taught here for several years, one of the most cultured and charismatic among us. I respect him. When I see him, I have the habit of telling him what I have been doing in class.

That morning I had concluded, in my final-year STEAM group, a teaching experiment on Joyce’s Dubliners — specifically Eveline. I had asked the students to have a large language model of their choice rewrite the ending of the story, attempting to maintain Joyce’s voice, his rhythm, his suspension. Not to produce literature — but to force the students to recognise, in the negative, where the machine fails. Where Joyce’s voice is inimitable. Where the statistical pattern breaks against something that has to do with art and is not optimisable. It is what I call algorithmic ghost writing — having the machine write so we learn to read the original better.

I was telling him this. The colleague was listening. And there, off the cuff — I had not prepared it — a sentence came out of me:

“School isn’t ready for the tsunami coming from the world of AI. It isn’t ready because it won’t know how to adapt.”

The colleague nodded. A brief comment, gentle as he is. The conversation moved on to the upcoming referendum, then to the school-leaving exams. The bell rang. Colleagues stood up. I had ten free minutes. I stayed seated.

And I heard the dissonance.

I had just described a sophisticated experiment in a four-year STEAM Cambridge International school — one of the most advanced school environments in northern Italy — and then I had erased the specificity of the scene with a generic sentence about school. I had spoken the way one speaks at a conference, not the way one speaks to a colleague over lunch. I had used the word tsunami, and in that moment I already knew — without quite knowing it yet — that it was the wrong word.

This newsletter is born from those ten minutes of dissonance. But it is not — and does not want to be — a mea culpa. That would be a kind of inverted vanity, dedicating a 3,800-word newsletter to one’s own unfortunate phrasings. What I want to do is different: take that impulsive sentence, take it apart with care, and use the deconstruction as an occasion to say something serious and constructive about schools — Italian, European, global — confronting AI. For my colleagues. For everyone who has invested twenty or thirty years of career in schools. For all those who have every right to feel this transition as exhausting, as unfair in its timing, as imposed from outside. And for students who, while we discuss, are already using AI every day, often badly, and rarely with the awareness they would need.

Let’s proceed in order.

The part of my sentence that was true: schools, today, are not ready

Let me begin where I was not wrong. School systems, taken as a whole, are not structurally ready today for a deep and governed integration of AI into teaching and learning. I am not the one saying it. Four converging sources are.

UNESCO, in its 2023 report AI and the Future of Education: Disruptions, Dilemmas, and Directions (Miao & Holmes, pages 24–31), documents that fewer than 10% of schools worldwide have formal guidance on the use of generative AI in their educational processes. The authors’ words are “largely unprepared.” In 2025 UNESCO published an update — AI and the Future of Education: Reimagined — and the diagnosis has not improved: the gap between the speed of technological development and the speed of institutional adoption has widened.

The OECD, through TALIS 2024–25 — Teaching and Learning International Survey: Results from the Fourth Cycle (OECD Publishing, Paris 2025, chapters 4 and 7 on teachers’ digital competencies), photographs an uncomfortable landscape across member states. The OECD-wide average is that only 36% of teachers report using AI in class at least occasionally. In Italy — to take one national example I know intimately — that figure is 25%. Across most national systems surveyed, more than two-thirds of teachers report they do not possess sufficient skills to integrate AI into their disciplinary teaching. Roughly a third report their schools lack adequate digital infrastructure.

The European Commission’s DESI 2025 — Digital Economy and Society Index (December 2025) registers similar gaps across EU member states. What strikes me most, however — looking at my own country’s data within that report — is the regional variance: 63.9% of schools in Lombardy meet basic digital infrastructure standards, against 27.2% in Calabria. The same kind of variance — between urban and rural, north and south, well-funded and underfunded districts — repeats itself in every national system I have examined, from the United States to Brazil, from the United Kingdom to Australia. Speaking of “schools” as a unitary subject, in matters of AI, is already an abstraction that erases reality on the ground.

The Politecnico di Milano’s Digital Innovation Observatory, in its October 2025 report Generative AI in Italian Schools, published a previously undisclosed snapshot: 64% of Italian teachers over 45 have never used a large language model. 68% of teachers under 35 use one at least weekly. This is an internal generational fracture within the teaching profession that has no precedent in recent school history — and from conversations with colleagues abroad, it appears to be playing out similarly across most national systems.

Four sources, four angles. They all converge on one point: schools, today, are not structurally ready. This part of my impulsive Monday sentence — “not ready” — is supported by the data. It is not opinion, it is documentation.

The mis-gap that should wake us up

There is one element, however, that the four sources do not tell on their own, and which is worth holding together. While roughly two-thirds of teachers report they do not have sufficient AI competencies, and while most teachers over 45 have never used an LLM, our students are already using AI. All of them. Daily. Every report on adolescents and digital media — from AGCOM in Italy to Common Sense Media in the United States, from Ofcom in the United Kingdom to ACMA in Australia — confirms it. Every secondary teacher who has graded a homework assignment in the last year confirms it.

Only they are often using it badly. They use it as an authoritative source — copying without verification. They use it as a substitute for critical thinking — asking directly for the answer instead of constructing one. They use it without recognising hallucinations — convinced that if ChatGPT says it, it must be true. They use it without knowing how to triangulate sources — because no one has taught them that asking the same question to three different models can yield three different answers, and that it is precisely in that discrepancy that critical thinking begins. Some of them, gradually, are learning to use it better — some of my final-year STEAM students are already capable of asking why the model responded a certain way, of interrogating the biases in its training data. But the majority are not. And in the meantime they go on using it.

This is the real mis-gap. It is not just between one national education system and another. It is internal to the classroom itself. On one side, a fifteen-year-old who makes forty interactions a day with an LLM and treats AI as part of their daily cognitive environment. On the other side, a fifty-year-old teacher — highly competent in their discipline, expert in classroom management, formed by an honest and demanding career — who has never used an LLM and who has not received, from the institutional system, the tools to do so. It is not the teacher’s fault. It is not the student’s fault. It is the fault of a system that has not yet built the bridge between the two banks.

When I said on Monday in the cafeteria that “school won’t know how to adapt,” I was speaking — without quite realising it — of this mis-gap. I was not talking about teachers as incapable. I was talking about the system as incapable of closing the distance, on a timeline compatible with that of the technology, between the competencies teachers have (many) and the competencies the profession will require in the next five years (some new, and profoundly different).

An unexpected voice: Marc Andreessen, and what to do with him

On 3 April 2026, Marc Andreessen — one of Silicon Valley’s most powerful and most ideologically declared venture capitalists, co-founder of Andreessen Horowitz, co-creator in the 1990s of Mosaic and Netscape, author in 2023 of the Techno-Optimist Manifesto — was a guest on the Latent Space podcast, hosted by Shawn Wang (Swyx) and Alessio Fanelli. In the last four minutes of the interview (timestamp 1:12:23, section “Why the Real Economy May Resist AI Longer Than Expected”), Andreessen formulated a thesis worth putting on the table for teachers anywhere.

Let me paraphrase, because I cannot reproduce it in full. Andreessen argues that both AI utopians and AI doomers are too optimistic about the speed of change. They agree, he says, on a wrong point: that the speed of social transformation will follow the speed of technical progress. For Andreessen it is the opposite: roughly 35% of the American economy is regulated by professional licences, professional orders, unions, civil service protections. The diffusion of AI will be slowed by these blocs of organised interests — which Andreessen calls, using the word in its technical-economic sense, cartels. On schools specifically, Andreessen is blunt: American K–12 public schools, he says, will never change, because they constitute a state monopoly that will resist any innovation from within. The only path, he concludes, is to create alternative schools outside the public perimeter (he cites Alpha School, a small private Texas chain).

Andreessen is wrong on three specific things, and we should say so with professional clarity.

First: talking about public school systems as a literal monopoly is a rhetorical simplification. In the United States there are roughly 10% private schools, plus charter schools, plus homeschooling. In Italy there are scuole paritarie. In the UK there are independent schools and academies. In every developed country there is a private sector alongside the public one. Literal monopoly is an ideological formulation, not a factual datum.

Second: saying that teachers are 100% opposed to AI is false. The Politecnico di Milano data I cited — 68% of Italian teachers under 35 using AI weekly — explicitly contradicts this reading. Comparable data from the United States, the United Kingdom, Australia, the Nordics tells a similar story. The difference is generational and structural, not monolithic.

Third: the Alpha School example as the only viable path is ideological. Alpha School is a small chain of a few sites, costs tens of thousands of dollars a year, and represents a pilot case set against a population of more than sixty million American students. Citing it as the solution is rhetoric, not analysis.

And yet. Andreessen — a man who shares nothing with public education, who indeed regards it as a structural adversary — catches a kernel of truth that deserves to be named openly among colleagues: public education systems, everywhere, have transformation timelines significantly longer than the technology that traverses them. It is a finding, not a judgement. And it is the point from which we must start to reason about how schools — not by bypassing the public system, as Andreessen would want, but by strengthening it from within — can cross the next five to ten years without being overwhelmed and without betraying their civic mandate.

Where my sentence was wrong

Three points.

First: the tsunami metaphor is wrong. A tsunami is a sudden, total, undifferentiated event. The diffusion of AI in education systems will be none of these things. It will be slow — here Andreessen is right. It will be fragmented — some schools yes, others no. It will be politically contested — unions, professional orders, parties, public opinion will all hold different and legitimately conflicting positions. I take back the metaphor. It was a podcast metaphor, not an analytical one.

Second: saying “schools are not ready” as a monolithic statement erases reality on the ground. Schools, in every country, are a field of forces in which different speeds, different experiences, different contexts coexist. Generalising means saying false things about most of them.

Third: we do not start from zero. Across most countries there are national plans for educational digital transformation, with all their imperfections and all their real spaces for change. There are international research programmes — like the European Erasmus+ KA3 AI4T (Artificial Intelligence for and by Teachers), 2021–2024, which involved hundreds of teachers across Italy, France, Ireland, Luxembourg and Slovenia in structured AI training. There are national-level policy frameworks emerging: Italy’s Ministerial Decree 166 of August 2025 establishing guidelines for AI in schools; the U.S. Department of Education’s 2023 AI report and its successor guidance; Australia’s Framework for Generative Artificial Intelligence in Schools (2024); Singapore’s Educators’ Generative AI Guidelines (2024). These frameworks are imperfect, surely improvable, but they exist. There is UNESCO and OECD work feeding into them. There are national pedagogical research bodies — IPRASE and INDIRE in Italy, the Chartered College of Teaching in the UK, the IFÉ in France, institutions like CRPE in the United States — all working on AI training for school leaders and teachers.

When on Monday I said “it won’t know how to adapt,” I was erasing with three words the work of thousands of colleagues who are already adapting. It is a judgement that, on factual grounds, is unjust toward them.

How schools should cross the transition

Here comes the constructive part — the one that, in all honesty, is missing from the public debate and that we must manage to bring to the foreground. It is not a recipe. It is a work map, articulated on five levels, that I have built reading five years of UNESCO, OECD, European Commission and national reports, and talking with dozens of colleagues in recent months.

First level — Teacher continuing professional development as a structural axis, not as an occasional grant programme

The main error of education policy on AI, to date, has been to treat professional development as a series of grants and ad-hoc projects — a digital recovery fund here, an AI module of a teacher training programme there, a course from a single training provider. They are useful but fragmentary interventions. What is needed is different: AI as a structural and obligatory content of in-service training, embedded within the existing professional development hours already mandated by national contracts and statutes.

Concretely, this means four things. One: every teacher should receive, in the next three academic years, at least 30 hours of basic AI training — technical literacy, conscious prompting, hallucination recognition, critical evaluation of generated sources. Two: every teacher should also receive at least 20 hours of disciplinary AI training — calibrated on their teaching subject, because AI in mathematics teaching is a different thing from AI in literature or foreign-language teaching. Three: training should be recognised in the personnel file and substantively evaluated, not merely certified in terms of hours of attendance. Four: training must be paid or recognised as service hours, because asking teachers to add 50 hours of AI training to their already-mandated professional development without contractual recognition is a formula that feeds only the union resistance Andreessen predicted.

The economic resources of various pandemic-recovery funds expire in 2026 across most jurisdictions. Beyond that, what is needed is ordinary — not extraordinary — funding for AI training, written into the standing education budget. Without this, any reasoning about teacher upskilling becomes rhetoric.

Second level — AI as an organic part of each school’s identity, not as a single project

Today, in most countries, technological innovation is typically entrusted to a system role (the digital coordinator, the digital team, sometimes an AI committee) within the school’s strategic plan. It works in some institutions, fails in many others. The reason is structural: entrusting technological innovation to a single role makes innovation into a sector, not a dimension of the institution’s identity.

What is needed is different. AI — and more generally cognitive innovation — should be one of the four or five load-bearing axes of every school’s strategic plan. At the same level as inclusion, guidance, assessment, openness to the local community. This means: every disciplinary department rethinks its curriculum planning, taking into account how AI changes its subject matter. Every year-group team dedicates at least two meetings a year to reflection on student AI use. Every faculty meeting has a fixed innovation and AI item on its agenda at least three times a year. The school leader — a figure ever more crucial in this phase — receives dedicated AI training, because the pedagogical leadership of an institution in the next five years will require competencies that traditional headship preparation has never included.

It is a change of organisational grammar, not of single interventions.

Third level — The pedagogical bridge between students and teachers

The mis-gap I described above is not closed by teacher training alone. It is closed by parallel work with students. This means: every disciplinary curriculum — from literature to mathematics, from foreign languages to history — should include explicit learning units on the critical use of AI within that specific subject. Not as a standalone subject (yet another mistake: AI is not a subject, it is a cognitive environment). But as a transversal dimension of every discipline.

Concretely: in a literature unit, the student learns to use AI to explore interpretive hypotheses, and then to verify them against the text. In a maths unit, the student learns to use AI to generate solution hypotheses, and then to prove them rigorously. In a history unit, the student learns to use AI to reconstruct context, and then to triangulate with primary sources. AI as helper, never as oracle. It is the model I call, in my own teaching, What If + AI + Creativity in 5 Phases — AI as provocateur, not as substitute.

On this front there are consolidated experiences to start from, in every national system. The work of Mary Helen Immordino-Yang on the neuroscience of learning; in Italy, Daniela Lucangeli on mediated learning and warm cognition; the OECD’s Future of Education and Skills 2030 framework; the studies emerging from Stanford’s Graduate School of Education and University College London’s Institute of Education on AI literacy. We are not starting from zero.

Fourth level — Continuous innovation, not episodic events

There is a phrase I have heard repeated in faculty meetings over the past three years that has always made me think. “Now that we have introduced the digital, we can move on.” It is the typical phrase of a system that thinks of innovation as a one-shot action — a series of projects that close. It does not work that way. The generative AI of today (May 2026) does not resemble that of three years ago, and will not resemble that of three years from now. The AI competencies of a teacher must be updated continuously, as those of a doctor, an engineer, a lawyer are updated.

What is needed, then, is a culture of continuous updating built into the school’s structure. It means: every institution should have — on par with disciplinary departments — a cognitive innovation observatory, made up of five to eight volunteer teachers, that meets every two months, monitors technological developments, critically evaluates them, returns them to the faculty in the form of brief reports, and proposes adjustments to the strategic plan. It means: every teacher should have, in their annual professional development plan, at least 10 hours dedicated to self-directed AI learning — readings, podcasts, personal teaching experiments, exchanges with colleagues in other schools. It means: schools should build thematic networks on AI didactics, as they have historically built them on inclusion or guidance, sharing materials, case studies, mistakes.

It is a model that requires time, patience, and — above all — the institutional recognition that the teacher’s work has changed. We do not teach the way we taught in 2015. We will not teach in 2030 the way we teach in 2026. This is a truth that school systems must build into their structures, not endure as a sequence of emergencies.

Fifth level — Defence of school as a public good

Everything I have said above makes sense only inside a clear political frame: the public school — accessible to all — is a public good to defend, not a business to disrupt. It is exactly the opposite of what Andreessen says. He proposes to bypass the public system by building Alpha Schools next to it, for the few. That, in any country, would be a civic disaster. It would transform AI from an opportunity for mass cognitive literacy into a privilege for those who can pay. It would destroy the principal social elevator of every modern democracy — ever more imperfect, ever more under pressure, but still functioning in many of its parts.

The challenge is double. On one hand, seriously modernising public schools on AI, with the four levels described above. On the other, defending the principle that modernisation is not done against the public system, but through the public system — even if it is slower, even if it is more tiring, even if it requires more patience. It is what UNESCO IESALC, in its 2022 report Thinking Higher and Beyond — Perspectives on the Futures of Higher Education to 2050, calls a global public good. Three words together: good, because it has intrinsic value; public, because it belongs to everyone; global, because knowledge has no borders.

A word to teachers reading this newsletter

I want to close — before the final considerations — with a direct word to the teachers reading me, wherever you are. Some of you have twenty years of career behind you. Some thirty. You have shaped generations of students, you have crossed five reforms of your national school system, you have learned to manage increasingly complex classrooms, you have built serious teaching competencies that are not learned from a manual and that no AI will ever know how to replicate. What is asked of you today — to add AI literacy to your competencies, to rethink your teaching, to keep pace with a technology that changes every six months — is a difficult task. It is fair to feel it as difficult. It is fair to feel it, also, as unjust in its timing: no serious reform happens in three years, and AI is asking us to do in three years what school systems have historically done in twenty.

What is not fair — and here yes, it is worth saying openly — is using that difficulty as a shield for not moving. Because our students are not waiting. They are using AI now. They are often using it badly. And if we do not accompany them to use it well — we, who have behind us disciplinary training, experience, the ability to build critical thinking — we will leave them alone. Alone with a statistical oracle from which they ask for answers instead of constructing them. It is exactly the opposite of what public education has done over the last two centuries.

The AI transition in schools is not made against teachers. It is made with teachers, and also for teachers — because the truth is that whoever masters AI in their profession today will have powerful tools to teach better, evaluate better, build better materials, reduce bureaucratic fatigue. Well-used AI, for a teacher, is a liberation of cognitive time to dedicate to what machines will never know how to do: looking a single student in the eye and understanding why today she is not working the way she was yesterday.

Monday’s coffee

Next Monday at ten in the morning, I will go and bring a coffee to my Italian and history colleague. I will tell him that last Monday in the cafeteria I said a generic, slightly hasty sentence, and that it has cost me five days of work to put it right. He will probably smile. He will probably tell me it was not hasty, only simplified. He will probably — I know him — quote from memory a passage from Manzoni in which someone says something similar and then regrets it, and between the lines he will place a joke about school and one about AI.

And perhaps — this is the thing I really hope for — we will talk about how, in the next three years, our school could build a structured AI training plan for the entire faculty. Not a single grant. Not a single training module. A three-year plan, integrated into the school’s strategic vision, evaluated and rethought annually, with recognised resources and recognised time. It is what should happen, by 2029, in every secondary school across the developed world.

Public schools — made up of teachers like my colleague, like me, like millions of others worldwide — have all the intellectual resources to cross the AI transition without being overwhelmed and without betraying their civic mandate. What they lack, today, is a clear institutional framework — a national plan that builds AI training into service hours, ordinary funding that continues beyond pandemic-recovery measures, contractual recognition of the new teaching work, structural evaluation of institutions on AI integration. These are things that depend on ministries of education, unions, regional authorities. They are things that can be done. They are things that, if not done, will lead schools exactly to the point at which Andreessen thinks they are destined to arrive. And that would be a historic failure.

Schools are not a tsunami. They are not a cartel. They are not an Alpha School waiting to be born. They are a patient, slow, stubborn place, in which every day millions of adults attempt to do well a serious job in front of generations of adolescents who deserve to be accompanied. It is an adult profession. It is a civic institution.

Facing AI, it needs time. But time, today, is precisely what it is being told it cannot quite have. The challenge — the real challenge of the next decade — is to find the way to accelerate without betraying, to modernise without disintegrating, to integrate AI without it becoming the new statistical god of a generation that, without critical thinking, will be poorer, not richer. That challenge will not be won in Brussels, and not in Washington, and not in Rome. It will be won in the school cafeterias of Rovereto and Lyon and Leeds and Chicago and Sydney and Singapore — between a plate of pasta and a school bell, when one colleague listens to another colleague describe an experiment, and someone, after some thought, decides to take it seriously.

Sources

1. Miao, F. & Holmes, W. (2023). AI and the Future of Education: Disruptions, Dilemmas, and Directions. UNESCO Publishing, Paris.

2. UNESCO (2025). AI and the Future of Education: Reimagined. Update report, Paris.

3. OECD (2025). TALIS 2024–25 — Teaching and Learning International Survey: Results from the Fourth Cycle. OECD Publishing, Paris, chapters 4 and 7.

4. European Commission (December 2025). Digital Economy and Society Index (DESI) 2025 — Italy Country Report. DG CNECT, Brussels.

5. Digital Innovation Observatory, Politecnico di Milano (October 2025). Generative Artificial Intelligence in Italian Schools. Milan.

6. Italian Ministry of Education and Merit (5 August 2025). Ministerial Decree no. 166: Guidelines for the Introduction of Artificial Intelligence in School Institutions. Rome.

7. U.S. Department of Education, Office of Educational Technology (2023). Artificial Intelligence and the Future of Teaching and Learning. Washington, DC.

8. Australian Government Department of Education (2024). Australian Framework for Generative Artificial Intelligence in Schools.

9. Ministry of Education Singapore (2024). Educators’ Generative AI Guidelines.

10. European Programme AI4T — Artificial Intelligence for and by Teachers. Erasmus+ KA3, 2021–2024.

11. UNESCO (2021). Reimagining Our Futures Together: A New Social Contract for Education. Paris.

12. UNESCO IESALC (2022). Thinking Higher and Beyond: Perspectives on the Futures of Higher Education to 2050. Paris.

13. Andreessen, M. Latent Space Podcast, episode of 3 April 2026, interview by Shawn Wang and Alessio Fanelli, timestamp 1:12:23, section “Why the Real Economy May Resist AI Longer Than Expected.” Available on YouTube and latent.space.

14. Andreessen, M. (October 2023). The Techno-Optimist Manifesto. Published on a16z.com.

15. Immordino-Yang, M. H. (2016). Emotions, Learning, and the Brain: Exploring the Educational Implications of Affective Neuroscience. Norton, New York.

16. Lucangeli, D. (2019). Cinque lezioni leggere sull’emozione di apprendere. Erickson, Trento.

17. OECD (2018). The Future of Education and Skills: Education 2030 Framework. OECD Publishing, Paris.

18. AGCOM (2025). Children and Adolescents in the Digital Age: Annual Report. Rome.

19. Common Sense Media (2024). The Common Sense Census: Media Use by Tweens and Teens. San Francisco.

20. Ofcom (2024). Children and Parents: Media Use and Attitudes Report. London.