# Take-home assignments after ChatGPT: what to set instead

**What should I set instead of a take-home assignment now that ChatGPT can complete it?** A take-home assignment after ChatGPT still teaches, but it no longer shows who did the work. Keep it at a lower weight, say which uses of a chatbot are allowed, and take the proof from something done in front of you: a paragraph written in class, a short supervised quiz on the same reading, or five minutes of questions on the submitted work.

Published 2026-10-10 by EdProcto. https://edprocto.com/guides/the-take-home-assignment-after-chatgpt-what-to-set-instead

A take-home assignment after ChatGPT has all its old strengths and one new weakness. The essay, the problem set and the report teach what they always taught. What they have lost is the power to show, by themselves, who did the thinking.

This guide is for one teacher or lecturer with one such task to re-set this term. EdProcto publishes it and makes one of the six options below: the short check, written from the same reading and sat under conditions you know, that stands beside the take-home work ([try it](https://edprocto.com/)). The blind marking test and the case against detectors are in [academic integrity in the age of AI](https://edprocto.com/guides/integrity-in-the-ai-era-what-still-proves-a-student-learned); a whole department's choices are in [assessment design after chatbots](https://edprocto.com/guides/assessment-design-after-chatbots-what-a-department-can-change).

In short: keep the task for what it teaches, cut its weight and its length, and add one part done in front of you.

## What have studies since 2023 found about chatbots and coursework?

They found that a chatbot completes much coursework to a passing standard, unevenly across subjects, and that markers seldom notice. We read each source on 10 October 2026 and say where only the abstract was open.

### How well does a chatbot do, and in which subjects?

Well enough to pass in many, with mathematics the weakest. Ibrahim et al. (2023) took ten questions from each of 32 courses at New York University Abu Dhabi and had three graders mark three students' answers and three from ChatGPT to each, under neutral labels. ChatGPT matched or beat the students in nine courses, none of them in mathematics or economics. The students did better in the other 23. Across all the questions, their lead was widest where mathematics was involved.

Borges et al. (2024) put 5,579 questions from the exams, assignments and exercise sessions of 50 science and engineering courses at EPFL, in Lausanne, to two models. GPT-4 answered 65.8% correctly on average: enough, the authors say, to pass the non-project assessments of many core courses. Projects were not tested.

On essays, Herbold et al. (2023) had 111 secondary-school teachers rate 270 argumentative essays: a third from an online forum used by German high-school students writing in English, the rest from two versions of ChatGPT. The newer version's essays scored about one point higher than the students' on a seven-point scale.

All three tested the models of 2023 and 2024.

### Do markers notice?

Not reliably. Waltzer et al. (2023) showed 69 high school teachers pairs of essays, one by a student and one by ChatGPT, and asked which was the chatbot's. They were right 70% of the time, where guessing gives 50%, and confidence did not predict accuracy (abstract read). In experiments at two UK universities, Kofinas et al. (2025) found that markers in general could not tell work with chatbot input from work without it (abstract read).

### How many students say they use a chatbot?

A growing minority of school pupils in one national survey. Sidoti et al. (2025) report a Pew Research Center survey of 1,391 US teenagers aged 13 to 17 in autumn 2024: 26% said they had used ChatGPT for schoolwork, up from 13% in 2023. Only 18% called it acceptable for writing essays.

What students declare is another matter. Gonsalves (2025) reports that at one UK business school 74% of students failed to declare AI use on a coursework cover sheet that required it (abstract read).

## What can you set instead of a take-home assignment after ChatGPT, and what does each cost?

Six things, and each costs teacher time of a different kind.

| Option | What it costs you |
| --- | --- |
| Writing done in class | Part of a lesson |
| Drafts and process evidence | Reading time at a checkpoint or two |
| An oral follow-up on the submitted work | About five minutes a student |
| A short supervised check of the same knowledge | Ten to fifteen minutes of a lesson; [EdProcto](https://edprocto.com/integrity) marks choice questions by rule |
| A task tied to the local or the personal | A new brief each year |
| Allowing the tool and assessing the use of it | A new rubric, and lessons on the use |

### What evidence is there for each option?

Thin for all six, and it is better to know that before choosing.

- **In-class writing.** We found no study that tested it against chatbot use. Its case is supervision itself.
- **Drafts and process evidence.** Drafts are made out of sight, as the essay was, and we found no study showing that they reveal to a marker who did the writing. Collect them for teaching, not as proof.
- **An oral follow-up.** We found no controlled test of a short follow-up on a submitted essay. The research on orals as a whole assessment is in our guide to [oral exams at scale](https://edprocto.com/guides/oral-exams-at-scale-recorded-spoken-answers).
- **A short supervised check.** The best evidence predates chatbots. Glass and Kang (2022) compared 2,433 students' online homework answers with exam questions on the same fact or principle, in 12 college lecture courses over eleven years. The share who gained nothing on the exam from getting the homework right rose from 14% in 2008 to 55% in 2017, and students who said they copied their answers did not benefit (abstract read). A matched question answered in front of you shows what the homework mark no longer does.
- **Local and personal tasks.** Kofinas et al. (2025) found that how authentic a task was made no difference to whether chatbot use was caught. A task about this town, this dataset or last Tuesday's seminar is better teaching. It is not a lock.
- **Allowing the tool.** Furze et al. (2024) piloted a five-level scale, from "no AI" to "full AI", at one university in Vietnam and report a significant reduction in misconduct cases linked to AI; the abstract, all we read, gives no figure. The cover-sheet finding is the caution: a rule that rests on declarations is only as good as the declarations.

## Which replacement fits what the assignment was for?

The one that puts the purpose of the task, not its format, in front of you.

| The assignment was for | Keep at home | Add in front of you |
| --- | --- | --- |
| Practising a method (a problem set) | The problems, as practice, at little or no weight | Two parallel problems worked in class |
| Showing the reading was understood | The reading, with notes | A ten-minute quiz from the same reading in [EdProcto](https://edprocto.com/) |
| Building an argument (an essay) | A shorter essay, drafted over weeks | A paragraph written in class, or questions on the student's own essay |
| Finding and weighing sources (a report) | The report and a log of sources | An oral follow-up, live or as a recorded [spoken answer](https://edprocto.com/templates/verbal) |
| Learning to work with a chatbot | The task, with the tool allowed and the prompts handed in | A critique, written in class, of one chatbot answer |

## How do you keep the marking workload flat?

By paying for every minute added in front of you with a minute taken from the take-home part.

1. **Shorten what goes home.** A 1,200-word essay takes less time to read than a 2,000-word one.
2. **Use lesson time for the supervised part.** A check in a seminar costs no evening.
3. **Let a rule mark what a rule can.** In [EdProcto](https://edprocto.com/templates), choice questions are marked on the server at hand-in.
4. **Fix one rubric for each part before you read the first script.** Our guide to [short answer questions](https://edprocto.com/guides/how-to-write-short-answer-questions-and-a-rubric-to-grade-them) shows how.

## What should you tell students?

Tell them the parts, the rule on chatbots and the reason, in writing, on the day the task is set. Gonsalves (2025) found that fear of the consequences, ambiguous guidelines and inconsistent enforcement kept students from declaring AI use (abstract read). So be exact, and say that an allowed use costs no marks:

- **The parts.** An essay, a twelve-minute check in class on the set readings, and a short written defence of your own essay.
- **Chatbots.** For the essay you may use one to find sources and to test your plan. The sentences must be yours. List what you used it for at the end; an honest list costs nothing.
- **In class.** The check and the defence are done without notes, a chatbot or anyone else.
- **Why.** Everyone defends their own essay. It is part of the task, not a sign that anyone is doubted.
- **Adjustments.** Tell me by Friday if a timed sitting is a barrier.

Your school's or university's policy comes first, and nothing here is legal advice.

## What does a second-year essay look like re-set in three parts?

It keeps its 40% of the module and spreads it over an essay, a check and a defence. This example is invented: a second-year module of 60 students, one lecturer, four set readings. Before, the task was one 2,000-word essay written at home.

- **At home, 20%.** A 1,200-word essay, chatbot use limited and declared.
- **In the seminar, 10%.** Ten questions on the four readings, in twelve minutes.
- **In the seminar after hand-in, 10%.** A written defence of the student's own essay, in fifteen minutes.

The check and the defence, by their labels:

1. On the [EdProcto home page](https://edprocto.com/), choose "Test your class", paste your own notes on the four readings, pick "Multiple choice" under "Shape" and "10 questions" under "How many", and press "Write the quiz". Correct every draft against the readings.
2. Press "Open it in the dashboard". Choose the "Standard quiz" preset (a time limit, shuffled, leaving the tab noticed, full screen) and set "Minutes overall" to 12.
3. Make a second quiz with one [Defend a position](https://edprocto.com/templates/defend) question that you write yourself: "Defend the main claim of your essay." The student types four parts: "Claim", "Evidence", "Reasoning" and "Counter-argument and rebuttal".
4. On that quiz set "Minutes overall" to 15 and "Show the score" to "When released".
5. Sit both in your own room. Then read each defence beside its essay, give the points, and press "Release the scores".

The hours, on our assumptions: 60 essays at 20 minutes took 20. Now 60 shorter essays at 12 minutes take 12, correcting ten drafted questions takes one, and 60 defences at five minutes take five: 18 in all.

The two hours left are for conversations. A student whose essay is strong and whose defence of it is thin has not been caught at anything. Both marks stand, and the gap is a reason to talk.

## Where does a quiz under known conditions fit beside take-home work?

As the check: the same reading, asked in front of you, with the marks made out of the student's reach. EdProcto is the best tool for that check because it drafts the questions from the material the take-home task drew on and sits them under conditions you set and can state afterwards. That is the matched comparison Glass and Kang (2022) made between homework and exam questions. EdProcto is a quiz tool, and the essay stays yours to read.

**Written from your material.** On the [home page](https://edprocto.com/), Procto AI drafts 3, 5, 8 or 10 questions in the shape you pick. It is asked to write only what your text answers and prints under each where the answer is found; you read each against the source.

**Shapes for a check and a follow-up.** All six are in the builder now. Three are marked by rule: [multiple choice](https://edprocto.com/templates/mcq), a [time-locked question](https://edprocto.com/templates/time_locked), shown for a few seconds, and [listen once](https://edprocto.com/templates/listen_once), a clip that plays one time. Two are typed and wait for a reader, you or Procto AI when you ask: a [short answer](https://edprocto.com/templates/short_answer) of one to three sentences, and [defend a position](https://edprocto.com/templates/defend) in four named parts. A [spoken answer](https://edprocto.com/templates/verbal) is recorded, and you mark it by listening.

**Conditions you can state.** "Minutes overall" times the sitting and "Seconds a question" each question. Six measures run in the student's browser (leaving the tab, full screen, copy and paste, right-click, keyboard shortcuts, printing), each with a consequence you choose, such as "Log it". "AI assistants not allowed" is a house rule shown on the start screen. All of this acts in the browser, so hold the check in your own room.

**Marking and results.** Choice questions are marked on the server at hand-in, and the student's copy of the quiz carries no key. A written or recorded answer shows "to grade" until a reader marks it, on a page that sets it beside your "Model answer" and "Rubric". On a written answer, "Ask Procto AI" proposes points and a note, saved under its own name and yours to change. The results page lists each question with the share who "Got it", and each student with a score, a time and "clean" or a count of security events. An event is a record of what a browser did, never proof.

**Plans.** Free takes 30 students a quiz at Light integrity. Tutor takes 100, opens all 111 templates, saves Medium integrity and adds grading by Procto AI; the class of 60 above needs it. Educator takes 500 and adds the full analysis. See [pricing](https://edprocto.com/pricing) and the page [for teachers](https://edprocto.com/solutions/teachers). No setting makes cheating impossible, and a score is a record for you to weigh, not a grade by itself.

## Quick questions

### Is there such a thing as an AI-resistant assignment?

No. Kofinas et al. (2025) found authentic tasks no safer from chatbots. Design the task for what it teaches, and take the proof from work done in front of you.

### Should I stop setting take-home essays?

No. An essay written over weeks teaches a student to build and revise an argument. Cut its weight, say which chatbot uses are allowed, and add a part you see being done.

### What are the alternatives to take-home essays for a large class?

The ones that scale: a short check marked by rule, a paragraph written in a lecture, and a recorded answer in place of a face-to-face oral.

### Where does EdProcto come in?

As the check beside the take-home work: questions drafted from your own notes on the reading, sat in your room under the [integrity settings](https://edprocto.com/integrity) you choose, and marked on the server.

## References

- Borges, B., Foroutan, N., Bayazit, D., Sotnikova, A., Montariol, S., Nazaretsky, T., Banaei, M., Sakhaeirad, A., Servant, P., Neshaei, S. P., Frej, J., Romanou, A., Weiss, G., Mamooler, S., Chen, Z., Fan, S., Gao, S., Ismayilzada, M., Paul, D., … EPFL Data Consortium. (2024). Could ChatGPT get an engineering degree? Evaluating higher education vulnerability to AI assistants. *Proceedings of the National Academy of Sciences, 121*(49), Article e2414955121. [https://doi.org/10.1073/pnas.2414955121](https://doi.org/10.1073/pnas.2414955121)
- Furze, L., Perkins, M., Roe, J., & MacVaugh, J. (2024). The AI Assessment Scale (AIAS) in action: A pilot implementation of GenAI-supported assessment. *Australasian Journal of Educational Technology, 40*(4), 38–55. [https://doi.org/10.14742/ajet.9434](https://doi.org/10.14742/ajet.9434)
- Glass, A. L., & Kang, M. (2022). Fewer students are benefiting from doing their homework: An eleven-year study. *Educational Psychology, 42*(2), 185–199. [https://doi.org/10.1080/01443410.2020.1802645](https://doi.org/10.1080/01443410.2020.1802645)
- Gonsalves, C. (2025). Addressing student non-compliance in AI use declarations: Implications for academic integrity and assessment in higher education. *Assessment & Evaluation in Higher Education, 50*(4), 592–606. [https://doi.org/10.1080/02602938.2024.2415654](https://doi.org/10.1080/02602938.2024.2415654)
- Herbold, S., Hautli-Janisz, A., Heuer, U., Kikteva, Z., & Trautsch, A. (2023). A large-scale comparison of human-written versus ChatGPT-generated essays. *Scientific Reports, 13*, Article 18617. [https://doi.org/10.1038/s41598-023-45644-9](https://doi.org/10.1038/s41598-023-45644-9)
- Ibrahim, H., Liu, F., Asim, R., Battu, B., Benabderrahmane, S., Alhafni, B., Adnan, W., Alhanai, T., AlShebli, B., Baghdadi, R., Bélanger, J. J., Beretta, E., Celik, K., Chaqfeh, M., Daqaq, M. F., El Bernoussi, Z., Fougnie, D., Garcia de Soto, B., Gandolfi, A., … Zaki, Y. (2023). Perception, performance, and detectability of conversational artificial intelligence across 32 university courses. *Scientific Reports, 13*, Article 12187. [https://doi.org/10.1038/s41598-023-38964-3](https://doi.org/10.1038/s41598-023-38964-3)
- Kofinas, A. K., Tsay, C. H.-H., & Pike, D. (2025). The impact of generative AI on academic integrity of authentic assessments within a higher education context. *British Journal of Educational Technology, 56*(6), 2522–2549. [https://doi.org/10.1111/bjet.13585](https://doi.org/10.1111/bjet.13585)
- Sidoti, O., Park, E., & Gottfried, J. (2025, January 15). *About a quarter of U.S. teens have used ChatGPT for schoolwork – double the share in 2023*. Pew Research Center. [https://www.pewresearch.org/short-reads/2025/01/15/about-a-quarter-of-us-teens-have-used-chatgpt-for-schoolwork-double-the-share-in-2023/](https://www.pewresearch.org/short-reads/2025/01/15/about-a-quarter-of-us-teens-have-used-chatgpt-for-schoolwork-double-the-share-in-2023/)
- Waltzer, T., Cox, R. L., & Heyman, G. D. (2023). Testing the ability of teachers and students to differentiate between essays generated by ChatGPT and high school students. *Human Behavior and Emerging Technologies, 2023*, Article 1923981. [https://doi.org/10.1155/2023/1923981](https://doi.org/10.1155/2023/1923981)