Reading, writing, arithmetic, AI.

In February 2024, I published the paper “Embracing the future of Artificial Intelligence in the classroom” (Walter, 2024). According to Google Scholar, it has since been cited around 2,500 times, and it has been accessed more than a quarter of a million times. For a scientific publication, these are exceptionally high numbers. The paper hit like a bombshell because it takes on a hot-button issue: how do we use AI without giving up our own thinking? Recent studies show just how relevant this question remains today.

My answer: it takes three skills that only work together. The first is AI literacy: a basic understanding of how these systems work, what they do well and where they fail. The second is prompt engineering, the ability to give an AI precise instructions. The third is critical thinking. Our new series “Thinking Required” builds on this foundation. Over five articles, we explore this for schools, for personal practice and for professional life. We start with schools, and with two experiments that could hardly be more different.

For six weeks, first-year senior secondary students in Nigeria practiced English with the support of Microsoft Copilot. A World Bank study compared them with a randomly assigned control group. On the final test, which covered AI knowledge and digital skills as well as English, the participants scored 0.31 standard deviations higher, and 0.23 in English alone. In education research, the standard deviation is a common measure of the size of an effect. According to the researchers, this learning gain corresponds to 1.5 to 2 years of regular schooling, and the program is among the most cost-effective ones studied so far. Girls and students who were already performing well at the start benefited most (De Simone et al., 2025).

In Turkey, things went the other way. Nearly a thousand high school students practiced math with a standard version of GPT. As long as they were allowed to use the AI, they solved the problems much better. In the exam without AI, however, they scored 17 percent lower than students who never had access to the AI. A tutor version that gave hints instead of finished solutions largely prevented this effect (Bastani et al., 2025).

Same technology, opposite effect. This brings us to the central question of this first part: what do schools need to build so that AI strengthens learning instead of replacing it?

Some of the most impressive evidence of the potential so far comes from an experiment at Harvard University. Physics instructors developed an AI tutor based on the same teaching principles they use in their in-class lessons, in which students actively work on problems. The tutor guided students through each part of a problem in a fixed order. It relied on solutions the instructors had worked out in advance instead of inventing its own, and it answered questions with individual feedback. In a randomized study with 194 students, learning gains with the tutor were more than twice as large as in class. Students needed a median of 49 minutes for this, compared with about 60 in class. They also felt more engaged and more motivated (Kestin et al., 2025).

The authors also describe what did not work. A system prompt alone, that is, a fixed basic instruction to the AI, was not enough to keep multi-part problems reliably in order. The structure had to be built into the learning platform itself. This shows that prompt engineering is also a teaching tool: how you instruct an AI is not a technical side issue but part of designing the lesson. It determines whether an AI delivers finished solutions or prompts thinking.

Education research knows several forms of so-called scaffolding, that is, targeted learning support by teachers: hints and specific questions, guided reflection, praise and feedback, alternative approaches, clear instruction, and modeling typical mistakes. Each of them can be built into an AI learning companion. But this only works well when it is done by people who know how to teach.

Understanding AI is becoming as fundamental as reading, writing and arithmetic. What sounded like a bold claim only a few years ago is now education policy. On June 18, 2026, the OECD and the European Commission presented the final AI Literacy Framework for primary and secondary education. It organizes AI literacy into four domains: Engage with AI, Create with AI, Manage AI and Shape AI. They cover engaging with AI consciously and critically, creating with AI, deciding responsibly about its use, and understanding how human values shape AI systems. The framework is deliberately not designed as a new school subject but is meant to be woven into existing subjects. Two of its statements could come from any staff room: AI systems are not human and cannot replace human judgment, care or relationships. And refusal or limitation can be a responsible and informed choice (OECD & European Commission, 2026).

In 2029, it becomes measurable. PISA, in which more than 90 countries and economies take part, will then assess Media and AI Literacy for the first time, that is, how well young people deal with digital content, including false and misleading content. According to the Australian Council for Educational Research (ACER), which develops and delivers PISA 2029 on behalf of the OECD, it is the first study of its kind to assess how equipped the next generation is to co-exist with AI (ACER, 2026).

Switzerland shows how great the need is. According to the JAMESfocus report by ZHAW and Swisscom, 71 percent of 12 to 19 year olds use AI tools, rising to 84 percent among 18 and 19 year olds. AI tools are widely used in young people's everyday lives, says co-author Jael Bernath, but a nuanced understanding of the risks and opportunities of AI is often missing. According to her, the same applies to adults (ZHAW, 2025).

The latest edition of the OECD's international teacher survey TALIS also asked how teachers use AI. On average across OECD countries, about 41 percent of lower secondary teachers use AI in their teaching, most often to learn about a topic or plan lessons. Of those who do not use AI, 75 percent cite a lack of knowledge and skills as the reason. Only 12 percent say their school prohibits it. On average, only about 38 percent had AI-related professional learning in the previous year (Education International, 2025; OECD, 2025).

The differences between countries are huge. In Singapore, almost 76 percent of teachers had AI training, compared with 9 percent in France. In Singapore, 75 percent of teachers use AI, in France 14 percent (Education International, 2025). So where more teachers are trained, more of them also use AI. Whether one causes the other, however, these figures do not show.

The investment also pays off for teachers themselves. A survey by Gallup and the Walton Family Foundation in the United States shows that teachers who use AI regularly save an average of 5.9 hours a week. Over a school year, that adds up to about six full weeks (Walton Family Foundation, 2025).

Universities know this too: when a university of applied sciences in Zurich introduced its first rules for using AI, a lack of knowledge was the biggest hurdle. Many lecturers were afraid of using AI incorrectly because they knew too little about it. There were no workshops, no budget and no position responsible for the topic. The TALIS figures show that this pattern is widespread around the world.

None of these steps is new. What is new is that research now shows why they are needed.

Being honest also means naming the open questions. Long-term studies on the effects of AI on learning, teachers and well-being are still missing because the technology is too new. The first randomized experiments exist, but they last weeks, not years. The Harvard tutor was tested in a single course, and the Nigerian program ran for six weeks. Whether the effects last remains open. Schools should therefore design their own pilot projects so that they can learn from them: with a clear question, a comparison group and an evaluation.

For school leaders and authorities

• Plan AI training as a fixed part of staff development, with a budget and working time.
• Appoint a responsible person or team to follow developments and support colleagues.
• Start with a few well-evaluated pilot projects instead of a school-wide rollout.

For teachers

• Try AI for your own preparation first. Those who know the tools also recognize their limits.
• Use study modes or your own instructions to the AI so that it asks students questions instead of giving them solutions.
• Make AI itself a topic in class: have students compare, check and critique AI answers.

Part two is about you. Why do those who know AI best misjudge themselves the most? And how does AI turn from ghostwriter into sparring partner?