Update from the book
AI and education: policies and data
Guidelines, the digital divide, what the other countries are doing, what is taught and how it is assessed. The book's chapter is about people who work; the school is here.
Two measures sum up Italy's position, and they point in opposite directions. On basic digital skills the gap with Europe has almost halved in two years. On ICT specialists Italy has lost 65,600 of them over the same period, while the Union gained seven hundred thousand. Below you will find those figures with the year of the survey next to them, the decree that governs AI in Italian schools together with the state of its implementation, what the other countries have done, what the experiments on learning have measured and, since the book's current edition, what is taught and how it is assessed.
Skills, in Italy
The European indicator on digital skills is surveyed every two years: 2021, 2023, 2025. The reports come out every year, the data does not, and almost all the percentages circulating today are still those of the previous round.
| Indicator | Italy | EU 27 | Survey |
|---|---|---|---|
| People aged 16-74 with at least basic digital skills | 54.27% | 60.40% | 2025 |
| the same measure, previous round | 45.75% | 55.56% | 2023 |
| ICT specialists as a share of total employment | 3.8% | 5.0% | 2025 |
| ICT specialists, number | 904,800 | 10,450,800 | 2025 |
| Share of ICT specialists holding a degree | 41.8% | 68.6% | 2025 |
Eurostat, databases on digital skills and on ICT specialists, queried again on 19 September 2026: the series have not moved, and the latest update Eurostat declares on both is 17 April 2026. The definition of ICT specialist is the OECD one applied to the ISCO-08 classification of occupations.
On the first row Italy has recovered almost four points in two years and the gap with Europe has fallen from 9.8 to 6.1 points. The recovery does not reach where it is most needed: among people with the lowest level of education the share stops at 28.9%.
On the other rows the movement runs backwards. Italian ICT specialists were 970,400 in 2023 and are 904,800 in 2025, while in the Union they rose from 9.7 to 10.4 million. And Italy is the only country in the Union where the majority of ICT specialists do not hold a degree: 41.8%, against 68.6% in Europe, with Cyprus, Spain, Ireland, Lithuania, France and Romania above 80%. The Italian ICT workforce is also among the oldest in Europe, with 70.6% of specialists over thirty-five.
The sentence about firms giving up on AI for lack of skills has to be handled with care. The figure exists and comes from Istat: among Italian firms with at least ten employees that considered AI without adopting it, 58.6% list the lack of skills among the obstacles. Three things have to be held together. It is one box ticked out of six, alongside legislative clarity (47.3%), data quality (45.2%), privacy (43.2%), cost (43.0%) and ethics (25.7%). The denominator is worth 9.4% of firms with at least ten employees, not all Italian companies. And the same percentage in Europe is 70.3%: Italian firms flag that obstacle less often than European ones, so the sentence does not stand as evidence of an Italian lag.
The phenomenon seen from outside: in OECD countries with data available the share of firms using AI goes from 8.7% in 2023 to 20.2% in 2025, and in 2025 those with more than 250 employees are at 52% against 17.4% of those with between ten and forty-nine. How many Italian firms actually use it, and with what gaps by size, is on the page librovivoia.it/en/r/adoption.
The Italian decree, and what it is missing
The measure is ministerial decree no. 166 of 9 August 2025, which approves the Guidelines for the introduction of artificial intelligence in schools, version 1.0. It is not published in the Gazzetta Ufficiale: article 10 provides that it enters into force the day after publication on the Ministry's website, which took place at the end of August 2025. It carries the favourable opinion of the Italian data protection authority of 4 August 2025.
The principles listed in chapter 2 of the Guidelines are six: centrality of the person, fairness, ethical and responsible innovation, sustainability, protection of fundamental rights and freedoms, security of AI systems and models. If you find five listed - ethics, security, inclusion, transparency, human oversight - they come from the 2019 European Ethics Guidelines for Trustworthy AI, which are seven and which the decree refers to in article 8. The operational path has five phases: definition, planning, adoption, monitoring, conclusion. Risk management follows the Council of Europe's HUDERIA methodology.
The high-risk classification of systems used for admission, for the assessment of learning, for determining the level of education a person will be able to access and for detecting prohibited behaviour during tests is not an Italian choice: the Guidelines take it from Annex III of the AI Act. The word "proctoring" does not appear in the document. And the classification is not triggered by purpose alone: it also requires the profiling of natural persons and the absence of the exclusion conditions in article 6.
| Measure | Date | Status as at 19 September 2026 |
|---|---|---|
| DM 166, Guidelines on AI in schools | 9 Aug 2025 | in force since September 2025 |
| Implementing decree provided for by art. 10 of the same Guidelines | — | not adopted. The policy directive of 27 February 2026 announces that it will be submitted to the data protection authority, which as at 19 September 2026 has published no opinion on the matter |
| Circular no. 3392, ban on smartphones in upper secondary schools | 16 Jun 2025 | in force since the 2025/26 school year |
| PNRR call for teacher training on AI, 100 million | 27 Mar 2026 | closed and awarded: 99.96 million to 2,100 schools on 19 May 2026 |
| National Guidelines for the first cycle, DM 221, with "a first approach to artificial intelligence" | 9 Dec 2025 | in force: the 2026/27 school year started in September |
| National Guidelines for the academic upper secondary schools, where AI is assigned to mathematics | 22 Apr 2026 | still out for consultation: the Ministry has not adopted the text |
| Legislative decrees implementing Law 132/2025, with 100 million for training and for AI in civic education | 4 Aug 2026 | approved on final examination by the Council of Ministers and not yet in the Gazzetta Ufficiale. The only legislative decree on AI published, no. 160 of 9 September 2026, covers policing and civil and criminal liability, and does not touch school |
| AI Act obligations on high-risk systems, education included | 2 Dec 2027 | not yet applicable. They were set for 2 August 2026 and the Omnibus moved them |
The AI Act deadlines, the full calendar and the two penalty ceilings are on the page librovivoia.it/en/r/rules.
The ministerial monitoring of digital use, published on 31 March 2026, asks secondary schools whether they have taken up the Guidelines: 45.9% reply that they have adopted them, 50.8% that they are in the process of adopting them, 3.3% that they have not. On the smartphone ban, over 90% of institutions declare they have written it into their own rules. The report does not say how many schools were approached, how many replied, in what period and by what method, and it measures the formal adoption of an act, not what happens in the classroom. As at 19 September 2026 no new edition has come out.
No general obligation to train all school staff on AI came into force on 2 August 2026. This is what the Regional Education Office for Lombardy writes in a note of 2 July 2026, which attributes the spread of the opposite claim to "commercial contexts". Article 4 of the AI Act falls on those who provide or use AI systems in the course of their activity, not on every school by the mere fact of existing.
The smartphone ban has three exceptions, and they are these: pupils with disabilities or specific learning disorders, where the device is provided for by the individualised education plan or by the personalised teaching plan; substantiated personal needs; technical programmes in computing and telecommunications. A general derogation for "authorised teaching use" does not exist, and the ban applies "including for teaching purposes". Computers, tablets and the interactive whiteboard remain allowed.
The money, and who trains the trainers
Scuola 4.0 is worth 2.1 billion euro, an amount never changed by the revisions of the Plan. The rest of what is written about this measure needs correcting. The European target no longer asks for one hundred thousand classrooms to be transformed: with the revision of November 2025 it became the adoption of allocation decrees for at least eight thousand schools, it was brought forward to the second quarter of 2025 and it is recorded as met by a decree of 2022. The one hundred thousand classrooms remain the Italian programme objective and are no longer subject to European verification.
As at 30 April 2026 the administration declares 81.5% of the measure spent. Payments recorded in June 2026 are 1.14 billion out of 1.89 of funding, 54.1%, and completed projects are 7,216 out of 17,362. Transformed classrooms declared by schools are 121,959 out of 173,540 planned. The deadline for reporting expenditure has been moved to 15 October 2026. These figures were not remeasured on 19 September 2026, because the Italia Domani open data catalogue would not open: they remain the ones read on 14 August.
The 100 million for teacher training on AI does not come from Scuola 4.0 but from the PNRR investment in digital teaching and staff training. The call was awarded on 19 May 2026 to 2,100 schools, with 48.4% of the resources going to the Mezzogiorno and at least fifty certificates per project.
The European AI4T trial, which the book cites, closed on 27 February 2024. In Italy it involved 91 upper secondary schools and 438 teachers, within a sample of 302 schools in five countries. It is one of the most serious controlled experiments on teacher training, and its result deserves to be read in full: knowledge of AI grew in all three countries analysed, in Italy classroom use grew as well, but in no country did use become frequent. Teachers tried the tools and did not integrate them.
On the training of Italian teachers there is no official survey. In the United States the EdWeek Research Center has been measuring the same question for three years, and the full series is this: 13% in 2023, 29% between January and March 2024, 42% in October 2024, 50% between September and November 2025. The threshold is low - it counts anyone who has done at least one session, even a single one and even a basic one - and the jump "from 29 to 50" that circulates is obtained by skipping the middle wave.
What the other countries have done
In this table the column that counts is the third. An approved plan and a programme that has started look the same in a newspaper headline, and not in the life of a school.
| Country | What, and when | Where it stands |
|---|---|---|
| China | Guidelines for a spiral curriculum from primary to upper secondary, 12 May 2025. National plan "AI plus education" signed by five administrations, 2 April 2026 | being implemented. The number of hours is set by local authorities and not by the national plan: Beijing asks for at least eight hours a year, Shandong six in the first two years of school and eight from the third to the ninth |
| South Korea | Digital textbooks with AI introduced in March 2025, downgraded by law to optional teaching material in August 2025 after a government veto | adoption down from 32.3% to about 19% of schools in one semester; from 2026 the ministry absorbs them into a broader, non-compulsory category. On 27 August 2026 it starts again somewhere else: sixteen AI subjects, six in primary, two in lower secondary and eight in upper secondary, with approved textbooks written by teachers, in class from March 2028 |
| France | National framework for the use of AI in education, June 2025. Two Pix courses on AI made compulsory in 4e and in 2de générale, that is at thirteen and at fifteen, and in the first year of CAP | compulsory from the September 2026 return to school, to a population of over 1.5 million students a year |
| United States | Executive order on AI education for young people, 23 April 2025, with a task force at the White House | active. Awards for the national student challenge on 9 June 2026 |
| United States | National Academy for AI Instruction, 8 July 2025, 23 million dollars from the teachers' union with Microsoft, OpenAI and Anthropic | stated target of 400,000 educators in five years. As at 19 September 2026 no progress data, neither on the academy's website nor among the union's press releases |
| United States | 50 million dollars on AI within a 169 million fund, awarded on 5 January 2026 | these are funds for universities and colleges, not for schools |
| Germany | DigitalPakt 2.0, 5 billion euro for 2026-2030, half federal and half Länder. Political agreement of 18 December 2025 | started. All the Länder have signed, the federal minister on 31 August 2026, and the pact applies from 1 September 2026 with retroactive effect to 1 January of the same year |
| Finland | UNESCO, UNICEF and ITU Charter for public digital learning platforms, launched in Helsinki on 19 March 2026 | seven guiding principles, without signatories and without binding force. The seventh asks that AI tools be treated as machines and not as entities with human values |
| Eight countries plus Italy | OpenAI's "Education for Countries" programme, announced on 21 January 2026. For Italy the partner is the Italian rectors' conference, not the government | on 20 May 2026 the company declares figures for five countries: Estonia, Jordan, Kazakhstan, Slovakia and Greece. On Italy no number of universities, teachers or students, not even there |
The Korean row is the only case in which a country has reversed course after putting AI at the centre of school by law. The vehicle was digital textbooks with a built-in tutor, launched in March 2025 as official material. Parliament downgraded them to optional material, the government vetoed, Parliament approved again. In the following semester real use showed itself for what it was: on average 8.1% of students opened them for ten days or more, and about 60% never accessed them at all.
In the United States schooling is run by the states, and on the number of those that have given themselves guidelines for K-12 there is no federal census: the two private trackers that keep count do not agree, and neither of them is recent. AI for Education counts thirty-three plus Puerto Rico, with a declared update of 28 October 2025; Ballotpedia counts twenty-seven as at August 2025. A figure higher than thirty-three, as at 19 September 2026, comes from neither of them.
On commercial tutors, the most quoted figure is Khanmigo's: two million users in the 2024/25 school year. The Khan Academy report says "students, educators and parents", and specifies that almost half were teachers, who use it free of charge. The 731% growth measures sign-ups, not use. For the following school year there is no figure yet: in mid-September 2026 the annual report published is still the one for 2024/25.
What the experiments have measured
Controlled studies on generative AI at school are still few and small, and the results depend on how the tool is built and on what is measured. These are the four experiments published in journals that have a control group, plus the most recent meta-analysis.
| Study | Result | On whom, and where |
|---|---|---|
| Bastani and others, PNAS, 25 June 2025 | With an unconstrained chatbot: +48% on the marks of practice exercises, -17% at the exam. With a tutor version that gives hints instead of answers: +127% on practice, negative effect at the exam cancelled out and no positive effect | about a thousand students at a single school in Turkey, mathematics only, four 90-minute sessions in autumn 2023 |
| Fütterer and others, Educational Psychology Review, 14 April 2026 | No measurable advantage on knowledge, engagement and study strategies compared with those using ChatGPT without support. Only perceived usefulness grows | 371 students in the second and third year of lower secondary and the first year of upper secondary, physics and English, six 45-minute sessions |
| Kestin and others, Scientific Reports, 3 June 2025 | With a tutor designed for the course students learn more and in less time than with active teaching in the classroom | university physics course at Harvard |
| De Simone, Tiberti and others, World Bank, May 2025 | +0.31 standard deviations on the final test | nine state secondary schools in Nigeria, six-week after-school programme |
| Wu and others, meta-analysis, 26 March 2026 | Positive average effect on a standardised measure, 0.670 | 35 experimental studies, 4,193 participants, outcomes almost all short-term |
The negative result that travelled around the world is the first one, and it is narrower than the way it gets quoted. The authors write, in the discussion, that they measured short-term outcomes because of a constraint set by the school hosting them, and they point to the study of the long term as work still to be done. The exam comes right after the practice and with computers closed; the sessions were supervised by a teacher and by a member of staff; the models were those of autumn 2023. Using it as proof that AI makes people unlearn attributes to it a measurement its own authors state they did not make.
The comparison the OECD itself places next to the positive results is as useful as the results: the 0.31 of the Nigerian experiment sits below the average of the intelligent tutoring systems that predate generative AI, which reviews place between 0.42 and 0.66. What the experiments separate out is the task: where the student asks for explanations, the effects are positive; where the student asks for the finished text, they disappear.
PISA 2025 came out on 8 September 2026, and for the first time it asks fifteen-year-olds how they use artificial intelligence. The first volume, "Future-Ready Students", measures over 760,000 students in 91 countries and economies and puts the use of AI among the behaviours surveyed, alongside curiosity and perseverance. Italy is above the OECD average in all three subjects: 483 in science against 482, 474 in reading against 461, 468 in mathematics against 463. On AI, in the way the OECD presents the results, students who do not use it for schoolwork do better in science than those who do, with one exception: those who use it once a week to have things explained to them. The figures circulating in the press, some twenty points of gap in science and almost thirty on summarising a reading, are not here: the body of the report does not open from the public record and I have not verified them.
On teachers, the international figure is the one from the OECD's TALIS survey, conducted in 2024 among lower secondary teachers: 36.3% state they used AI in their work in the previous twelve months, an arithmetic average of twenty-seven systems. 71.9% agree that AI allows students to pass off other people's work as their own. Worth bearing in mind before quoting them: TALIS asks about artificial intelligence in a broad sense, and in the questionnaire the definition goes beyond generative AI and language models.
The Guidelines on the Unica platform →
What is taught, and how it is assessed
In the third Italian edition this part was chapter 12 of the book, "AI at school: building the skills of the future". From the current one - the fourth in Italian, the second in English - that chapter is gone: the QR code that leads here sits in chapter 11, "The rules", next to the Italian law, and school teaching lives on this page, where it can be updated with every ministerial circular. The criterion has not changed: in secondary schools AI is taught on three levels at once, the concepts, the practice, critical judgement.
Within the subjects that already exist. The underlying concepts - algorithms, data, learning models - are introduced step by step, and the hard part is doing it without an extra hour on the timetable:
- In mathematics and computing: data visualisation workshops, simulating trends with open datasets.
- In philosophy: debates on ethical dilemmas linked to AI (the responsibility of a self-driving car, bias in recruitment algorithms).
- In history: text analysis and content generation to reconstruct historical relationships or build multimedia narratives.
- In foreign languages: generative applications for creative projects and for practising conversation.
- In science: simulations of complex models, from the evolution of the climate to the spread of epidemics.
Five skills are needed, and only the first is technical. Understanding how the models work and where their errors come from, without a programmer's skills. Judging how reliable their output is, because they produce quality content and disinformation in the same tone. Ethics and governance: privacy, algorithmic surveillance, the responsibility of whoever decides. Using generative AI to amplify creativity, handing over the repetitive tasks. And working in mixed teams, where the person leads and the agent carries out.
An essay written in thirty seconds no longer measures what it used to measure. A ban cannot be checked, and an assignment that can be done with AI without anyone noticing is a badly designed assignment. Four ways of assessing that hold up:
- Documented project work: students use AI but have to document the process, explain their choices and separate their own contribution from the AI's.
- Practical tests in real time: classroom exercises where the use of AI is allowed but supervised, to assess the ability to interact with it critically.
- Oral examinations and presentations: students present and defend their own work, and show that they understand the reasoning behind the solutions.
- Integrated, progressive assessments: tests that combine independent research, guided use of AI, synthesis and critical reflection.
What gets rewarded is the process: how the student analysed the problem, which sources they used, where they stopped to doubt.
AI also helps the people who teach, and above all it helps to personalise. Five uses that already work today: individual paths calibrated to each student's pace and level; alternative explanations, examples set in context and interactive materials for the teacher; a native-speaker partner who answers at any hour and corrects you while you speak; simulations for exploring scenarios and variables safely; structured feedback on coherence, style and argument in the humanities. AI gives most where classes are small and mixed, and it helps inclusion: text-to-speech, rewriting in plain language and, less mature for now, translation into sign language.
Three things tell you whether a plan - national or for a single school - is serious: whether it provides recurring training for the people who teach, and not only equipment; whether it states what students may not do with these tools, and at what age; whether someone measures the results and publishes them. Plans that announce only investment and platforms all look alike, and the table of countries above is to be read with this criterion.
What has changed
19 September 2026 · revision
The decree that carries the hundred million to teacher training is not in the Gazzetta Ufficiale, and the one that has come out does not talk about school. The first implementing decree of law 132/2025 is legislative decree no. 160 of 9 September 2026, in the Gazzetta on 15 September and in force from the 30th: it covers the use of AI in policing and civil and criminal liability. The decree approved definitively on 4 August, the one with the hundred million and with AI inside civic education, as at 19 September is not recorded as published. Anyone reading "AI enters civic education" is reading a Council of Ministers press release, not a rule in force.
The school year has started, and the novelty applies to lower secondary school and not to the academic upper secondary schools. The National Guidelines for the first cycle, with their "first approach to artificial intelligence", apply from this September. The text for the academic upper secondary schools, which assigns AI to mathematics, came out on 22 April and five months on has still not been adopted.
PISA 2025 measures for the first time how students use AI, and the INDIRE survey says how many use it in Italy. The two things this page listed as due in September have arrived: the INDIRE survey on 4 September, PISA 2025 on the 8th. They touch the section on the experiments and the one on the gaps. On Italian teachers the gap remains, and nobody has yet measured how many use AI and with what training.
14 August 2026 · first publication
From September 2026 artificial intelligence enters the Italian curriculum for the first cycle. The new National Guidelines, approved in December 2025 and applicable from the next school year, provide for "a first approach to artificial intelligence". For the academic upper secondary schools the text is still out for consultation, and it assigns the subject to mathematics without a dedicated space. Anyone working in lower secondary school has an operational novelty from September; anyone working in the academic upper secondary schools does not.
The European obligations on the automated assessment of students have slipped by sixteen months. They were due to apply from 2 August 2026 and they apply from 2 December 2027. A school that in the meantime adopts tools to admit, assess or monitor does not yet fall under the high-risk obligations, and remains subject to the GDPR, to the transparency obligations and to Italian law.
The money to train teachers has been awarded, the decree that governs the projects has not. One hundred million went to 2,100 schools in May 2026. Since August 2025 the Guidelines have been deferring to a decree setting out how schools design their AI initiatives, and a year on that decree is not there.
Nine figures not to repeat. They all circulate, and none of them holds up at source.
1. "Only 34% of Italian teachers feel prepared to use AI". It has never been published: it is a subtraction made in a newspaper headline, on a figure that concerned a different question, taken from a survey by an online tutoring company with three hundred teachers in six countries, perhaps fifty of them Italian. The regional rankings derived from it rest on a handful of replies per region.
2. "97% of schools have adopted the Guidelines". It is the sum of two different answers to the same questionnaire: 45.9% have adopted them, 50.8% state they are in the process of adopting them.
3. "57% of teachers say AI helps to write lessons". In TALIS it does not exist in any perimeter. The values are 53.0% across all teachers and 63.7% among teachers who use AI, and they are two different questions.
4. "37% of teachers use AI". It is 36.3%. The 37 appears only in the editorial of the OECD report of January 2026: the chapter that reports the data writes 36.
5. "According to ETS, the training of American teachers has gone from 29% to 50%". The source is the EdWeek Research Center, and the series passes through the 42% of October 2024.
6. "China requires at least eight hours a year of AI education". It is not a national requirement: the plan of April 2026 leaves the number of hours to local authorities, and it is they who set it. Beijing asks for at least eight hours a year, Shandong six in the first two years and eight from the third to the ninth.
7. "Korea integrated AI into the national curriculum at all levels from 2025". AI entered Korean classrooms in March 2025 inside a single resource, which the following August lost its status as a textbook by law.
8. "Four hundred thousand American teachers trained on AI". It is the five-year target announced in July 2025. No progress data has been published.
9. "The meta-analysis that gives ChatGPT an effect of 0.867 on learning". Retracted by the journal on 22 April 2026 for discrepancies in the analysis. It was the most quoted one in favour.
Five things I looked for and did not find. I write them down because they are absences verified at source, and because in at least two cases someone tells them as if they existed.
An official Italian survey of how many teachers use AI and with what training. The only public figure concerns individuals: according to Istat, in 2025 19.9% of Italians aged between 16 and 74 had used generative AI in the previous three months, and among those aged between 14 and 19 the share is 51.2%. Since 4 September 2026 there is also a figure on students, and it is not an official survey: the INDIRE study with the Festival dell'Innovazione Scolastica, 5,342 secondary school pupils between March and June 2026, says that 94% use AI at least occasionally and that 84% use it to study. The authors declare a convenience sample, not a representative one. On teachers there is still nothing.
The report on the Italian trial in four regions, closed with the 2025/26 school year. Press claims about improvements in learning do not rest on a published document. In August 2026 the ANSA news agency places it in fifteen schools in Lombardy, Tuscany, Lazio and Calabria, started in October 2024, and quotes the president of INVALSI: the outcomes "will presumably be ready in December".
The outcomes of the data protection authority's inspections on AI at school, scheduled between January and July 2026.
The current uptake of the Finnish ViLLE platform. The two figures in circulation, half the schools and a third of them, are not compatible with each other and neither is recent: the first is a statement by the university that developed it, from 2021.
The data from the map of trials and from the monitoring dashboard that the Italian decree provides for inside the Unica platform. What this page listed as due in September has arrived: the INDIRE survey on the 4th of the month, the first outcomes of PISA 2025 on the 8th. Still awaited are the second volume of PISA, which carries the new domain on learning in the digital world, and the INVALSI evaluation of the Italian trial.
Sources
- Eurostat, digital skills of individuals, ICT specialists by occupation and by level of education, use of AI in enterprises: databases queried on 14 August and again on 19 September 2026, with the update dates of the tables read on the dissemination API. European Commission, Italy country report of the Digital Decade 2026, updated to 17 June 2026.
- Istat, "Imprese e ICT - Anno 2025", 15 December 2025, and "Cittadini e ICT - Anno 2025", 22 April 2026; no new edition by 19 September 2026. INDIRE and Festival dell'Innovazione Scolastica, survey "Io e l'Intelligenza Artificiale", presented on 4 September 2026. ANSA, August 2026, on the Italian trial and on INVALSI's timing.
- Italian Ministry of Education and Merit, ministerial decree no. 166 of 9 August 2025 and the attached Guidelines for the introduction of AI in schools, version 1.0; circular no. 3392 of 16 June 2025 on smartphones in upper secondary schools; monitoring of digital use of 31 March 2026; policy and institutional directive 2026 of 27 February 2026; DM no. 221 of 9 December 2025 and implementing note no. 1312 of 12 March 2026 on the National Guidelines for the first cycle.
- Italian Ministry of Education and Merit, PNRR call of 27 March 2026 on training for AI and directorial decree no. 34 of 19 May 2026 on the award. Regional Education Office for Lombardy, note of 2 July 2026 on the absence of a general training obligation.
- Italia Domani, open data on the PNRR for the Scuola 4.0 measure: financial framework, targets and milestones, spending by measure, projects and payments of mission 4, consulted on 14 August 2026; the catalogue would not open on 19 September 2026 and those figures were not remeasured.
- Presidency of the Council of Ministers, Council of Ministers press releases no. 177 of 10 June 2026 and no. 185 of 4 August 2026. Legislative decree no. 160 of 9 September 2026, Gazzetta Ufficiale general series no. 214 of 15 September 2026, ELI record read on 19 September 2026. Regulation (EU) 2026/1744 of 8 July 2026 deferring the high-risk obligations.
- AI4T project, Italian national report and European comparative evaluation report, July 2024; Erasmus+ project fact sheet.
- Chinese Ministry of Education, guidelines of 12 May 2025 and "AI plus education" plan of 2 April 2026; Beijing Municipal Education Commission, 2025-2027 plan. Korean Ministry of Education, 2026 work plan of 12 December 2025, government briefing of 26 December 2024 and press release of 27 August 2026 on the AI subjects. People's Daily, 24 August 2026, on the number of hours by province. Pix and the French Ministry of Education, framework of use of June 2025 and press release of 30 January 2026 on the requirement from the 2026 return to school. German federal and Länder ministries, press releases of 18 December 2025 on DigitalPakt 2.0 and of 1 September 2026 on the signature of all the Länder.
- American Federation of Teachers, press release of 8 July 2025; White House, executive order 14277 of 23 April 2025; Department of Education, press releases of 10 November 2025 and 5 January 2026. Khan Academy, annual report on the 2024/25 school year, reread on 19 September 2026: there is none for 2025/26. UNESCO, UNICEF and ITU, Charter for public digital learning platforms, 19 March 2026. OpenAI, press releases of 21 January and 20 May 2026.
- OECD, "Digital Education Outlook 2026", 19 January 2026, and "Results from TALIS 2024: The State of Teaching", 6 October 2025, with the tables annexed to chapter 1; "PISA 2025 Results (Volume I): Future-Ready Students", 8 September 2026, and the Italian scores on Education GPS. Trackers on US state guidance from AI for Education and from Ballotpedia, with their respective update dates.
- Bastani and others, "Generative AI without guardrails can harm learning", PNAS, 25 June 2025; Fütterer and others, Educational Psychology Review, 14 April 2026; Kestin and others, Scientific Reports, 3 June 2025; De Simone, Tiberti and others, World Bank, working paper 11125, May 2025; Wu and others, meta-analysis in Humanities and Social Sciences Communications, 26 March 2026; retraction notice for the Wang and Fan meta-analysis, 22 April 2026.