Whitepaper

Assignments That Show the Learning

Designing homework, projects and tests that keep student thinking visible in the age of generative AI

By Frank van Bokhorst and Romana Finžgar

1. WHY DOQENT WROTE THIS PAPER

DOQENT builds a platform for teachers. In conversations with the teachers who use it, one question comes up more than any other: now that a student can produce a fluent essay, a tidy summary or a plausible lab report in seconds, how do I still know what my students can actually do?

This paper is our answer. It is deliberately practical. It does not argue about whether generative AI belongs in schools, because it is already there, in students' pockets and on their laptops, and increasingly in teachers' own preparation. It asks a narrower and more useful question: how do we design assignments so that the learning stays visible, whatever tools a student happens to use?

Our starting position is simple. An assignment is not "AI-resilient" because it forbids AI. It is AI-resilient because it cannot be completed well without the student thinking, choosing, gathering, explaining and revising in ways a teacher can see. Once that is true, the question of whether a chatbot was involved becomes far less important than the question the assignment was meant to answer in the first place: what has this student learned?

The name DOQENT comes from the Latin docēre, to teach, to show, to cause to know. Showing is the heart of it. Good assignments make learning show. That is what this paper is about.

2. WHAT GENERATIVE AI HAS CHANGED, AND WHAT IT HAS NOT

What has changed

For most of the history of schooling, producing a written product was expensive. A student who handed in a structured, error-free, well-sourced report had almost certainly done a lot of reading, selecting, ordering and drafting to get there. The product was a reasonable proxy for the process. Teachers could grade the product and infer the learning.

That proxy has broken. A general assignment with a clear expected structure and no specific context can now be turned into a finished product in under a minute. The product still looks like evidence of learning. It is no longer reliable evidence of anything except access to a chatbot and a willingness to use it.

Teachers notice this in familiar patterns: written work that is suddenly far stronger than classroom contributions; vocabulary and argumentation a student cannot explain when asked; homework that is consistently excellent alongside tests that are not; reflections that could have been written about any assignment at all.

What has not changed

Learning itself has not changed. Students still need to understand ideas, apply them to new situations, weigh alternatives, gather and judge evidence, explain their reasoning to others and improve their work in response to feedback. None of that can be outsourced without the student ending up with less than they started with.

Nor is the underlying problem new. Parents have always helped with homework. Older siblings have always written the occasional essay. Copying from encyclopaedias, then from websites, has always been possible. Schools adapted every time, usually by moving assessment closer to the classroom and closer to the student's own voice. Generative AI is a larger version of the same pressure, and it rewards the same response.

3. THE FOUR LENSES: POSITION, PLACE, PATH, PROOF

DOQENT uses four lenses to judge, and to strengthen, any assignment. Each lens asks one question. An assignment that answers all four convincingly is one where a student's thinking has to show, regardless of the tools used along the way.

The four lenses are not a scoring system. They are a way of looking. Hold an assignment up to each lens in turn and ask whether it passes. Most traditional assignments pass none of them: the topic is general, the structure is given, nothing is seen until the deadline, and nothing is said aloud. Most can be made to pass two or three with small changes, and that is usually enough.

A short illustration

Notice that the stronger version does not ban anything. A student may well ask a chatbot to explain a cause, or to check their grammar. What they cannot do is skip the thinking, because the thinking is what is being collected.

4. REDESIGNING AN ASSIGNMENT IN THIRTY MINUTES

You do not need to rewrite your course. Most existing assignments can be strengthened in half an hour by adding one or two elements. This section is the procedure DOQENT recommends.

1. Name the learning goal in one sentence. Not the topic, the goal. "Students can weigh competing explanations using evidence" is a goal. "The Industrial Revolution" is a topic. If you cannot state the goal, the assignment cannot be redesigned yet.
2. Hold it up to the four lenses. Write P, P, P, P on a sheet and tick the ones the current assignment already passes. Be honest. Most first attempts tick none.
3. Add Position first. It is the cheapest change. Turn "describe" into "choose and justify". Turn "list three" into "which of these three matters most, and why".
4. Add one anchor of Place. Something that exists only here: a class discussion, a measurement in the corridor, a conversation with a real adult, a photograph the student took. One anchor is enough.
5. Insert one visible checkpoint on the Path. A first paragraph written in class with devices closed. A list of sources handed in before the draft. A revision note explaining what changed after feedback.
6. Attach a short moment of Proof. Ninety seconds spoken. One question from a classmate. A hand-drawn diagram of the argument. It need not be graded heavily; it needs to exist.
7. Say where AI may and may not be used. Not as a threat, as an instruction. See section 10 for the model DOQENT recommends.

5. WORKED EXAMPLES ACROSS SUBJECTS

The lenses apply in every subject. The examples below are original DOQENT designs. Each pairs a common assignment with a strengthened version, and notes which lenses do the work.

Mathematics: probability, 13 to 14 year olds

Physics: motion and forces, 15 to 16 year olds

Biology: plant growth, 12 to 13 year olds

History: local heritage, 14 to 15 year olds

Language and literature: writing, 15 to 16 year olds

Chemistry: reaction rates, 15 to 16 year olds

Civics and economics: prices, 13 to 14 year olds

A pattern is visible across all seven examples. The stronger versions are not longer for the teacher to mark. They are often shorter. What they add is a small amount of observed reality, one or two visible stages, and a spoken moment. Those three ingredients do most of the work.

6. HOMEWORK THAT STILL TELLS YOU SOMETHING

Homework has always been the least controllable form of school work. Help from home was already invisible. Generative AI makes the problem more acute but does not change its nature. The answer is not more surveillance; it is homework that is designed to feed the next lesson rather than to replace it.

Three principles for homework design

Homework prepares; the lesson collects

If the product of homework is something that happens in the classroom, the incentive to outsource it collapses. Ask students to prepare a question, a position, a set of observations or a first attempt, and then use the lesson to work with what they bring. A student who arrives with nothing of their own has nothing to contribute, and both they and you will know it immediately.

Ask for the selection, not the set

"Complete the exercises" invites a chatbot. "Complete the exercises and tell me which two you would set as test questions, and why" invites judgement. The selection is short to write, quick to read and impossible to fake convincingly.

Anchor in the student's own material

Ask students to connect the content to something only they have: their sport, their commute, their family, their last week. The link need not be deep. Its purpose is to make the homework belong to the student rather than to the topic.

Examples

7. RESEARCH PROJECTS AND EXTENDED WORK

The traditional research project (choose a topic, gather information, write a report) is the format most exposed to generative AI. It is also one of the most valuable formats in a school, because it is where students learn to sustain an inquiry. The right response is not to abandon it but to redesign it so that the inquiry, not the report, is what gets assessed.

Five design moves

Require first-hand data

A project that must include something the student collected themselves cannot be generated. Interviews, observations, small surveys, measurements, site visits and experiments all qualify. Even one such element changes the character of the whole project, because everything else must be reconciled with it.

Narrow the perspective

"Sustainability" is a topic. "How well does our school building manage heat in June, based on measurements in three classrooms" is a question with a place, a method and a possible answer. Narrow questions are harder to outsource and, in our experience, more interesting to students.

Stage the path

Collect the question, the source list, an early draft and a revision note as separate deliverables with separate dates. Each one is small. Together they make the student's route visible and make the final report almost impossible to fake, because it has to be consistent with everything that came before it.

Compare kinds of evidence

Ask students to set their own observations against a published source and to comment on where they agree and where they do not. This is genuine academic work, it is very hard to simulate, and it produces some of the best thinking you will see from students.

End with a moment of proof

After the report is submitted, give each student five minutes and a blank sheet to draw the structure of their argument, then two minutes to talk it through. Students who did the work find this easy and often enjoyable. Students who did not find it impossible. The teacher learns more in those seven minutes than from many pages of text.

A staged project template

8. TESTS AND IN-CLASS ASSESSMENT

Tests are the format least affected by AI, because they happen under supervision. But a test that only asks students to reproduce material they could have generated at home is still a weak instrument. The same four lenses improve tests as much as they improve homework.

Test formats that make thinking visible

• Questions tied to lesson-specific material. "Using the example we worked through on Tuesday, explain why the method failed at step three." This material does not exist outside the classroom.
• Open-notes questions that require judgement. Allow class notes, then ask questions that notes alone cannot answer: compare, rank, predict, decide.
• Show-your-route questions. Award part of the mark for the reasoning path rather than the final answer, and say so on the paper.
• Own-example questions. "Give one example from your own experience of this principle, and explain the link."
• Case questions. A short unfamiliar scenario the student must analyse with the tools of the course.
• Defence questions. "Here is a claim. Do you agree? Give one reason for and one against, then decide."
• Paired written and spoken assessment. A short written test followed by a two-minute conversation with the teacher about one answer, chosen by the teacher on the spot.

Spoken proof and students with support needs

Proof moments are meant to reveal understanding, not fluency, confidence or first-language ability. A student with a formal support plan, a language background different from the language of instruction, a speech difficulty or acute anxiety in front of a group is entitled to demonstrate understanding in a form that works for them, and schools are legally expected to provide it. The Proof lens is satisfied by any live, own-words demonstration that cannot be prepared in advance and pasted in.

• Change the audience, not the task. One-to-one with the teacher, or with a single trusted partner, instead of in front of the class.
• Change the medium, not the task. A hand-drawn diagram, a labelled sketch, a short written explanation completed in class without devices, or a demonstration with objects.
• Allow preparation notes in the student's home language, with the explanation itself given in the language of instruction where that is the learning goal, or in the home language where the goal is the content.
• Give time and warning where the support plan specifies it, and accept that ninety seconds may become five minutes.
• Follow the plan. Where a student has a documented support plan, its provisions apply to Proof moments exactly as they apply to any other assessment. The teacher does not have to invent an adaptation; the plan already describes it.

These adaptations are not a weaker form of proof. A student who explains the collision model one-to-one with a sketch has shown exactly what the lens asks for.

The purpose of these formats is not to exclude AI. It is to make sure the test measures what the course was for: understanding, application, judgement and explanation. A test designed this way is also fairer, because it rewards students who understood the material over students who merely memorised it.

9. ASSESSING THE PROCESS: THE DOQENT PROCESS RUBRIC

Once an assignment makes the process visible, the process can be assessed. This is fairer than product-only grading, because it rewards students who understood the work but produced a modest final piece, and it tells students what actually matters. DOQENT recommends a four-criterion rubric with the final product carrying no more than a third of the weight.

Performance levels

The rubric is short by design. Students should be able to hold it in their heads, and teachers should be able to apply it in minutes. Share it before the assignment begins. Students who know that their choices, their evidence and their revisions carry seventy percent of the mark behave very differently from students who believe only the product counts.

10. AGREEMENTS ABOUT AI: CLASSROOM AND SCHOOL LEVEL

A ban on AI does not work, for the same reason a ban on parental help never worked: it cannot be enforced, it pushes use underground and it forfeits the chance to teach responsible use. What works is a clear agreement, at classroom and school level, about where AI is welcome and where the student's own thinking is required.

The three-zone model

DOQENT recommends dividing every assignment into three zones and telling students explicitly which parts fall where. The model is deliberately simple, so that it can be printed on the assignment sheet itself.

The calculator analogy is worth using with students. In a mathematics test, nobody is confused by the instruction "calculator permitted for questions 4 to 8 only". The three-zone model is the same instruction for a different tool.

One line across the school

Agreements are far more effective when the whole school holds them. A student who is allowed everything in one classroom and forbidden everything in the next learns only that the rules are arbitrary. A school does not need identical assignments in every subject, but it does need a shared baseline: the same three zones, the same expectation that use is declared, and the same understanding that spoken and in-class work is always the student's own.

A shared baseline also helps teachers. It turns a hundred individual judgement calls into one team decision, and it makes conversations with students and parents about AI use short and unemotional, because everyone is referring to the same document.

11. TEN PITFALLS WE SEE IN PRACTICE

These are the mistakes DOQENT sees most often when schools first respond to generative AI. Each is understandable. Each has a better alternative.

12. THE DESIGN CHECKLIST

Use this before setting any assignment that will be graded. Ten questions, two minutes. Aim for at least six yes answers. If the last two are not, the assignment needs work regardless of the rest.

13. PROMPTS TEACHERS CAN USE

Generative AI is a capable assistant for the teacher who is designing assignments, provided the teacher stays in charge. The prompts below are DOQENT originals built around the four lenses and the three-zone model. Fill in the brackets, read the output critically, and keep what fits your class. You know your students, your subject and your goals. The prompt does not.

14. HOW THE DOQENT PLATFORM SUPPORTS THIS APPROACH

Everything in this paper can be done with paper, a pencil and a timetable. DOQENT exists to make it lighter to do consistently, across a full teaching load, without losing the pedagogy described above. This section describes where the platform fits. It is not a feature list; it is a description of the workflow.

Lesson preparation with the lenses built in

When a teacher prepares a lesson or assignment in DOQENT, the platform works from the curriculum the teacher actually teaches, aligned to the relevant national framework. Assignments can be drafted with the four lenses and the three-zone AI agreement as part of the structure, so that Position, Place, Path and Proof are considered at the point of design rather than remembered afterwards. Differentiated versions for different levels within the same class are produced from the same design, keeping the pedagogical intent identical while adjusting the demand.

Assessment generation that favours application

DOQENT's assessment generation is built to produce the kinds of questions described in section 8: scenario-based, own-example, show-your-route and defence questions, tied to the material a teacher has actually covered. Because the platform holds the teacher's own lesson content, it can reference examples that were used in class, which is exactly the lesson-specific material that generic question banks cannot provide.

AI-assisted grading against a process rubric

The DOQENT process rubric in section 9, or a teacher's own variant, can be applied through the platform's AI-assisted grading. The teacher remains the grader; the platform prepares a structured first reading against each criterion, flags where evidence for a criterion appears to be missing, and drafts feedback in the teacher's own language and register. This is where the staged deliverables of a project earn their keep: the platform can hold the question, the source log, the draft and the revision note together, so that consistency between stages is easy to see.

Keeping the whole route in one place

Intermediate deliverables are only useful if they are easy to collect and easy to find again. Within DOQENT, the stages of an assignment sit together in the student's record for that subject, so that a teacher reviewing a final report can see the path that led to it without searching through email or a shared drive.

Communication with parents

The three-zone agreement, the process rubric and the expectations for first-hand work are only effective if parents understand them. DOQENT's school-to-home communication carries the assignment brief, the zones and the assessment weighting to parents in the language of their choice, so that help at home works with the assignment rather than against it.

Languages and access

DOQENT operates in Slovenian, English, German, Dutch, Croatian and French. The platform is free for teachers and for students. It is designed to be used by a school, not by an individual working around a school, which is why the shared agreements described in section 10 sit naturally within it.

15. CLOSING NOTE

Generative AI has changed what a written product can tell a teacher. It has not changed what learning is. Students still need to think, choose, gather, explain, defend and improve, and the job of an assignment is still to make that happen and to make it visible.

The good news is that the changes required are small. A choice to defend. One thing observed in the real world. One stage seen before the deadline. Ninety seconds spoken aloud. A clear line about where AI is welcome and where it is not. None of this requires a new theory or new budget. Most of it requires less marking, not more, because it replaces pages of generic text with short, specific evidence of thinking.

There is also something to be gained beyond resilience. Students who have to choose, gather, explain and revise develop ownership of their work in a way that reproduction never produces. They finish with something they made, not something they obtained. The teacher who designs for that is not defending against technology. They are doing what good teachers have always done, with slightly sharper tools.

The role of the teacher in all of this is not diminished. It is the centre of it. At DOQENT we call the teacher the heartbeat of the school. Everything we build exists to help teachers, students and parents work together, so that the learning that happens is real, visible and the student's own.

DOQENT, August 2026

APPENDIX A. ONE-PAGE SUMMARY FOR THE STAFF ROOM

THE FOUR LENSES

THE THREE AI ZONES

THE PROCESS RUBRIC

THE THIRTY-MINUTE REDESIGN

Name the goal. Check the four lenses. Add Position. Add one anchor of Place. Insert one Path checkpoint. Attach ninety seconds of Proof. Print the three zones on the sheet.

TWO RULES THAT SIT ABOVE EVERYTHING

Detector results are a reason for a conversation, never evidence. Student data never goes into public AI services.

DOQENT · doqent.ai · From first bell to final grade.