September 25, 2026

A detector score should never decide whether a student cheated

Schools are trying to protect learning while AI becomes part of ordinary life. Recent moves in California, Denmark and Norway point toward clearer rules, better evidence and age-appropriate use.

By Jay Cadmus

The short version

A student turns in an essay. Software says AI probably wrote it. The student says the work is original. What happens next can affect a grade, a disciplinary record and the student's trust in school.

California has now put a useful boundary around that decision. Its new, nonmandatory model policy for K-12 schools says AI detection software should never be the sole basis for discipline or a grade penalty. A school should look at the student's engagement, use another way to assess the learning and review documented evidence of authorship. Students and families should also know how to challenge an accusation.

That protects innocent students without giving actual cheating a pass. Submitting generated work as your own still violates academic integrity. Schools need evidence strong enough to tell those situations apart.

Other countries are changing the work itself. Denmark announced last week that students ages 16 to 19 will have to defend major essays orally. Schools will also bring more writing into controlled settings so teachers can see how the work develops. Norway is taking an age-based approach: children ages 6 to 13 generally won't use generative AI in school, students ages 14 to 16 may use it cautiously with teachers, and older students will learn appropriate use for work and further education.

These decisions share a practical goal. A school should be able to see what a student understands, protect a child from unsafe systems and prepare older students for a world where AI is common. Suspicion alone can't do that. Neither can unlimited access.

A detector score needs evidence

AI detectors estimate whether patterns in a piece of writing resemble machine-generated text. They don't watch the work being created. They can't know whether a student drafted a paragraph over three days, asked a parent to edit it or used a chatbot to produce the whole thing.

Research has also found uneven performance. A Stanford-led study tested several widely used detectors on writing by native and non-native English speakers. The tools frequently misclassified the non-native writers, while simple changes to AI prompts could help generated text avoid detection. That creates a bad bargain: some honest students face more scrutiny while a student who knows how to revise generated text may pass unnoticed.

The harm begins before a formal punishment. A student accused of cheating may feel that their teacher no longer recognizes their voice. A multilingual student may start making sentences more complicated to appear human. A careful writer may save every keystroke out of fear. Those responses consume attention that should be going toward learning.

California's model policy offers a fairer process. A detector can raise a concern. The teacher then looks for evidence such as drafts, notes, source choices, revision history and the student's ability to explain the work. An alternative assessment can show whether the student understands the material. The student gets a clear way to respond.

This process also catches cases that a detector misses. A student who submits polished AI output but can't explain the argument, sources or calculations has revealed a learning gap. The student's demonstrated work becomes the focus, while a detector percentage remains one weak input.

Make the learning visible

Denmark's oral-defense requirement changes what a major essay proves. About 9,000 upper-secondary students complete the affected assignment each year. They will now have to discuss and defend work written at home. Schools are also being encouraged to complete more assignments in class and follow the writing process.

An oral defense won't fit every student or subject. Anxiety, disability and language differences can affect how someone performs in a live conversation. Schools need other ways to show understanding, including annotated drafts, short conferences, worked examples and recorded explanations.

The underlying idea is strong. A finished document has become easier to generate. The path to that document still reveals a lot. Why did the student choose this source? What changed after feedback? Which claim is weak? Can the student apply the same idea to a new example?

Teachers can make that path visible without turning every classroom into a surveillance room. A research assignment might include a topic proposal, source check, early outline, draft and brief reflection. A math or science task can ask a student to explain one incorrect approach before showing the correct one. A coding project can include a short demonstration and questions about the student's choices.

Each assignment should also state the AI rule in plain language. Some work should be completed without AI because the student is practicing a foundational skill. Some may allow AI for brainstorming, feedback or translation while keeping the final reasoning with the student. Other work may require students to use AI, test its output and disclose what they changed. A schoolwide slogan can't make those distinctions for every lesson.

Different ages need different rules

Norway's new standards draw a sharp line by age. Beginning with the new school year, students ages 6 to 13 generally won't use generative AI in school. Students ages 14 to 16 may use it cautiously under teacher supervision. Students ages 17 to 19 will be taught appropriate use as preparation for higher education and work.

The reasoning is easy to understand. Young children still need to build fluency in reading, writing and mathematics. A tool that produces an answer too early can remove the productive struggle where those skills develop. Older students need practice making decisions about AI because colleges and employers will expect it.

Age changes the safety requirements too. A child may share a full name, school, medical concern or family problem without understanding where that information goes. A conversational system may sound certain, friendly or caring even when it is wrong. A student can encounter harmful content, biased advice or an interaction designed to keep them engaged.

UNICEF's latest guidance calls for child-centered AI that protects privacy, safety, fairness, development and well-being while preparing children for current and future uses. In a school, that starts with approved accounts and tools, limited data collection, clear guardian information and an adult route for reporting something unsafe. Children shouldn't have to understand a vendor's business model before receiving protection.

Responsible use also needs to be taught. A student should know that a fluent answer can contain a false claim, personal information should stay out of open chatbots, generated images can harm real people and AI help must be disclosed when an assignment requires it. These are practical habits. They belong beside research skills, media literacy and online safety.

Trust has to work both ways

Many school conversations begin with what students might hide. California's policy also asks educators to disclose their own AI use. That could include AI-assisted feedback, a generated example or material drafted for a lesson.

The disclosure matters because students can see the standard being applied fairly. A teacher can say where AI saved time, where the output was wrong and what changed before the material reached the class. Students can then describe their own use without every admission becoming a confession.

Clear rules still need consequences. A student who submits generated work as original has avoided the practice the assignment was designed to provide. The response should connect back to that missing learning. Requiring the student to redo the work, explain the topic and reflect on the choice may produce more value than a punishment that ends the conversation.

Schools also need to separate confusion from deception. One teacher may permit brainstorming while another prohibits it. A writing aid may add generative features during an update. A student may use an approved accessibility tool without realizing that part of its output counts as AI assistance. Expectations should appear on the assignment, not remain buried in a district policy.

AI can support learning without hiding it

Used carefully, AI can give a student another way to get unstuck. A student can ask for three counterarguments to a draft, compare them with credible sources and decide which one deserves a response. A language learner can practice a conversation, then review the wording with a teacher. A student preparing for an oral defense can ask a chatbot to challenge the argument and reveal weak spots.

Teachers can use approved tools to draft practice questions, adjust the reading level of a passage or create examples for different interests. The teacher still checks accuracy, bias and fit. Student records and identifiable work stay out of consumer accounts.

The measure is the student's growth. Did the student understand the material, improve the work and explain the final decision? Speed alone is a weak result when the tool completes the thinking that school was meant to develop.

Opportunity radar

Student-data reviews for school AI tools

Schools are approving tutoring, writing, translation and classroom-management tools that may collect student work, conversations and identifiers. Many districts lack staff who can review every vendor's retention terms, training practices, age rules and deletion process.

A privacy or education-technology specialist could review one district's highest-use AI tools and produce a plain-language record of what each service collects, who can see it and how the data can be removed. The buyer would be a district, charter network or regional education agency. A pilot should verify vendor claims against contracts and product settings, then measure how many risky tools are changed, restricted or replaced. The useful result is fewer risky tools in use and a clear deletion path.

What you can do with this

If you teach

Put the AI rule on each assignment. Ask for one or two signs of process that fit the work, such as a source note, draft or short explanation. Treat a detector result as a reason to look closer and keep the final decision with people.

If you're a student or parent

Keep ordinary evidence of the work, including notes, sources and meaningful drafts. Ask which AI uses are allowed before starting. If an accusation arises, request the evidence, the review process and another way to demonstrate understanding.

If you lead a school

Give teachers approved tools, privacy rules and a fair authorship-review process. Train staff before enforcement begins. Track who gets accused, which evidence changes the decision and whether multilingual students or students with disabilities face uneven outcomes.

The bigger picture

Schools have two jobs here. They must protect the work of learning, and they must prepare young people to use a technology they will encounter outside the classroom. Those jobs require different limits for a second grader, a middle-school student and a senior preparing for college or work.

Fair evidence, visible learning and clear rules can carry most of the load. Students should know when AI help is allowed, teachers should know how to verify understanding and families should know how a decision can be challenged. A school earns trust when it can address cheating without treating every unusual sentence as proof.

References

• California Department of Education: Model policy for artificial intelligence in K-12 schools, reviewed June 26, 2026
• The Guardian: Denmark requires oral defenses and more supervised writing in upper-secondary schools, August 6, 2026
• Reuters: Norway sets age-based limits on generative AI in schools, June 19, 2026
• UNICEF: Guidance on AI and children, version 3.0, December 2025
• Patterns research: AI detectors show bias against non-native English writers