AI in education: learning, authorship and assessment

Asking an AI assistant to solve a problem set takes seconds. The question that matters to students, families and teachers is a different one: after that, did the student learn?

This article brings together what studies and official bodies say about AI in education on three fronts: learning, authorship and assessment. At the end, it summarizes what changes in Brazil with the guidelines from the Conselho Nacional de Educação (CNE, Brazil’s National Education Council).

Use is already widespread

In November 2025, Cetic.br (the Brazilian center that studies internet use) released a qualitative study on AI in Brazilian education. According to a report by Agência Brasil, the TIC Educação survey indicates that 70% of high school students and 58% of teachers use generative AI tools in school activities.

The same study records concerns on both sides. Students report fear of losing the ability to learn and of becoming dependent on the technology. Teachers report a decline in writing. And both say they use AI, in many cases, without guidance or rules from the school.

Learning: the effect depends on how AI is used

When AI becomes a crutch

An experiment published in PNAS in June 2025 followed about a thousand high school math students. One group used an interface similar to standard ChatGPT. Another used a “tutor” configured with instructions designed to safeguard learning.

During the exercises with access to AI, performance rose 48% in the first group and 127% in the second. The important result came afterward. On the exam taken without AI, those who had used the standard interface scored 17% lower than students who never had access to the tool. In the group with the safeguarded tutor, this negative effect was largely avoided. The authors describe unrestricted use as a “crutch”.

When AI is designed to teach

A randomized trial published in Scientific Reports in June 2025 compared university students who learned with an AI tutor and students in in-person active learning classes. The tutor was built with the same pedagogical best practices used in the classroom. According to the authors, those who used the tutor learned significantly more, in less time, and reported greater engagement.

The two studies point in the same direction. The tool alone does not determine the outcome. What matters is the pedagogical design: whether the AI does the work for the student or guides them to think. In the editorial team’s assessment, this shifts the question from “use it or not” to “use it how, for what and with what mediation”.

The results apply to the contexts studied and do not automatically extend to every subject, age group or school.

Authorship: what belongs to the student

When a text is produced with the help of AI, the question of authorship becomes less obvious. Checking spelling with a tool is different from asking for a finished text. Between one extreme and the other there are many intermediate uses: generating ideas, building an outline, rewriting paragraphs.

UNESCO’s guidance for generative AI in education and research, from September 2023, advocates a human-centered approach. It recommends setting a minimum age for independent conversations with these platforms and requiring protection of users’ data, and it notes that schools are largely unprepared to validate the tools.

In practice, authorship rules tend to work better when they are agreed on before the activity. A good rule says what is allowed, what must be disclosed and how to disclose it. Knowing how to write clear requests helps (see prompt engineering), but it does not replace verification: language models can invent facts and references, as the article on AI hallucinations explains.

Assessment: why detectors are not the answer

The most immediate reaction to AI-generated text was to look for tools that detect it. Research shows the limits of that approach.

A study by Stanford researchers, published in 2023, tested commonly used AI-generated text detectors. They consistently classified texts by people writing in English as a second language as machine-generated, while texts by native speakers were correctly recognized. The authors also showed that simple rewriting strategies were enough to fool the detectors. The conclusion was a warning against using these tools in assessment settings.

The problem has two sides. A false positive, when the detector flags a human-written text, can lead to punishing someone who did nothing wrong. A false negative, when the detector misses AI text, gives a false sense of security. That is why many assessment proposals bet on other approaches: in-class activities, oral presentation of the work, records of the writing process and questions that require connecting the content to the student’s own experience.

What changes in Brazil with the CNE guidelines

On September 1, 2026, the CNE plenary approved guidelines for the use of AI in schools and universities, according to a report by Jornal do Comércio. The text is an advisory opinion (parecer) and, according to the reports consulted, was sent to the Ministry of Education for approval (homologação), the step that gives it regulatory force. Check the current status with the Ministry of Education before applying the rules.

According to coverage by the newspaper O Hoje, the main points are:

• Four risk levels. Low (organization, accessibility, translation, lesson planning), moderate (virtual tutors and writing assistants, with mandatory human review), high (grading of objective tests and biometric monitoring, with impact assessment and continuous supervision) and excessive, which is prohibited.
• Essay grading. Using AI to grade essays and open-ended questions is prohibited, even if only to suggest a grade.
• Detectors. The result of an AI detector cannot, on its own, justify academic punishment.
• Early years. Children up to the 5th grade of elementary school may not have direct, unsupervised access to generative or conversational AI. Any interaction must be mediated by teachers.
• Student data. Targeted advertising, commercial profiling and commercial exploitation of data are prohibited.
• Deadlines. The prohibitions apply immediately. For the rest, school systems and institutions have 12 months to adapt after approval by the ministry.

In the preliminary version approved in May 2026 for public consultation, the excessive-risk category included emotional surveillance, social scoring, psychological profiling for disciplinary purposes and fully automated decisions on passing or failing students, according to Poder360.

What students and teachers need to learn about AI

Using AI with judgment requires knowledge. In 2024, UNESCO published two AI competency frameworks. The one for students has four competencies: human-centered mindset, AI ethics, AI techniques and applications, and AI system design. The one for teachers has five areas: the same mindset and ethics, AI foundations and applications, AI pedagogy, and AI for professional development.

A practical skill for any age is checking what the machine says. The criteria in how to evaluate AI responses are a good starting point.

Key takeaways

AI can help people learn, but it can also replace the effort that produces learning. The difference lies in the design of the activity and in a teacher’s mediation, not in the tool itself.

On authorship, clear rules agreed on in advance work better than trying to catch use after the fact. On assessment, text detectors are not reliable enough to punish anyone on their own, and the CNE guidelines follow that understanding. For schools and families, the starting point is the same: decide what AI will be used for and who oversees that use.