At TEDxUniversity of Tartu on June 2, 2026, Estonia’s Minister of Education and Research Kristina Kallas opened a talk about artificial intelligence by saying she was not going to talk about AI. She was going to talk about humans.
That is the right starting point for AI in education. Schools do not exist to maximize software adoption, produce the longest chatbot sessions, or turn every assignment into a prompt. They exist to change what a learner can understand and do after the lesson is over—especially when the tool is no longer beside them.
Kallas’s argument is that AI puts education under cognitive pressure. If machines can retrieve, summarize, calculate, and draft on demand, then a system built mainly around recalling information and reproducing known procedures becomes easier to bypass. Estonia’s response is AI Leap, or TI-Hüpe: a national attempt to use AI as a Socratic tutor while training teachers to push students toward analysis, evaluation, creativity, and independent judgment.
The slogan is excellent. The implementation is more interesting. And the evidence, one year in, is promising but still incomplete.
TL;DR: what is Estonia actually testing?
Question
Direct answer
What is AI Leap?
A national public-private program combining teacher development, AI access, student literacy, a learning app, and research
Who joined the first year?
154 of 156 upper-secondary schools, about 5,000 teachers, and 20,000 grade 10–11 students
Is the student tool just ChatGPT?
No. ITI is designed to ask questions and support learning instead of immediately completing the task
What is the educational goal?
More analysis, evaluation, transfer, creation, self-regulation, and AI literacy—not more chatbot hours
What has been measured?
Strong teacher adoption and reported teaching redesign; long-term student learning effects are still under study
Cognitive offloading: the student submits an answer while the model, not the student, performed the learning
What should other schools copy?
Teacher-led lesson design, answer friction, explicit learning outcomes, and evaluation after AI support is removed
Estonia’s education minister argues that schools should use AI to strengthen analysis, evaluation, and creation—not replace human thinking.
The real question is not “Should students use AI?”
Students already use it. The useful policy question is: What kind of thinking does the assignment still require from the student after AI enters the room?
An essay can be completed with a model while producing almost no durable change in the learner. The same model can also make an essay harder in the right way: challenge the thesis, demand evidence, surface a contradiction, ask for a counterargument, or require the student to defend why one interpretation is stronger than another.
The interface looks similar in both cases. The learning process is not.
This is why bans and blanket adoption mandates both miss the point. A ban measures whether the technology appeared. An adoption mandate measures whether it was used. Neither measures whether the learner formed a better mental model, noticed an error, persisted through confusion, or transferred an idea to a new problem.
Our AI homework house rules make the same distinction at family scale: AI may help create questions, feedback, and alternative explanations, but the child should still produce the core answer. Estonia is trying to turn that principle into infrastructure for an entire school system.
Lower-order and higher-order thinking: useful model, wrong metaphor
Kallas’s presentation uses the familiar six-part sequence:
Foundational processes
More complex processes
Remember
Analyze
Understand
Evaluate
Apply
Create
This is revised Bloom’s taxonomy, an instructional framework for describing learning objectives. It is useful because it makes teachers ask whether an assessment merely checks recall or requires a student to compare evidence, make a judgment, and build something new.
But it should not be mistaken for literal neuroscience. Humans do not switch between a clean “lower brain mode” and “higher brain mode,” and difficult thought is not simply a separate metabolic gear. Cognitive tasks recruit overlapping systems; complexity depends on prior knowledge, task design, working memory, and expertise. Research on higher-order questions also warns that Bloom labels alone do not tell us whether a question is difficult or whether it produces deep understanding.
The correction matters because foundational knowledge is not obsolete. You cannot evaluate a historical claim if you know no history. You cannot spot a fabricated citation if you do not understand what credible evidence looks like. You cannot critique generated code if you never formed a model of how the program works—a problem we see in AI-driven developer de-skilling.
The better goal is not to skip remembering, understanding, and applying. It is to stop treating them as the finish line.
What Estonia’s AI Leap actually built
Estonia announced AI Leap in February 2025 as a public-private program inspired by its 1990s Tiger Leap, which brought computers and internet connectivity into schools. The initial plan paired national access with teacher preparation instead of dropping a consumer chatbot into classrooms and hoping for the best.
About 5,000 teachers and 20,000 grade 10–11 students were involved
94% of participating teachers used AI in their work
63% said they had redesigned teaching as a result
More than 12,000 students had used the ITI learning app
Around 4,000 students were using ITI actively each week by year-end
A study across about 100 schools and 9,000+ students began measuring learning and AI-literacy outcomes
OpenAI’s February 2025 announcement described a nationwide ChatGPT Edu collaboration. The design evolved into a broader stack: paid model access for teachers, training and professional learning communities, student AI-literacy activities, and ITI, a tutor-style application developed with Estonian researchers and OpenAI.
This distinction matters. Access is not pedagogy. Malta’s national program, covered in our AI literacy and ChatGPT access analysis, makes training a gateway to the tool. Estonia goes one step further by changing the behavior of the learning interface itself.
Why the ITI tutor is deliberately annoying
Students told the AI Leap Foundation that ITI could be annoying because it asks many questions and does not immediately give the answer. Program leaders said that was the point.
Normalizes error and confusion as part of learning
Encourages transfer into a new context
Offers another opportunity to practice
Avoids responses so long that they overload working memory
Remains logical, relevant, and factually accurate
That list is more important than the model name. It describes an interaction contract: the AI should preserve productive struggle while shortening unproductive waiting.
A general-purpose assistant is optimized to be helpful, and “helpful” often means finishing. A tutor should be optimized for what the student becomes capable of doing. Those objectives conflict whenever the fastest answer removes the mental operation the lesson was meant to train.
The Dartmouth Phosphor study we analyzed in our AI tutor evidence review points in the same direction. Its strongest signal came from written-response practice with feedback, not open-ended chatbot use. When the human had to generate an answer, engagement tracked learning. When quizzes became multiple choice, that relationship disappeared.
The OECD number is powerful—but should be stated precisely
Kallas references a 2017 OECD exercise comparing computer capabilities with adult literacy, numeracy, and digital problem-solving tasks. The broad warning is valid: machines were already competitive with large shares of adults on standardized information-processing tasks before ChatGPT.
The precise data is more nuanced than “two-thirds of the workforce had lower skills than computers.” The 2017 exploratory report estimated that 62% of workers used at least one of the assessed skills daily at a level computers were projected to approach by 2026; only 13% clearly exceeded that projected benchmark. It did not say 62% of workers’ overall skill or intelligence was already below computers in 2017.
The OECD’s later 2023 follow-up reported that expert-rated AI performance could potentially outperform 90% of adults in literacy and 57–88% in numeracy, depending on the proficiency threshold. It also estimated that 59% of workers used literacy skills daily at a level comparable to or below computers, while 27–44% did numeracy tasks at or below the assessed AI level.
These were expert judgments on PIAAC test items, not proof that computers were globally “more intelligent” than those workers. But the policy implication remains sharp: education cannot define human capability as whatever machines have not automated yet.
Our take: the scarce skill is cognitive agency
Kallas frames AI as evolutionary pressure. We would phrase it differently: AI creates an incentive-design problem.
The model offers immediate relief from uncertainty. The learner gets an answer, the teacher gets a clean submission, and the institution gets completion metrics. Everyone receives a short-term reward for a process that may quietly remove learning.
The skill schools need to protect is cognitive agency: the ability to form a view before consulting the machine, notice what one does not understand, test a claim, change one’s mind for a reason, and take responsibility for the final judgment.
That is why the correct unit of evaluation is not “quality of the AI-assisted assignment.” It is:
What can the student explain without the chat transcript?
Can they solve a different problem using the same concept?
Can they identify when the model’s answer is wrong or incomplete?
Can they defend the choices in the final work?
Do they know when not to use AI?
This is also the deeper connection to Thinking Machines Lab’s human-centered AI argument: the meaningful benchmark is not what the model accomplishes alone, but whether the human-machine system strengthens human judgment over time.
A classroom pattern teachers can use now
Schools do not need to wait for a national platform to copy the core design. Use a five-step attempt–question–critique–transfer–reflect loop.
1. Attempt before assistance
Require a first answer, sketch, hypothesis, or solution path before the student opens AI. It can be incomplete. The purpose is to activate prior knowledge and create something the learner can compare against.
2. Ask for questions, not completion
Use a tutor prompt such as:
text
Act as a Socratic tutor. Do not solve the task for me.
First, ask what I already know and what I have tried.
Then ask one question at a time that helps me find the next step.
If I make a mistake, point to the conflict without replacing my answer.
At the end, ask me to explain the idea in my own words.
3. Critique the machine
Give students one generated answer containing a subtle error, missing assumption, weak source, or ethical trade-off. Grade the diagnosis and correction—not the ability to produce polished prose.
After supported practice, change the numbers, setting, source, or constraints and remove AI. Transfer reveals whether the learner acquired a portable concept or only navigated one conversation.
5. Reflect on tool use
Ask three short questions: What did the AI help you notice? Where did it make the task too easy? What can you now do without it?
That reflection turns AI literacy from “prompt tips” into metacognition.
What the first-year results do—and do not—prove
The adoption numbers are meaningful. Teacher use at 94%, reported redesign at 63%, and professional learning communities in roughly two-thirds of schools show that Estonia changed school behavior at national scale. That is rare.
But those figures do not yet prove stronger student learning. They are implementation signals and self-reports. ITI usage also fell well below total account availability, and “students found it annoying” can mean productive friction or a product that fails to sustain engagement. Only outcome data can distinguish the two.
The program’s evaluation plan is therefore the part to watch. It includes autonomous motivation, task persistence, learning strategies, self-regulation, prior-knowledge activation, conceptual change, chat-log analysis, standardized testing, and a large randomized trial. OpenAI says it is working with AI Leap, the University of Tartu, and Stanford and has committed to share findings publicly.
Until those results arrive, “world-first experiment” is fair. “Proven model for AI learning” is not.
The policy risks Estonia still has to manage
Even a learning-centered rollout carries hard questions:
Vendor dependence: A national learning layer should remain portable across model providers rather than inheriting one company’s pricing, moderation, and product roadmap.
Student privacy: Chat logs are unusually sensitive records of confusion, beliefs, ability, and behavior. Research access, retention, consent, and deletion rules need public clarity.
Foundational inequality: A Socratic tutor helps only if a learner has enough background knowledge to participate. Students with weaker foundations may need more direct instruction, not simply more questions.
Teacher workload: Designing high-quality AI-supported assignments is difficult. The reported two hours saved weekly matters only if some of that time returns to feedback and lesson design.
Assessment validity: If AI is always available during assessed work, schools need separate ways to measure what each student can independently retrieve, reason through, and create.
For younger learners, staged access is especially sensible. Our middle-school AI curriculum starts with training data, model limitations, media literacy, and supervised creation before asking students to rely on a chatbot for academic work.
The education system should optimize for better humans
Kallas’s most useful move is refusing to make the machine the protagonist.
The purpose of education in the AI era is not to beat a model at remembering every fact or generating the first plausible draft. It is to develop people who can decide what matters, recognize when a fluent answer is weak, connect knowledge across contexts, make ethical judgments, and create work they can defend.
AI can help train those capacities. It can also quietly perform them on the student’s behalf.
Estonia’s real contribution is not putting ChatGPT in schools. It is making that difference—the difference between thinking with AI and letting AI think instead—the central design requirement.
Program participation and first-year figures are accurate as of July 13, 2026. Estonia’s long-term learning study was still in progress; implementation metrics should not be read as causal proof of improved student outcomes.
