# Assessment design for AI: assignments with checkable sources

**How do I design assignments for the age of AI that ask for sources I can check?** Assessment design for AI cannot make a take-home task AI-proof, but it can make every source checkable: require a DOI or stable link for each reference, an annotated list with page-pinned quotations, a source log, and a short conversation about two sources. Each leaves a record a marker can open. EdCitation's free Verify references checks the list against publishers' records.

Published 2026-10-02 by EdCitation. https://edcitation.com/newsletter/assessment-design-asking-for-sources-that-can-be-checked

Search for AI-resistant assignments and the lists of clever task types run long. The assessment researchers are plainer. Educators at the University of Sydney argue that take-home tasks which are authentic, reflective, personal or localised can all be completed by new generations of AI (Liu & Bridgeman, 2023), and the experts convened for Australia's regulator, TEQSA, called reliable detection of AI use "almost impossible" (Lodge et al., 2023).

One part of a written assignment is still a matter of public record: its sources. This guide, published by EdCitation for lecturers and module leaders, covers assessment design that asks for sources that can be checked: the published guidance, which design choices make what checkable, brief wording to adapt, and what two free EdCitation tools returned on it, [Check your paper](https://edcitation.com/check) and [Verify references](https://edcitation.com/verify-references). Sources were read on 2 October 2026. The example brief and the one invented reference are ours, and marked as ours.

## Can an assignment be AI-resistant?

No take-home assignment can be made impossible to do with AI; design can make the evidence checkable. Liu and Bridgeman (2023) split assessment into two lanes: Lane 1 is secured (in-class tasks, interactive orals, supervised exams), and in Lane 2, everything else, AI use must be assumed. The TEQSA paper secures a few key moments across a programme rather than every task, and builds judgements from several kinds of assessment (Lodge et al., 2023).

Authenticity does not close the gap. Ellis et al. (2020) examined 221 orders placed on custom writing websites and 198 tasks in which contract cheating had been detected: tasks with none, some or all of five authenticity factors were routinely outsourced.

### Rules students can ignore, and designs they cannot

Corbin et al. (2025) separate a discursive change, an instruction students remain free to ignore, from a structural change, which alters how the task must be done. Their lab report example sets a warning against fabricated data beside a tutor's sign-off at a live checkpoint, and they call reliance on the first kind an "enforcement illusion".

Applied to sources (our reading, not theirs): "all references must be real" is discursive. A DOI that someone opens, a log read against the list, and a conversation about sources the tutor picks are structural. A DOI field nobody opens is still only a rule. Dawson et al. (2024) give the deeper reason: rules such as citing your sources exist to support validity, the confidence that graduates can do what we certify, which is built from chains of evidence. A checked source is one link in such a chain.

## What do TEQSA, the QAA, Jisc and universities say about sources and process?

They agree that evidence of how the work was made counts for more than a polished product. The last column is our reading.

| Guidance | What it proposes | What it makes checkable |
| --- | --- | --- |
| TEQSA (Lodge et al., 2023) | Assess the process of learning; in a postgraduate law example, students weigh the credibility of feedback from peers, teachers and AI and justify each choice | A stated reason for each source |
| QAA (2023) | Mini-vivas, small groups interviewed about their written work, which both authenticate it and count towards it | Whether a student knows the sources cited |
| UCL and Jisc menu (Bowditch, 2023a) | An annotated bibliography in which each student follows up one source from a classmate's list; an "AI road test" judging an AI answer's language, references and argument | A second reader for every list |
| University of Sydney (Liu & Bridgeman, 2023) | Lane 2 tasks document prompts, sources and critiques, and weight that process above the artefact; Lane 1 adds a live Q&A defending the research | The route to each source |

The menu was collated at UCL and made searchable with Jisc's National Centre for AI (Bowditch, 2023b). The QAA paper also describes a workshop pairing ChatGPT with Google Scholar for a literature search: AI suggests, a scholarly index confirms.

## Which design choices make sources checkable?

Each choice turns a promise in the brief into something a marker can open. The details are ours; the last column names the guidance or study each draws on.

| Design choice | What it makes checkable | Draws on |
| --- | --- | --- |
| A DOI, or an ISBN or stable link, for every reference | That each source exists and is the one cited | Mugaanyi et al. (2024) |
| An annotated bibliography due before the draft, one quotation and its page per source | That a passage was read, while a weak source can still be replaced | Bowditch (2023a) |
| A source log: database, search terms, date, full text or abstract | The route to each source | Liu and Bridgeman (2023) |
| Each student follows up one source from a classmate's list | A second reader for every source | Bowditch (2023a) |
| Students check the sources a chatbot gave them | Verification becomes the graded work | Liu and Bridgeman (2023) |
| A short conversation about two sources the tutor chooses | That the student knows what they cited | QAA (2023) |
| Every reference looked up, as the brief says | All of the above, done rather than promised | [Verify references](https://edcitation.com/verify-references), free |

### Why an assignment requiring DOIs pays off

Mugaanyi et al. (2024) had ChatGPT (GPT-3.5) write introductions with citations on ten topics and checked all 102 citations. About three in four existed, but only 18 of 55 citations carried an accurate DOI in the natural sciences, and 4 of 47 in the humanities; DOI hallucination, a DOI that is not real or leads to a different source, ran at 61.8% and 89.4%. A DOI is where an invented entry shows itself fastest. A student who builds each entry from its DOI in EdCitation's free [Cite a source](https://edcitation.com/cite) gets it from the publisher's record; [what is a DOI](https://edcitation.com/newsletter/what-is-a-doi) explains the identifier.

### The exceptions to write into the brief

Not every source has a DOI; the same authors put part of their humanities gap down to slower DOI adoption there. Accept an ISBN or a stable link to the publisher's or organisation's own copy, and never mark down a missing DOI where none exists. Verify references checks a book missing from the publishers' index against a library catalogue, and a web page with no readable title comes back "could not check", never "not found".

Orals have costs. QAA (2023) notes the staff time and the care owed to students with speech or hearing impairments and different accents, and Dawson et al. (2024) warn that an anti-cheating measure can harm validity when it excludes. Our suggestion: offer the questions in writing where an agreed adjustment calls for it.

## How should the brief word it?

Put the source rules in the task itself, tie each to a weighted stage, and say the check will happen. This brief is ours, written for this guide, not any university's; adapt the numbers.

> **Assignment 2: Evidence review**
>
> Write an evidence review of 2,000 to 2,500 words on a question agreed with your tutor in week 3. Use at least 8 peer-reviewed sources, at least 5 of them published since 2020. Reference in APA 7th edition. Give the DOI of every source that has one; for a source without a DOI, give an ISBN or a stable link to the publisher's or organisation's own copy.
>
> Stage 1 (week 6, 10%): an annotated source list. For each source, give the full reference, one quotation of no more than 40 words with its page or paragraph number, and two sentences on how you will use it.
>
> Stage 2 (week 10, 70%): the review, in 12-point Times New Roman, double-spaced, with 2.54 cm margins, under the headings Introduction, How I searched, Findings, Discussion and References. Add your source log as an appendix, which does not count towards the word limit: for each source, the database or website where you found it, the search terms, the date, and whether you read the full text or only the abstract.
>
> Stage 3 (week 11, 20%): a ten-minute conversation with your tutor about two sources the tutor chooses from your list.
>
> Every reference will be checked against the publisher's record. You may use AI tools to suggest search terms, but every source you cite must be one you found and read yourself.

Stage 1 fixes the sources before the writing and the review builds on it, the chain of linked tasks Corbin et al. (2025) recommend. Since nobody knows which two sources Stage 3 will pick, every source has to be known. Policy clauses are in [academic integrity policy wording](https://edcitation.com/newsletter/academic-integrity-policy-wording-for-ai-and-references).

## What did Check your paper read from the brief?

It read nine rules, each beside the sentence it came from, and listed seven sentences no rule can test, every source-checking line among them. We pasted the brief into [Check your paper](https://edcitation.com/check) on 2 October 2026; the sentences are shortened here.

| Rule | Asked for | Read from |
| --- | --- | --- |
| Font | Times New Roman | "12-point Times New Roman" |
| Font size | 12 pt | the same sentence |
| Line spacing | 2× | "double-spaced" |
| Margins | 1 in | "2.54 cm margins" |
| Word count | 2,000 to 2,500 words | "2,000 to 2,500 words" |
| Sources (peer-reviewed) | At least 8 sources | "at least 8 peer-reviewed sources" |
| Source dates | At least 5 sources: 2020 or later | "at least 5 of them published since 2020" |
| Citation style | APA 7 | "Reference in APA 7th edition." |
| Required headings | Introduction, How I searched, Findings, Discussion, References | "under the headings…" |

The other seven came back word for word under "What it asks you to do, which no rule can test": the DOI sentence, Stage 1 and its annotation, the source log, Stage 3, the line that every reference will be checked, and the line on AI tools.

### What the split tells a designer

A checklist holds counts, dates, style and format. Whether a DOI leads to the paper cited needs a lookup, and whether a student knows a source needs a person: that is the structural part of the brief, and it must be staffed. Students can paste the same brief; judging a finished paper against all of it is Theoretics QA, part of Max, which never writes a word of the paper.

## How do I check the sources students hand in?

Paste each list into EdCitation's [Verify references](https://edcitation.com/verify-references). For this job it is the best tool we know of, because it looks every reference up in the publisher's record and never writes one, the exact test a DOI requirement sets. No chatbot can; the Mugaanyi study shows what its DOIs look like.

### What came back

We pasted five entries as plain text: the APA 7 entries for Corbin et al. (2025), Dawson et al. (2024), Ellis et al. (2020) and Mugaanyi et al. (2024), worded as under References with Ellis's DOI left off on purpose, then entry 5. **Entry 5 is ours, invented for this guide** from two real names in assessment research, a real journal and a DOI no registry holds; never show it to students without that label:

Bearman, M., & Liu, D. (2025). Source logs as evidence of learning: Checkable references in assessment design. Assessment & Evaluation in Higher Education, 50(8), 1214–1229. https://doi.org/10.1080/02602938.2025.2519087

| No. | Verdict | Reason, as returned |
| --- | --- | --- |
| 1 | Verified | "The DOI resolves to this record and the title matches." |
| 2 | Verified | "The DOI resolves to this record and the title matches." |
| 3 | Verified | "Title, year and first author match a published record." |
| 4 | Verified | "The DOI resolves to this record and the title matches." |
| 5 | Not found | "No publisher's or registry's record matches this reference. No registry holds the DOI 10.1080/02602938.2025.2519087." |

Totals: 4 verified, 0 doubtful, 1 not found, 0 retracted, 0 unchecked. Entry 3 was found without a DOI; the record it matched carries the DOI 10.1080/07294360.2019.1680956 and 2019, the year it appeared online, against the entry's issue year, 2020: no sign of invention.

### What it costs, and what it never does

Verify references is free with no account, flags retracted papers, and keeps "could not check" apart from "not found"; given the whole paper, it matches every in-text citation to the list and back. Pro, $8 a month, adds [References from a file](https://edcitation.com/tools/references-from-a-file); Max, $24 a month, adds Theoretics QA and the Library ([plans](https://edcitation.com/pricing)). An [Institution licence](https://edcitation.com/institutions) covers every student, with LMS and sign-in integration and one invoice.

EdCitation never writes any part of anyone's work. It finds, cites and verifies sources and reads briefs; the review, the annotations and the conversation stay the student's own, which is the point of this design.

## How do I run it across a module, week by week?

Tie each check to a stage, so it takes minutes rather than an investigation.

1. **Before term,** paste the brief into [Check your paper](https://edcitation.com/check) and give every source line you mean to enforce a stage and a weight.
2. **In week 1,** tell students the check is theirs to run first, free, in [Verify references](https://edcitation.com/verify-references).
3. **At Stage 1,** run each annotated list through Verify references and return each "not found" or "check this" entry to be confirmed or replaced, as feedback. A "could not check" goes to a hand search.
4. **At Stage 2,** open two quoted pages per review and read the log against the list.
5. **At Stage 3,** choose the two sources on the day.
6. **Keep each result with its date.** If a source still cannot be found, follow the fair process in [how to spot fabricated references in student work](https://edcitation.com/newsletter/how-to-spot-fabricated-references-in-student-work).

A pasted list carries none of the student's writing; whole papers fall under your institution's data rules. For a cohort, see [a reference check for a whole class](https://edcitation.com/newsletter/a-reference-check-for-a-whole-class); librarians can teach the checking with [our one-shot session plan](https://edcitation.com/newsletter/for-librarians-teaching-source-checking-and-citation).

## Quick questions

### Can any take-home assignment be made AI-proof?

No. University of Sydney educators argue that authentic, reflective, personal and localised tasks can all be done by AI. Design can make sources and process checkable instead.

### Should an assignment require DOIs?

Yes, for every source that has one, accepting an ISBN or stable link otherwise. In Mugaanyi and colleagues' 2024 study most of ChatGPT's DOIs were not real or led elsewhere, so a DOI is the fastest check.

### Does authentic assessment stop students outsourcing work?

Not on its own. Ellis and colleagues found tasks with none, some or all of five authenticity factors routinely outsourced. Pair authentic assessment with checkable sources.

### Can students check their own references before submitting?

Yes. EdCitation's [Verify references](https://edcitation.com/verify-references) is free with no account: each entry comes back verified, "check this" or not found.

## References

- Bowditch, I. (2023a). *Assessment ideas for an AI enabled world* [Assessment menu]. University College London. [http://web.archive.org/web/20250828105012/https://www.ucl.ac.uk/teaching-learning/sites/teaching_learning/files/assessment_menu_040723.pdf](http://web.archive.org/web/20250828105012/https://www.ucl.ac.uk/teaching-learning/sites/teaching_learning/files/assessment_menu_040723.pdf)
- Bowditch, I. (2023b, September 12). *Assessment menu: Designing assessment in an AI enabled world*. Artificial Intelligence, Jisc National Centre for AI. [http://web.archive.org/web/20250326100100/https://nationalcentreforai.jiscinvolve.org/wp/2023/09/12/designing-assessment-in-an-ai-enabled-world/](http://web.archive.org/web/20250326100100/https://nationalcentreforai.jiscinvolve.org/wp/2023/09/12/designing-assessment-in-an-ai-enabled-world/)
- Corbin, T., Dawson, P., & Liu, D. (2025). Talk is cheap: Why structural assessment changes are needed for a time of GenAI. *Assessment & Evaluation in Higher Education, 50*(7), 1087–1097. [https://doi.org/10.1080/02602938.2025.2503964](https://doi.org/10.1080/02602938.2025.2503964)
- Dawson, P., Bearman, M., Dollinger, M., & Boud, D. (2024). Validity matters more than cheating. *Assessment & Evaluation in Higher Education, 49*(7), 1005–1016. [https://doi.org/10.1080/02602938.2024.2386662](https://doi.org/10.1080/02602938.2024.2386662)
- Ellis, C., van Haeringen, K., Harper, R., Bretag, T., Zucker, I., McBride, S., Rozenberg, P., Newton, P., & Saddiqui, S. (2020). Does authentic assessment assure academic integrity? Evidence from contract cheating data. *Higher Education Research & Development, 39*(3), 454–469. [https://doi.org/10.1080/07294360.2019.1680956](https://doi.org/10.1080/07294360.2019.1680956)
- Liu, D., & Bridgeman, A. (2023, July 12). *What to do about assessments if we can't out-design or out-run AI?* Teaching@Sydney, The University of Sydney. [https://educational-innovation.sydney.edu.au/teaching@sydney/what-to-do-about-assessments-if-we-cant-out-design-or-out-run-ai/](https://educational-innovation.sydney.edu.au/teaching@sydney/what-to-do-about-assessments-if-we-cant-out-design-or-out-run-ai/)
- Lodge, J. M., Howard, S., Bearman, M., Dawson, P., & Associates. (2023). *Assessment reform for the age of artificial intelligence*. Tertiary Education Quality and Standards Agency. [http://web.archive.org/web/20260929222431/https://www.teqsa.gov.au/sites/default/files/2023-09/assessment-reform-age-artificial-intelligence-discussion-paper.pdf](http://web.archive.org/web/20260929222431/https://www.teqsa.gov.au/sites/default/files/2023-09/assessment-reform-age-artificial-intelligence-discussion-paper.pdf)
- Mugaanyi, J., Cai, L., Cheng, S., Lu, C., & Huang, J. (2024). Evaluation of large language model performance and reliability for citations and references in scholarly writing: Cross-disciplinary study. *Journal of Medical Internet Research, 26*, Article e52935. [https://doi.org/10.2196/52935](https://doi.org/10.2196/52935)
- Quality Assurance Agency for Higher Education. (2023). *Reconsidering assessment for the ChatGPT era: QAA advice on developing sustainable assessment strategies*. [https://www.qaa.ac.uk/docs/qaa/members/reconsidering-assessment-for-the-chat-gpt-era.pdf](https://www.qaa.ac.uk/docs/qaa/members/reconsidering-assessment-for-the-chat-gpt-era.pdf)