What Students Should Still Be Able to Do Without AI

Education & Careers9 min read

A student submits an impeccable paper on the causes of the Revolutions of 1848. It has structure, examples and a conclusion. From the text alone, the teacher cannot know with certainty how much was written by the student, how much was rewritten with the help of a chatbot, and how much was copied almost verbatim. But that is not the only question—and not the most useful one.

The pedagogically significant question is different: if the chatbot is closed, what can the student still do with the ideas in the paper?

Can they explain why they chose one cause as central? Can they interpret a new document? Can they verify a factual claim provided by AI? If so, the tool may have supported learning. If not, the paper remains a product that is better than the competence it claims to demonstrate.

This distinction already matters in Romania. The OECD’s PISA 2025 country note, published on 8 September 2026, shows that 40% of Romanian 15-year-olds say they use AI chatbots for learning at least weekly. Some 33% report using them for initial research and for drafting text for an assignment, while 31% use them to summarise a text. These are self-reported data: they describe reported frequency, not how much students learn or the effect of AI on their results.

In a large classroom, the practical response cannot be a lengthy oral defence for every assignment, nor a hunt for supposed “signs” of AI-generated text. Schools need tasks that make understanding visible without turning every student into a suspect.

Two different questions: authorship and learning

A sound AI policy begins by separating two objectives.

What the school wants to find out	The right question	Useful evidence
Attributing authorship	Who contributed to the submitted product, and how?	A brief statement of use, outlines, draft versions, sources.
Checking learning	What can the student do now, without the chatbot’s support?	An explanation, transfer to a new case, verification of a claim, justification of a choice.

The first objective may matter for fairness in assessment. The second is essential to the meaning of the grade. A student may honestly declare that they used AI and still have learned a great deal; another may claim not to have used it but be unable to defend the ideas in their paper. A declaration is therefore not a verdict, and an AI detector should not be treated as decisive evidence.

In its 2025 consultation document on upper-secondary curricula, the Ministry of Education and Research sets out the same direction: assessment should emphasise reasoning, process and the use of knowledge, rather than the final product alone; AI should facilitate learning, not substitute students’ autonomous thinking, reasoning and intellectual effort. This is a curricular direction under consultation, not evidence of a rule applied uniformly across all upper-secondary schools.

What we know—and do not know—about AI support

A randomised experiment conducted in a Turkish upper-secondary school, involving nearly 1,000 students and published in the Proceedings of the National Academy of Sciences, offers a precise warning. In mathematics, access to a GPT-4 interface without pedagogical guardrails improved performance during practice exercises. In the subsequent test, without resources, that group performed significantly worse than the group without AI. In the same study, a tutor configured to provide hints and avoid giving direct answers largely eliminated this negative effect.

For more on digital skills, see also “AI-resistant” jobs: what matters more than the name of the profession, according to the data, which offers complementary details on this topic.

This does not mean that every chatbot “harms” learning. The experiment concerns upper-secondary mathematics, in a Turkish context, using particular versions of GPT-4. Another randomised experiment, conducted at university level in a physics course, reported better results for an AI tutor explicitly designed around active-learning principles. The responsible conclusion is narrower: the technology itself does not determine the outcome; what matters is the kind of help it provides and the task the student must then carry forward.

The European report “Friend or foe? Evidence from the use of generative artificial intelligence in learning and teaching” likewise warns that benefits and risks depend on how AI is used: active student engagement can support learning, while mechanically copying answers can weaken it. Long-term evidence remains insufficient.

A complete example: a history assignment after AI has been permitted

Original assignment: “Explain, in 500 words, why the Revolutions of 1848 had different repercussions in the Romanian lands. Use two identifiable sources.”

Permitted use of AI: the student may ask for research questions, feedback on their own outline and counterarguments. They may not use the chatbot’s response as a source or cite facts or quotations without verification.

The final paper may be good even if AI helped with its structure. To assess understanding, however, the teacher can ask for three pieces of evidence, in no more than five minutes or in a short form.

Explanation: “Why did you place the political factor before the economic one in your argument?” The student must articulate their own cause-and-effect relationship, rather than recite a submitted sentence.
Transfer: the teacher provides a new excerpt from a petition or proclamation. The student identifies which idea from their paper helps interpret that document and what it would qualify.
Verification: the student shows a claim received from or formulated with AI, the source in which they checked it, and what they corrected, retained or qualified.

These checks do not mathematically establish what percentage belongs to the student. They do something more important for a grade: they show whether the student can work with the knowledge displayed in the paper.

The post-AI autonomy test: an InfoBuz proposal, not a standardised test

The proposal below is not a universally validated method and cannot by itself determine whether a piece of work is authentic. It is a principle for designing assessment, inspired by the distinction between performance achieved with available assistance and competence demonstrated after that assistance is withdrawn.

A short checklist: three checks, two process traces

1. Explain: What is the decisive idea or step in your work?

2. Transfer: How does the answer change if I give you a new case, date, condition or document?

3. Verify: Which important claim did you check in another source, and what did you find?

For substantial projects: What did you change after AI feedback? What did you use AI for, and what did you not use it for?

The forms of evidence should be equivalent, not necessarily oral. A student who is anxious, has difficulty with oral expression or has a relevant educational need may respond through an annotated diagram, a 100–150-word written paragraph, a solution with steps recorded, or a short audio recording. Assessment should evaluate the reasoning required by the subject, not stage confidence.

A mini-rubric proportionate to the stakes

Stakes of the task	What is required	What is not required
Routine, low-stakes homework	One explanation question or one step redone in class.	A complete conversation history or an individual defence.
Assessed essay or assignment	An explanation, one source check and an 80–120-word note on AI use.	Policing every phrase or automatically assuming misconduct.
Substantial project or high-stakes grade	Two process traces, a new in-class transfer item and selective verification.	A single check based solely on an AI detector or an impression of writing style.

A quick marking rubric can have four observable criteria: explains the central idea; applies the idea in a new situation; verifies a claim responsibly; transparently declares AI’s contribution. For low-stakes homework, the teacher may assess only the first criterion. For a project, they may use all four.

Rules that protect students, not just assessment

Transparency does not mean submitting the complete history of conversations with a commercial service. As a rule, a declaration limited to the task is sufficient: the tool used, the purpose of use, one idea accepted or rejected, and the sources consulted for verification. Students should not be pushed to submit personal data, conversations unrelated to the assignment or screenshots that expose other people’s information.

Before students begin work, the teacher should announce three things: when AI is permitted, what kind of assistance is permitted, and what evidence of understanding will later be required. A rule announced after submission is closer to a trap than to assessment.

In Romania, this need for clarity arises in a system where policy is still evolving. On 23 April 2026, the Ministry announced that 32 upper-secondary schools would apply alternative framework curricula from the 2026–2027 school year; artificial intelligence is among the possible areas of the student-choice curriculum. This is a limited pilot, not evidence that a uniform practice already exists or that a particular model of AI use has been pedagogically validated at national scale.

What teachers and parents can do tomorrow

Teachers can change just one instruction: instead of “submit the paper”, they can ask students to “submit the paper, explain your main choice and verify one claim in two sources”. In mathematics and science, a similar item with changed data, solved in class, may be worth more than trying to guess whether homework was generated.

Parents can avoid technical surveillance and ask: “What was the idea you changed after checking it?” or “Show me how you would solve it if this condition changed.” Getting stuck is not evidence of guilt; it is a signal that the step deserves to be revisited together.

Students can use AI more intelligently by asking for hints, checking questions and counterarguments, rather than a finished product to submit. A simple rule helps: before accepting an answer, try to explain it; before citing it, verify it; before submitting it, change one condition and see whether you can still use it.

The underlying rule is this: access to an answer is not what is graded. What is graded is the cognitive contribution the student can demonstrate after the answer is no longer on the screen.

Frequently asked questions

If a student cannot explain orally, does that mean they have not learned?

No. Oral explanation is only one possible form. The teacher may accept an annotated diagram, a short written answer, a solution with explained steps or an audio recording. What matters is that the chosen form makes the reasoning required by the task observable, without unnecessarily penalising anxiety or difficulties with expression.

Must a student submit their complete chatbot conversation history?

Generally, no. For transparency, a declaration limited to the assignment is enough: the tool used, its purpose, an intervention accepted or rejected, and the sources used for verification. Requiring the full history may expose personal data and conversations unrelated to the assessment.

Why are AI detectors not the main solution?

Because they do not directly show what the student has learned and should not be treated as the sole verdict on authorship. A combination of process traces, verification of a claim and a transfer item provides evidence closer to the competence the grade claims to measure.