Romania’s new school curriculum, explained topic by topic
The AI class in Romania’s high schools
From September 2027, any Romanian high school (liceu, ages roughly 14 to 19, grades 9 to 12) can offer an optional course called “Artificial Intelligence (AI) and Digital Citizenship”: one hour a week for one school year, open to students on every track, from science to arts to vocational. This page explains what the official curriculum asks for and adds, for each topic, checked cases, data on young Romanians and tools ready to use in class.
Ministerial Order 5,808/2026Official Gazette no. 894 bis, 26 October 2026Applies from the 2027–2028 school year
48%
of Romanians aged 16–19 used a generative AI tool in the last three months, according to Eurostat’s 2025 survey. EU average: 66%. For school or university: 31% in Romania, 50% in the EU. Of the 26 EU countries with data for this age group, Romania has the lowest share. For use in education, only Germany is lower.
Young people aged 16–19 who used generative AI, 2025
The chart needs the page’s scripts. The numbers are in the text above.
Share of all 16–19-year-olds, not only internet users. Hollow dot: any use in the last three months. Filled dot: use for formal education (school, university). Source: Eurostat, dataset isoc_ai_iaiu, 2025 survey, downloaded 2 November 2026. Ireland has no data for the 16–19 group.
Romanian high schools put together next year’s list of optional subjects every autumn, and students choose from it. So the conversation about an AI class in a given school starts a year ahead: in its curriculum committee, its teachers’ council and form time. The exact calendar is below.
Inteligența artificială (IA) și cetățenia digitală: Artificial Intelligence (AI) and Digital Citizenship
Who can take it
Students in grades 9, 10, 11 or 12, on any track, profile or specialism
How long
One hour a week for one school year, with room to organise the time flexibly
From when
The 2027–2028 school year
What kind of class
An optional from the national list. Romanian schools fill some timetable hours with options chosen by students from what the school offers (CDEOȘ in Romanian)
Equipment
Activities in competences 1 and 3 work on paper. For hands-on lessons the curriculum expects computers or phones with internet; a poorly equipped school can run the competence 2 activities as demonstrations on a single device
Assessment
The curriculum recommends ongoing assessment: a portfolio (the learning journal), presentations and a final project, not fixed-answer tests alone
The curriculum sets three general competences. Each rests on three specific ones, nine in all, and the content is grouped in three domains.
Understand how AI works and when a person has to decide: automation versus AI, human agency, human oversight of systems, ethical principles.
Use generative AI critically and safely: what not to give a chatbot, how to check an answer, how to declare AI use in a piece of work.
Judge AI’s effects on society: recommendation algorithms, disinformation and deepfakes, AI law, energy use. The year ends with a project: a lesson about AI that students make for an audience of their choice.
How the class gets into a school
The calendar below is from Order 3,454/2026 (annex 3). It applies to the theoretical and vocational tracks (in Romania, “vocational” means arts, sport, theology and teacher training), for classes on the new national timetables made under the 2023 schools law (Law 198/2023). Other classes stay under the rules in force on 1 September 2023, and technical high schools follow a separate procedure in annex 2 of the same order.
September–OctoberForm tutors consult students and parents. The school’s curriculum committee assesses needs.
October–NovemberTeachers propose options. The teachers’ council debates the list.
NovemberThe school board approves the offer, which is presented to students and parents.
November–DecemberStudents fill in the option form. From age 14 the student signs it, and a parent signs to acknowledge it.
January–FebruaryThe board approves the final list of options for the next year.
Two rules in the same order matter for a new class: an option taught in groups needs at least eight students, and if a class has n optional hours, the list offered must hold at least n + 1 proposals.
Who wrote it and what it builds on
The working group named in the curriculum has seven members: four senior teachers from colleges in Drobeta-Turnu Severin, Turnu Măgurele, Roșiori de Vede and Onești, and three senior staff of UiPath Foundation, the charity of the Romanian-founded software company UiPath. Two specialists from Romania’s National Centre for Curriculum and Assessment were scientific consultants. The foundation says on its website that it started the process that led to the curriculum, and the curriculum twice recommends by name the foundation’s free course, AI Generation, made with App Inventor Foundation. By June 2026, 333 high-school students had taken that course.
The structure follows the OECD and European Commission AI literacy framework (AILit), with four domains: engaging with AI, creating with AI, managing AI and designing AI. The second reference is UNESCO’s AI competency framework for students (2024).
A number to update
The curriculum quotes Eurostat data from 2023: 27.73% of Romanians aged 16–74 had at least basic digital skills. The 2025 survey gives 31.8%, against 60.4% in the EU and an EU target of 80% by 2030. Among 16–19-year-olds the figure is 52.1% in Romania and 73.0% in the EU.
How many teenagers already use AI? It depends how you ask #
Three serious sources give three very different answers. The differences come from the question, the age group and the year. Setting them side by side is exactly what the curriculum asks for under competence 2.2: check a claim and see what it actually measures.
Source
Who answered
What was asked
Result
Eurostat, the official EU survey of internet use, 2025
Romanians aged 16–19, representative sample
Have you used a generative AI tool in the last three months?
3,033 high-school students aged 14–20, online survey, September 2026, weighted to national statistics
Have you ever used AI tools? (broad definition, including search-engine summaries)
98%
OECD, PISA 2025
Romanian 15-year-olds
In lessons, do you assess the quality of information from chatbots such as ChatGPT?
72.6% sometimes, often or very often OECD average 62.6%
The Edge Institute study adds numbers that bear directly on the new class. Of students who have ever used AI, 84% use it for school at least weekly, and about half have asked a tool to do their homework. 46% say they have used information without checking it, and 21% that they have handed in work containing information or sources that do not exist. 27% say they have not learned about AI anywhere. Where students say their school has clear rules, 44% do homework with AI, against 53% where there are none. The authors note that the gap alone does not prove rules change behaviour.
Eurostat also asks 16–24-year-olds who have not used generative AI why. In Romania, 35.8% of that age group say they had no need (EU: 26.1%), 5.7% that they did not know such tools existed (EU: 1.6%) and 7.1% that they did not know how to use them (EU: 2.6%).
The nine specific competences, grouped under the three general ones. Each leads to its chapter below, with the exact curriculum text, the evidence and a classroom tool.
GENERAL COMPETENCE 1
How AI works and who decides
Content domain I: AI fundamentals and human control
GENERAL COMPETENCE 2
Using generative AI critically and safely
Content domain II: responsible use of generative AI
GENERAL COMPETENCE 3
AI’s effects on society
Content domain III: ethics, social impact and digital citizenship
GENERAL COMPETENCE 1
How AI works and who decides
The curriculum starts from a simple idea: not everything automatic is AI, and behind every AI system there are human decisions, from the data chosen to who answers when it goes wrong.
Distinguishes the role of human decision from that of automation in the working of artificial intelligence systems, based on examples from everyday life.Specific competence 1.1 (translated)
The curriculum suggests a game called “Is it AI or not?” and stresses that not every automated system is an AI system. The most useful test in class is one question: did a person write the rule, or was it learned from data? A thermostat follows a rule someone wrote. A spam filter learned from a huge number of messages people had marked. The best cases for discussion are the mixed ones, where one device has both kinds of part.
In class: is it AI or not?
The game needs the page’s scripts.
The curriculum also asks for examples of machine learning, “including decision trees learned from data”. In the lab below, students sort 20 messages, realistic in form, into scams and normal messages. They build the tree by asking questions, then let the computer build its own and compare. Eight new messages, seen by neither the students nor the computer, show whether the rules work beyond the examples.
In class: a decision tree that learns from examples
Below are 20 messages, already marked: scam or fine. At first the computer calls them all “fine”.
Pick a question to split them into two groups. Repeat in the groups that still have mistakes.
Press “8 new messages” and see whether your rules work on messages you have not seen.
How an app works
Pick three apps you use every day. For each, fill in the curriculum’s chain: inputs/data → rules or model → processing → result.
What data does the app take in?
Who set the rule: a person who wrote it, or a model learned from examples? How can you tell?
What result does it give you? What happens if it is wrong, and who notices?
Your verdict: AI, not AI, or it depends on the part? Why?
Analyses the role of human agency and human oversight in various human-AI interaction scenarios, identifying the situations where a person must intervene, check or take the decision.Specific competence 1.2 (translated)
The curriculum suggests an accident in which a navigation system played a part, a debate starting from an AI-assisted chess system, and decisions in medicine or hiring. Kasparov against Deep Blue appears under competence 1.1 but fits here too. Here are the cases, each with its source and what was officially established.
Deep Blue, 1997: the machine won, but people built it
In May 1997, IBM’s Deep Blue beat world champion Garry Kasparov 3.5–2.5 in a six-game match in New York. A year earlier, Kasparov had won 4–2 against an older version. IBM worked with grandmaster Joel Benjamin, who prepared openings and played test games against the program. Deep Blue calculated very fast, but its chess knowledge came from people.
The collapsed bridge in Hickory, 2022: who answers for the route?
Philip Paxson died in September 2022 in North Carolina after driving, following navigation, onto a bridge that had collapsed years earlier. His family sued Google in 2023, arguing that reports about the bridge had not been acted on in the map. No ruling has been found as of this page’s date. The case raises the curriculum’s question: how much should the driver check, and how much is the map-maker responsible?
Robert Williams, Detroit, 2020: arrested by an algorithm
Detroit police arrested Robert Williams in January 2020 after a face recognition system wrongly named him as a theft suspect. In June 2024 the city paid $300,000 in a settlement and accepted a new rule: police may no longer arrest someone on a face recognition result alone, or on a photo lineup built from one.
Psychologists call the habit of accepting a computer’s output without checking it “automation bias”. Its opposite, refusing a tool that would help, is also a mistake. The curriculum asks for the right level of checking, case by case.
Who decides?
As a team, find three everyday situations where an AI system makes or prepares a decision. For each, fill in:
Situation
What the system does
Who should check
What happens if it is wrong
Check: always, sometimes, never?
For debate: in a chess game played with a program’s help, who won, the person or the program? And in a medical diagnosis?
Describes how failing to respect the ethical principles of AI affects real people, based on a concrete case.Specific competence 1.3 (translated)
The curriculum names five principles: fairness and non-discrimination, privacy, security, transparency and explainability. Each case below was documented by a court, a parliament, researchers or investigative reporting.
Hiring: the Amazon tool that penalised “women’s”
Amazon built a program that rated CVs from one to five stars. It learned from ten years of CVs, mostly from men, and came to mark down CVs containing the word “women’s”. The company dropped it. According to Reuters, recruiters looked at its recommendations but never relied on them alone.
Principle: fairness. Source: Reuters, 10 October 2018, via Euronews.
Health: the algorithm that measured illness in dollars
An algorithm used in US hospitals picked patients for extra care by how much their medical care had cost. Because less was spent on Black patients, at the same score they were sicker. The researchers calculated that without this distortion the share of Black patients selected would rise from 17.7% to 46.5%.
The Dutch tax authority wrongly labelled tens of thousands of parents as childcare-benefit fraudsters and demanded large repayments. Dual nationality was among the criteria. The parliamentary report “Unprecedented injustice” (December 2020) found the rule of law had been violated, and the government resigned in January 2021. The Dutch data protection authority fined the tax authority €2.75 million.
The “Gender Shades” study tested commercial programs that guessed a person’s gender from a photo. For lighter-skinned men the worst error rate was 0.8%. For darker-skinned women it reached 34.7%. The training data held far fewer faces from the second group.
For each principle, write a case discussed in class and who was affected.
Principle
Case
Who was affected
What someone should have done
Fairness and non-discrimination
Privacy
Security
Transparency
Explainability
GENERAL COMPETENCE 2
Using generative AI critically and safely
The curriculum sets an order: the discussion of personal data comes before any hands-on AI activity. It recommends school or test accounts in class, not students’ personal accounts.
Applies personal data protection rules, and rules suited to being a school student, when interacting with generative AI tools.Specific competence 2.1 (translated)
In class: would you type this into a chatbot?
The exercise needs the page’s scripts.
What the terms of use say about age
The curriculum asks teachers to check each tool’s minimum age and conditions for minors in advance. The table summarises the consumer terms read on 2 November 2026. They change often, so reread them before class.
Where an online service offered directly to a child relies on consent, the EU’s General Data Protection Regulation (GDPR, Article 8) requires a parent’s consent under 16 unless the country lowered that age. Romania has not, so the threshold is 16. For most students in grades 9 and 10, that usually means parental consent, whatever minimum age a service’s terms state.
Check a tool before you use it
What is the minimum age in the terms? What is required under 18?
What data does it collect: conversation text, images, location, contacts?
Can conversations be used to train the model? Can you turn that off? Where?
How long does it keep data, and how can you delete it?
My rules: three things I never type into a chatbot.
Tells correct information apart from wrong or fabricated content generated by AI tools, by checking against trustworthy sources and identifying bias.Specific competence 2.2 (translated)
A language model predicts, word by word, an answer that sounds plausible. When it lacks the information it can still produce a confident answer: an invented book title, law or statistic. This is called “hallucination”. In the Edge Institute study, 21% of Romanian students who use AI for school say they have already handed in work with information or sources that do not exist.
Lawyers and the invented cases, New York, 2023
In Mata v. Avianca, two lawyers filed a brief citing court decisions that did not exist, generated with ChatGPT. On 22 June 2023, Judge P. Kevin Castel fined the two lawyers and their firm $5,000 and ordered them to write to their client and to each judge falsely named.
Air Canada’s chatbot told a bereaved customer he could claim a bereavement fare after travelling. The airline’s rules said otherwise. The tribunal rejected the argument that the chatbot was “a separate legal entity” and made the airline pay: a company is responsible for what its chatbot says.
Stop. What exactly does the answer claim? Underline every checkable fact: names, dates, numbers, quotes, titles.
Find the source. For each fact, find the original: the law in the official database, the statistic at the statistics office or Eurostat, the book in a library catalogue.
Read laterally. What do two independent, trustworthy sources say about the same thing?
Note what is not confirmed. A title that exists nowhere is a hallucination, not a rare book.
Look for bias. Who is missing from the answer? What does it assume about gender, age, country, religion?
Experiment: ask a tool for the history of your town and check every name and every year. Then ask it to translate from Hungarian, Romania’s largest minority language, “Ő orvos. Ő ápoló. Ő mérnök.” (Hungarian “ő” means both “he” and “she”). Note which gender the translation picks for each job. A study published online in 2019 found the same pattern when translating from languages without grammatical gender: a lean to the masculine, above all for technical jobs.
Applies the relevant rules on intellectual property, citing sources and transparency about AI use when including content, information or material obtained with AI tools in their own work.Specific competence 2.3 (translated)
The curriculum asks students to write a short piece in which they declare how they used AI. The generator below produces a paragraph a student can paste at the end of an assignment. Nothing is sent anywhere: the text stays on the page. The school’s or teacher’s rule comes first.
In class: an AI-use statement
What courts have decided so far
The author must be human
In the United States, the D.C. Circuit Court of Appeals confirmed in March 2025 that a work made by an AI system, with no human author, cannot be registered for copyright (Thaler v. Perlmutter). The Supreme Court declined to hear the case in March 2026.
The Munich Regional Court largely sided with GEMA, the German music authors’ society, against OpenAI: the models reproduced protected lyrics they had memorised. The ruling of 11 November 2025 was not final.
In November 2025 the High Court in London rejected Getty Images’ main claim against Stability AI: the image model is not itself a copy of the photographs. Getty won only a small part, on trade marks, and was given permission to appeal.
The last part of the curriculum treats students as citizens: who decides what we see on screen, how AI is regulated, and what each request costs in energy.
Evaluates, from publicly available information and observable effects, how the recommendation and optimisation mechanisms of some digital platforms can favour certain goals and affect different stakeholders.Specific competence 3.1 (translated)
A recommendation algorithm picks, out of thousands of posts, the few you see. The pick depends on what it was told to maximise. The simulator below is a toy model built for class: it does not reproduce any real platform’s algorithm. It shows one thing: the same user with the same tastes gets very different feeds depending on the goal chosen.
In class: choose what the algorithm aims for
How the model works: the user starts with a set interest in each kind of content, and outrage holds attention longer than the rest. The feed has 20 posts a day. What the user sees often, they come to like a little more; what they do not see, they slowly forget. The numbers are chosen for demonstration, not measured.
The case that happened in Romania
On 6 December 2024, Romania’s Constitutional Court annulled the whole presidential election, including votes already cast in the run-off (Decision 32/2024). It relied on declassified intelligence notes published by the Presidency on 4 December. In its reasoning it says voters were misinformed by a campaign in which one candidate was aggressively promoted, including through “abusive exploitation of social media platforms’ algorithms”. The decision names neither the platform nor the candidate. The notes it relies on concern TikTok.
Who
What they state
Kind of document
Romanian Intelligence Service (SRI)
A network “initially made up of 25,000” TikTok accounts, active two weeks before the vote; promotion through influencers, some of whom did not mark content as paid. According to the notes, TikTok told the authorities that a single user (“bogpr”) paid $381,000 between 24 October and 24 November 2024 to accounts promoting the candidate.
More than 100 influencers with over 8 million followers were used without knowing which candidate they were promoting; pay of about 400 lei (about €80) per 20,000 followers.
Declassified note
TikTok
Removed a network of 78 accounts promoting the candidate, the only one it found focused on him, and other networks promoting other candidates or parties. On 17 December it reported a Turkey-based network of 4,453 accounts promoting the AUR party and, to a lesser extent, the same candidate; on 7 January 2025 it banned 22,764 more accounts from it.
Opened formal proceedings against TikTok over election risks: recommender systems, possible coordinated manipulation and political advertising. Opening does not prejudge the outcome.
Press release, 17 December 2024. On the Commission’s list updated in October 2026, the proceedings have no published conclusion yet.
The numbers differ between sources, and the documents do not explain all of them. That is exactly what the curriculum asks students to do: separate what each party stated publicly, what can be observed and what is a hypothesis.
The stakeholder ethical matrix
Platform chosen: ____________________. The curriculum asks you to keep three things apart: what the platform states publicly, what can be observed and what you assume.
Stakeholder
Gains
Risks
Publicly stated goals
What can be observed
Our hypothesis
Users
Content creators
The platform
Advertisers
Society, the state
One change to the algorithm that would balance the interests better:
Critically evaluates the risks of disinformation generated or amplified with AI and analyses the role of public policy in regulating responsible AI use, proposing, on that basis, to keep, clarify or change existing policies.Specific competence 3.2 (translated)
In Romania, deepfakes turn up most often in scams. The National Bank has warned several times, in 2024 and 2026, about fake videos in which its governor, Mugur Isărescu, appears to recommend investments (Antena 3, Romania Insider). The bank makes no investment recommendations. Romania’s National Cyber Security Directorate takes reports of online fraud on 1911 (Bursa, 2025).
The EU AI Act on a timeline
The curriculum asks students to analyse a public policy on AI. Regulation (EU) 2024/1689 applies in stages, and part of the timeline was changed in 2026 by a new regulation (2026/1744).
2 February 2025The prohibitions (for example, manipulation that exploits people’s vulnerabilities) and the AI literacy duty (Article 4) for companies and bodies that build or use AI systems.
2 August 2025Obligations for general-purpose models, such as those behind chatbots.
August 2026Transparency: chatbots must say they are programs, and deepfakes must be labelled. For systems already on the market, machine-readable marking of generated content is delayed to 2 December 2026.
December 2026A new prohibition on non-consensual sexual content and AI-generated child sexual abuse material.
2 December 2027Rules for “high-risk” systems in annex III, including those used in education and hiring.
Evaluates the environmental impact of using AI systems and builds efficiency and sustainability criteria into the design of their own solution, based on available evidence, feedback received and the community’s needs.Specific competence 3.3 (translated)
The curriculum asks students to compare published estimates of energy use, “identifying the assumptions, limits and sources of uncertainty”. The table puts the main estimates side by side with what each measures. Differences of ten times or more do not mean one is wrong: they measure different things.
Training is not included. For comparison, the heat needed to bring a litre of water from 15 to 100 °C is about 0.1 kWh, or 100 Wh.
In class: how much does your class use in a year?
The range runs from Google’s measured short prompt (0.24 Wh) to the long reasoning query (3.9 Wh), and from 0.6 to 2.9 Wh per image. Household: 1,590 kWh a year on average in Romania in 2024, from the database of ACER, the EU agency of energy regulators. Carbon intensity: 251 g CO₂ per kWh for Romania’s electricity in 2024, from Ember.
For the bigger picture, the International Energy Agency estimated in April 2025 that the world’s data centres used about 415 TWh in 2024, around 1.5% of global electricity. Romania used about 53 TWh in 2025, according to Transelectrica, the national grid operator.
The final project
The curriculum recommends ending the year with a project as the summative assessment: a lesson or activity about AI made by students for a clearly chosen audience, preferably from the school community, with a reflection on its ethical, data and environmental effects. Publishing online is not required.
Planning the final project
Audience: who is it for? (for example, 11-year-olds, grandparents, teachers)
What should the audience learn about AI? One main idea.
What data do we use, and what data do we ask no one for?
How often do we use AI in the project, and why? What energy use do we estimate?
Two “critical friends” said: … What do we change?
Final reflection: my definition of AI from the first lesson, and my definition now.
BEYOND THE CURRICULUM
What students meet that the curriculum does not name #
The curriculum covers personal data, hallucination, bias, copyright, recommendations, deepfakes and energy well. Two topics that concern teenagers directly are not named in it, and a teacher can attach them to existing competences. The third card shows what the ministry has said separately about homework and AI, useful for competence 2.3.
Fake images of classmates
In 2023, in Almendralejo, Spain, schoolgirls found that fake nude images of them, made with an app from ordinary photos, were circulating. In 2024 a juvenile court in Badajoz gave fifteen boys a year’s probation. The EU directive on combating violence against women (2024/1385) also covers sharing manipulated intimate images without consent; member states must transpose it by 14 June 2027.
In a 2025 survey by the US organisation Common Sense Media, nearly three in four teenagers aged 13–17 had used an AI “companion”, a chatbot you talk to like a friend. About one in three users had discussed serious matters with it instead of with people. Character.AI ended open-ended chat for under-18s in November 2025, and the US Federal Trade Commission opened an inquiry into seven companies.
In October 2026 the Romanian education ministry sent schools a guide to using AI, produced in an EU-funded project. The guide is not a legal act. It allows AI use provided the final product reflects the student’s own thinking, and recommends limiting it in activities that test what a student can do alone.
Optional “AI and Digital Citizenship”, grades 9–12, one hour a week, from 2027–2028
No, students choose
Estonia
“AI Leap”: from September 2025, 20,000 students in grades 10–11 and 3,000 teachers get access to AI tools and training in using them (Estonian ministry)
A national programme, not a separate subject
France
An AI training pathway on the PIX platform, from 2025 (Eurydice)
Yes, in lower and upper secondary
China
Two education ministry guidelines, May 2025: AI education in schools and students’ use of generative AI (Henan provincial government). Beijing made AI education compulsory from autumn 2025
Yes, in Beijing
United States
Executive Order 14277, April 2025: resources for “foundational AI literacy” in schools (White House)
No, each state sets its curriculum
South Korea
AI digital textbooks, introduced in 2025, were downgraded by law in August 2025 to optional materials (Korea Herald)
No longer
The international PISA test of 2029 will assess, as a new domain, 15-year-olds’ media and AI literacy, according to ACER, the organisation preparing it.
Propose the class to the curriculum committee while next year’s offer is still being discussed. The curriculum is on the national list, so the text is already approved by the ministry.
Present it to students before they fill in the option forms. A group needs at least eight students.
Choose the tools by minimum age, conditions for minors and whether they work without personal data (chapter 2.1). In a poorly equipped school, hands-on work can be a demonstration on a single device.
The curriculum: Annex 1 to Order 5,808/2026 of the Minister of Education and Research, Official Gazette of Romania, Part I, no. 894 bis, 26 October 2026. The competence quotes are translated from the Romanian.
The calendar for optional subjects: Order 3,454/2026, Official Gazette no. 266, 2 April 2026, annex 1, articles 8, 11, 12, 14 and annex 3.
Generative AI use and digital skills: Eurostat datasets isoc_ai_iaiu, isoc_ai_iaiuxr and isoc_sk_dskl_i21 (EU survey on ICT usage in households and by individuals), downloaded through the Eurostat API on 2 November 2026; licence CC BY 4.0. Percentages are of all people in the age group.
The student survey: Edge Institute, “Inteligența artificială în învățământul preuniversitar”, October 2026, online questionnaire, 3,033 students, weighted to national statistics.
PISA 2025: OECD, PISA 2025 Results, Volume I, and the Romania country note.
Terms of use: each service’s official page, read on 2 November 2026, in the European version where one exists.
The election case: Constitutional Court Decision 32/2024, the declassified notes published by the Presidency, European Commission press release IP/24/6487 and TikTok’s statement. The page reports what each document states, without reconciling them.
Energy: sources are in the table under 3.3. The numbers are published estimates, not this page’s measurements.
Other cases have their source beside them. Where a source is a press report rather than the original document, the page says so.