If AI Can Do the Homework, What Was Homework Measuring?

On August 6, 2026, Denmark's Ministry of Children and Education announced emergency measures addressing AI cheating in upper-secondary education.

The package has three pillars.

First, an oral defense will accompany certain examination papers written at home. The initial target is the SSO, a large written project completed each year by approximately 9,000 students in the HF program. Students will have to explain their work and answer questions about it.

Second, schools will be encouraged to use tools that monitor computer activity during written examinations and systems that restrict network access.

Third, more assignments will be written in controlled school settings so that teachers can follow the process, give feedback, and assess the student's work more accurately.

This is not the abolition of homework. But it expresses distrust in a system that evaluates ability from a polished text completed outside school.

Generative AI can summarize, outline, compose, solve equations, and write programs in seconds. A 2026 RAND survey reported that the share of American secondary and college students using generative AI for homework rose from 48 percent in May 2025 to 62 percent in December.

Two extreme reactions follow.

One says, “Every student who uses AI is cheating. Monitor them and ban it.”

The other says, “If AI can do everything, homework, memorization, and writing are obsolete.”

Neither asks carefully enough why homework existed in the first place.

The central claim of this essay is:

Generative AI did not break homework. It broke the inference that seeing a finished submission is enough to see the student's learning.

Homework produces at least two kinds of result:

1. an external artifact—a report, calculation, or program; and
2. an internal change in the learner—ways of thinking, skills, memory, and judgment.

I will call these the two products of learning.

AI is powerful at producing the first. If schools truly want to cultivate and measure the second, they must stop relying on the artifact alone.

1. Why did Denmark turn to oral defense?

Denmark's response focuses less on mechanically detecting AI-generated language than on verifying that students understand what they submitted.

During an oral defense, a teacher can ask:

• Why did you choose this source?
• Can you explain this paragraph's claim through a different example?
• How would you answer an opposing view?
• What happens if we change the premise?
• At what point in the research did your thinking change?

Even if AI assisted with a draft, a student who examined the material and adopted the conclusion as their own judgment can answer. If the student merely pasted a finished output, the gap between textual polish and personal understanding becomes visible.

An oral defense does not directly detect whether AI was used. It tests whether the submission corresponds to the learner's understanding.

The distinction matters. If the only objective is detecting misconduct, students and teachers become mutual suspects. If the objective is verifying learning, teachers can ask—even when AI was used—what the student understood, what they delegated, and where they exercised judgment.

2. What was homework for in the first place?

The single word “homework” combines several purposes.

Practice

Calculation, vocabulary, music, and physical skills become automatic through repetition. The value lies in the learner performing the repetitions, not merely in the finished sheet.

Checking understanding

Teachers need to know whether students can explain a lesson and apply it to a problem. The answer is treated as evidence of understanding.

Inquiry

Students research, read, experiment, and create beyond what class time permits. They formulate questions, compare sources, and think over longer periods.

Self-management

Students learn to plan before a deadline and work independently.

Assessment

The submission contributes to grades, promotion, and selection.

AI affects each purpose differently.

AI as a foreign-language conversation partner can increase practice. Asking for background on a difficult text can support inquiry. Having AI perform arithmetic practice yields correct answers but no skill. Generating an entire paper used for grading destroys the submission's value as evidence of understanding.

“May students use AI for homework?” is therefore too broad a question.

What is this assignment trying to achieve? Relative to that purpose, is AI a scaffold or a substitute?

3. The two products of learning—an answer is not an ability

Imagine a baking course.

The assignment is to bake a loaf of bread. A student buys one from a famous bakery and submits it. Its appearance and taste may be perfect, but it proves nothing about the student's understanding of fermentation or temperature.

Another student consults a professional while baking, then explains why the first attempt failed. Learning may have occurred even with outside help.

Writing works the same way.

A perfect AI-generated paper is an excellent artifact. It does not establish that the student can read sources, construct an argument, or answer objections.

Conversely, a student may use AI to check grammar and expose weaknesses in an outline while retaining responsibility for the final reasoning. That can be the product of learning with a tool.

Schools are confused because they have long tried to measure both products through one sheet of paper.

Answers and abilities were never perfectly aligned before generative AI. Parents, tutors, friends, translation software, textbooks, and internet searches already intervened. AI made the separation cheap, fast, and scalable.

It did not create an entirely new hole in assessment. It widened a seam that was already there.

4. Plato—does a new technology of memory weaken thought?

Plato's Phaedrus contains a myth about the invention of writing. The inventor claims that writing will increase memory and wisdom. The king replies that dependence on writing will weaken memory and give people only the appearance of wisdom.

The exchange resembles fears about generative AI:

• Search means no one remembers.
• Calculators mean no one can calculate.
• AI writes, so no one thinks.

Yet we did not abandon writing. It externalized part of memory while making long arguments, communication across distance, law, and science possible. The issue is not whether to use an external tool, but which abilities to entrust to it and which to retain within ourselves.

A navigation app is useful, but we still need judgment when it sends us down a dangerous road. Machine translation is useful, but we must recognize a mistranslation that creates contractual risk. Writing AI is useful, but we must detect fluent misinformation.

The purpose of education is not purity from tools. It is the capacity to understand what has been delegated and to object when the tool should not be followed.

5. Dewey—learning is the experience of inquiry, not possession of the answer

John Dewey understood education not as transferring finished knowledge but as an experience of encountering a problem, forming a hypothesis, testing it, and revising thought in light of the result. Critical thinking is less the speed of finding an answer than the habit of examining reasons and updating judgment.

From this perspective, the value of homework is not confined to the final paragraph.

• Where did the learner become confused?
• Which source did they doubt?
• Why did the first idea fail?
• What changed after feedback?
• Can the insight be applied to another problem?

Generative AI can either support or erase this inquiry.

“Write the answer” shortcuts the encounter with the problem. “Show three weaknesses in my hypothesis,” “Give me a counterexample,” or “Give me only a hint” can turn AI into a partner in inquiry.

The same technology can play opposite roles in the learning process.

Schools should therefore teach which actions substitute for the learning objective and which support it, rather than regulating only the name of the tool.

6. Is using AI itself cheating?

Cheating is not determined by a tool's mere presence. It depends on the assignment's rules and the representation made at submission.

Using a calculator in a calculator-free examination is cheating. Using one in a statistics assignment where it is allowed is not. Having a friend write an individual paper is misconduct; collaboration is legitimate when the assignment requires joint research.

The same distinctions apply to AI.

Clear substitution

The student has AI generate an answer that was explicitly supposed to be written by the student, then presents it as personal work. The assessed ability has been misrepresented.

Permitted support

The student checks grammar, organizes ideas, or generates practice questions within a scope approved by the teacher and discloses the use.

The boundary zone

The assignment states no rule. The student sees AI as an extension of search, while the teacher regards it as ghostwriting. Punishment after the fact cannot erase the institution's failure to explain expectations.

The first step in preventing misconduct is to decompose the vague phrase “use your own ability.”

• May AI assist with brainstorming?
• May it suggest an outline?
• May it rewrite one sentence?
• May it translate?
• Must outputs be cited?
• Must conversation history be submitted?

Without clear rules, a school does not teach ethics. It creates invisible traps that differ from one teacher to another.

7. Kant—why is cheating wrong beyond breaking a rule?

Kantian ethics reveals two problems in cheating.

First, it uses teachers and other students as means to the cheater's end. The student pretends to participate in a shared practice of demonstrating ability while taking only the resulting grade.

Second, it weakens autonomy. Autonomy does not require acting without help. It means accepting responsibility for the reasons and judgments one adopts.

A student who submits AI language they do not understand places words under their own name without owning the reasons. They appear to speak while hollowing out their position as an agent of judgment.

AI use itself is not necessarily heteronomous. A student can examine an output, correct errors, explain why parts were accepted or rejected, and reach a conclusion under their own responsibility.

The important question is not only who typed every character. It is who can answer for the reasons behind the claim.

8. Rawls—was homework already measuring the home?

AI-assisted misconduct threatens fairness. But homework was not perfectly fair before AI.

• Some students have quiet private rooms; others perform household work, caregiving, or paid employment.
• Some families can buy expensive tutoring; others cannot.
• Some parents can edit academic prose; others face constraints of language and time.
• Some homes have fast devices and connections; others have little access outside school.

Take-home assignments intended to measure independent ability have always measured household resources as well.

From a Rawlsian perspective, banning AI and returning to the old system is insufficient. Providing time and support inside school can reduce household inequality as well as AI cheating.

New inequalities may also emerge between paid and free AI systems, family expertise in prompting, and access to devices. If AI is included in an assignment, schools must provide comparable tools, time, and instruction to everyone.

Fairness is not only a world in which nobody uses AI. Whether AI is used or not, students need comparable access to the ability being assessed.

9. Foucault—may schools watch students constantly to protect learning?

Denmark's emergency measures include screen monitoring and network restrictions during exams. Some control is necessary to preserve the integrity of high-stakes assessment.

Michel Foucault's analysis of surveillance reveals another danger.

When every click, screen, gaze, and textual revision is recorded and students are permanently treated as suspicious, school changes from a place of learning into a place of monitoring. Students begin acting not to understand, but to leave a trace that appears innocent.

Surveillance needs at least four conditions:

1. Limit it to necessary settings, such as high-stakes examinations.
2. Explain what is recorded and who can view it.
3. Prohibit secondary use and long-term retention.
4. Accommodate disability and legitimate learning needs.

Assessment should not depend on surveillance alone. Oral defenses, mid-process consultations, short in-class checks, and student reflections can verify understanding without observing every moment of home study.

Trust in education is protected neither by zero verification nor by limitless monitoring. It requires a balance between verifiability and respect for persons.

10. Oral defense is not a universal solution

The idea that speaking reveals understanding is powerful, but oral defense has biases of its own.

• Students prone to anxiety may be disadvantaged.
• A student learning in a second language may understand but struggle to answer immediately.
• Speech, hearing, and neurodevelopmental differences require accommodation.
• Question difficulty and grading may vary among teachers.
• Large cohorts require substantial time and staff.

An oral defense must not become a competition won by the smoothest speaker.

Schools can share model questions and rubrics, give preparation time, permit written or visual responses and assistive technology, and improve consistency through recording or multiple assessors. Students should explain major choices and reasons, not memorize the entire submission.

The purpose is not competitive debate. It is to verify the connection between the artifact and the learner's understanding.

11. Three modes of assignment for the AI age

The same AI rule need not govern every task. We can distinguish three modes by purpose.

Mode A: protected foundational skill

Vocabulary, mental arithmetic, basic proof, and short composition may measure abilities that need to become internal and automatic. These tasks are completed briefly, at school, without AI.

The reason for the restriction should be explained—not “because this is how school has always worked,” but because the skill is needed when tools are unavailable and forms a foundation for higher judgment.

Mode B: disclosed AI collaboration

Students use AI, search, and source material as they might outside school. They disclose significant prompts and outputs, explain what they accepted or rejected, and identify sources used for verification. Judgment is assessed alongside the artifact.

Mode C: inquiry and defense

Long papers, research, and creative work may permit support tools while adding consultations, drafts, demonstrations, and oral defense. The question is whether the student can take responsibility for the inquiry and conclusion.

This three-part approach avoids both contradictions: banning AI only inside school even though society uses it, and assuming that the existence of AI removes the need to learn foundations.

12. Substitution, scaffolding, and co-production—the boundary of AI use

AI's role can be divided into three categories that make rules easier to explain.

Substitution

AI performs the central skill being assessed and the student submits the result without understanding it—for example, generating the entire argument in an argumentation task or obtaining only the answer in an arithmetic-skills task.

Scaffolding

AI provides hints, examples, feedback, and practice so the student can perform the task. The final judgment and explanation remain with the learner.

Co-production

The ability to collaborate with AI and other tools is itself assessed. Managing, verifying, editing, and allocating responsibility for outputs become part of the task.

The same action changes category with the purpose of the course. Machine translation may be substitution in a class assessing English composition. Evaluating and revising machine translation may be co-production in a class on international communication.

Ethical AI education is not a universal blacklist. It teaches students to identify the role their use plays relative to the learning objective.

13. Seven tests for “evidence of learning”

We can assess homework in the AI age through seven questions.

1. Purpose

Can the teacher state in one sentence what the task teaches and measures?

2. Correspondence

Does the quality of the submission genuinely correspond to the student's ability being assessed?

3. Defensibility

Can the student explain major choices, reasons, and responses to objections in their own words or through an appropriate alternative medium?

4. Process

Can some part of the learning process—drafts, mistakes, revisions, or consultation—be observed?

5. Transfer

Can the student apply the knowledge to a new problem when numbers, conditions, or examples change?

6. Fairness

Do household income, time, devices, or parental assistance unfairly determine the result?

7. Proportionality

Is anti-cheating surveillance proportionate to the stakes, while protecting privacy and reasonable accommodation?

The more these conditions are satisfied, the less assessment depends on guessing whether AI was used and the more it measures the learner.

14. A teacher's job is not to police prose

As AI cheating rises, teachers spend more time searching for unnatural language, interrogating suspected students, and checking surveillance records. Time for lesson preparation and feedback shrinks.

Misconduct cannot simply be ignored. But a system that turns teachers into AI detectors is not sustainable.

Assignment design can reduce conflict:

• Formulate the question during class.
• Replace a single large paper with brief drafts and consultations.
• Use local experience, experimental data, and in-class discussion as specific materials.
• Provide an AI-use disclosure field so students have less incentive to hide.
• Allocate grades to evidence, revision, and explanation rather than fluency alone.
• In low-stakes practice, experiment with AI as a learning partner.

These changes let teachers return from being detectives of misconduct to experts who observe and support changes in thought.

15. Should homework disappear?

Some homework should disappear.

Purposeless repetition, worksheets that reward volume, research tasks that merely copy search results, and giant home projects that measure parental support had doubtful educational value even before AI.

But reading, observing, practicing, and creating long work outside class can remain valuable. Nor can every kind of learning be completed during school hours.

Preserving the institution called homework should not become an end in itself.

Assignments worth keeping meet the following conditions:

• Teachers can explain why they must be done at home.
• The permitted role of AI is clear in relation to the objective.
• Alternatives exist for disadvantages in the home environment.
• High-stakes decisions do not rest on the finished artifact alone.
• Students receive concrete feedback.

Reducing homework is not the same as reducing learning. Time removed from low-value submission rituals can return to dialogue, practice, reading, and inquiry.

16. Conclusion—assessment moves from authorship to responsibility

Generative AI can produce polished prose faster than a human being. We might conclude that human writing no longer has a purpose.

But education was never only about supplying text.

It is about judging whether a claim is true, selecting evidence, answering objections, admitting error and revising, accepting the consequences of one's words, and transferring knowledge to unfamiliar problems.

These abilities are difficult to see in a finished product alone.

Assessment in the AI age should move from a search for pure authorship—“Who typed every character?”—toward another question:

Who understands, verifies, chooses, and can explain this claim, and who can accept responsibility for its consequences?

Oral defenses, supervised tasks, and records of process are ways of checking that responsibility. They must be used according to purpose, not as excuses for limitless surveillance.

Education did not become unnecessary when AI learned to answer homework.

Instead, AI made something unmistakable: possessing a correct answer and possessing the ability to think are not the same.

If homework is ending, that need not mean the end of learning. It may mean the end of an assessment system that treated an artifact as identical to the learner's ability.

FAQ

Is every use of AI for homework cheating?

No. It depends on the purpose and the stated rule. Permitted grammar checks or brainstorming are not misconduct. Generating an entire paper in a task measuring the student's own writing and submitting it as personal work may be.

Would a complete AI ban solve the problem?

AI-free conditions are necessary for some foundational skills. But home use is difficult to verify, and a total ban would not teach the ability to evaluate AI in the real world. It is better to separate AI-free foundational checks from disclosed AI collaboration.

Can an oral defense always expose AI cheating?

No. Answers can be memorized, and speaking ability varies. Drafts, in-class work, transfer problems, and oral explanation should be combined with accommodation and common rubrics.

Should students submit their entire AI conversation history?

Not always. Disclosure of significant use can be justified for high-stakes assignments, but full logs may contain private information. Schools should request only what the purpose requires and explain retention and access.

Does AI necessarily reduce thinking ability?

It depends on the use. Substituting an answer removes practice. Using AI for counterexamples, hints, questions, and feedback while verifying the output can stimulate thought.

Will reducing homework lower achievement?

Quantity alone cannot answer the question. Schools should remove purposeless work while strengthening in-class practice, reading, repetition, and feedback. Effects must be examined by age, subject, and objective.

NOW IN QUESTION

One student submits a perfect paper but cannot explain it. Another discloses AI use, corrects errors in the output, and can defend the conclusion.

In which student is the “student's own ability” more fully present?

References

• New emergency package against AI cheating in upper-secondary schools (Danish Ministry of Children and Education, August 6, 2026)
• Danish pupils will have to orally defend essays in attempt to combat AI cheating (The Guardian, August 6, 2026)
• Student Use of AI for Homework Rises as Concerns Grow (RAND, March 17, 2026)
• What's worth measuring? The future of assessment in the AI age (UNESCO, October 27, 2025)
• Use of AI in education: Deciding on the future we want (UNESCO)
• Philosophy of Education (Stanford Encyclopedia of Philosophy)
• Critical Thinking (Stanford Encyclopedia of Philosophy)
• Plato, Phaedrus.
• John Dewey, How We Think, 1910; Democracy and Education, 1916.
• Immanuel Kant, An Answer to the Question: What Is Enlightenment?, 1784.
• John Rawls, A Theory of Justice, 1971.
• Michel Foucault, Discipline and Punish, 1975.