PISA 2025, the guidelines discussed by the CNE, and a new code of conduct from Microsoft AI each focus on different aspects of the relationship between people and artificial intelligence. Together , they raise an uncomfortable question for education: when we delegate a task to AI, are we eliminating operational work or the very experience of learning itself?

These three documents—published or debated in 2026—reflect the spirit of our times. An international assessment seeks to examine how students learn in digital environments; the National Education Council establishes guidelines for the use of AI in Brazilian institutions; and a company that develops models states how it intends to keep them under human control. These documents are of different natures, and that difference matters.

PISA does not set rules for Brazilian schools, the CNE does not prove that a particular teaching practice works, and Microsoft does not offer an independent theory of learning. The intersection of these three, however, allows us to more precisely define a problem that is already present in classrooms, teaching meetings, and administrative decisions.

PISA is interested in how students interact with the digital environment

In the “Learning in the Digital World” domain, PISA 2025 assesses two competencies: self-regulated learning and scientific and computational inquiry practices. Specifically , the assessment examines whether students track their own progress, identify gaps in their understanding, adjust their strategies, and use digital resources to explore systems, represent ideas, and solve problems. The focus, therefore, goes beyond simply knowing how to use an interface.

This changes the way we view a digital activity. Imagine two students working on a simulation. Both arrive at the correct answer. One of them tested different possibilities, realized that his hypothesis did not explain the results, and changed his strategy. The other kept trying until he received a confirmation that he was correct. The final result is similar; the learning processes are not.

We must avoid a common pitfall: PISA 2025 does not yet provide results in this area that would allow us to determine how Brazilian students performed. The OECD reports that the results of “Learning in the Digital World” are expected in 2027. For now, the document tells us what the assessment aims to investigate, not what it has already found out about Brazil.

The CNE shifts the conversation from individual choice to institutional responsibility

In schools, the use of AI often begins as an isolated decision: one teacher allows it, another prohibits it; one class reports its use, another doesn’t know what to report. The debate led by the CNE indicates that this is no longer sufficient.

The guidelines approved by the Council in September 2026 address human oversight, student protection, age-appropriate use, and limits on educational decisions made exclusively by automated systems. The legal application of each provision depends on the final text and its ratification status; approval by the Council and entry into force should not be confused.

See also:

• The issue isn't whether to use AI in research. It's knowing how to use it.

There is an educational implication to this discussion. If a school uses AI to suggest exercises, support planning, or analyze information, it must define who reviews the results and is accountable for the decisions. This becomes even more sensitive when assessment, academic progress, and the data of children and adolescents are at stake. Human oversight only makes sense if the person in charge is truly capable of understanding, challenging, and modifying what the system has produced.

Thus, the CNE does not answer the question “What is the best AI-powered class?” on its own . It requires us to ask some preliminary questions:

• Who authorized its use?
• For what purpose?
• Under what conditions?
• What are the options for appeal?

Microsoft proposes a subordinate AI — and admits it isn't there yet

The Humanist AI Code of Conduct, released by Microsoft AI on September 14, 2026, states that its models must remain under human control, promote people’s autonomy, and be transparent about their capabilities and limitations. Its most interesting provision regarding education lies in the distinction between helping someone perform a task and making them progressively dependent on that help to think and make decisions.

The status of this statement needs to be clarified. It is a document issued by the company itself regarding the behavior it intends to develop in its models. Microsoft states that the text is currently open for public comment and is not yet being used to train current models.

It also acknowledges that the effects of prolonged AI use on human cognition require further research. Therefore, the code can be viewed as a commitment and a vision for development—not as proof of current performance or of educational benefits that have already been measured.

There is even a productive tension for administrators and educators: if even those who design the systems say that preserving human autonomy is a goal that is still subject to evaluation, why should schools assume that this autonomy automatically arises whenever a student uses AI?

The documents are similar on one point, but they do not say the same thing

PISA asks what students can do to learn in a digital environment. The CNE asks under what responsibilities and limitations AI can play a role in education. Microsoft asks how a model should behave in order to serve people without slipping out of their control.

See also:

• The next educational inequality will be inequality in the questions asked

The connection between the three is an analytical interpretation: all of them highlight some form of human action, but each places it in a different context. In PISA, it appears in student competencies. In the CNE, it appears in the responsibilities of institutions and professionals. At Microsoft, it is a stated goal for the design and behavior of models. Confusing these positions would create a false convergence. None of the three documents, on its own, provides a universal rule that distinguishes the activities a student can delegate from those they must perform themselves.

This rule depends on the learning objective.

The question that should be brought back to the planning table

Asking AI to transcribe an interview can free up time for students to analyze the testimonies. Asking it to compare the testimonies can be equally useful if the class needs to examine the criteria for comparison. However, if learning to construct and justify these criteria was the goal of the activity, receiving the analysis ready-made eliminates precisely the intellectual exercise that was intended to be taught.

The same applies to writing. There is a difference between asking for feedback on an argument a student has constructed and asking the AI to formulate the argument for them. It is not the button that is pressed that determines the educational value of the use. It is the relationship between what has been delegated and what should be learned.

This reading also raises certain requirements for assessment. A fluent text, an accurate graph, or a convincing presentation demonstrate that a product exists. To assess reasoning, authorship, and learning, it may also be necessary to examine the hypotheses considered, the choices made, the suggestions rejected, and the ability to explain one’s own thought process. These procedures are possible pedagogical implications of the analysis, not identical prescriptions found in all three documents.

Staying up to date on AI in education, therefore, requires more than just learning about new applications or memorizing a list of prohibitions. It requires distinguishing between what we can delegate, what we need to monitor, and what students need to continue doing to develop. This triad is a framework for schools to consider, not a formula prescribed by PISA, the CNE, or Microsoft. Its value lies in forcing each activity to confront the decisive question: Has AI removed an obstacle to learning, or has it simply done the learning we expected of the student on their behalf?

About the author:

OECD. PISA 2025 Assessment and Analytical Framework. Paris: OECD Publishing, 2026. See the chapter “Learning in the Digital World.” DOI: 10.1787/86c36975-en.

BRAZIL. National Education Council. Reference document: Artificial Intelligence — public consultation. Brasília: CNE, 2026. Access the document. Editorial note: This is the version made available for public consultation, prior to the CNE’s approval of the report in September; it should be identified as such in the article.

MICROSOFT AI. Humanist AI Code of Conduct. Sept. 14, 2026. Access the document. Version published for public comment; the company states that it does not yet use it to train its current models.

Share on social media: