A Toolkit for Sustainable Assessment in the Age of Generative AI
Six practical guidelines for instructors and teaching leads who need assessments that hold up — whether students use generative AI or not. Built from evidence gathered across ten countries and five continents.
Why this project exists
Assessment only works if the work submitted is the student's own. Generative AI tools such as ChatGPT have made this assumption harder to hold in practice. Rewriting every assessment, every term is not something university instructors can sustain.
We treat this as a sustainability problem rather than a detection problem. Instead of asking which submissions were written by AI, we ask which assessment designs stop depending on the answer.
Find the vulnerabilities
Real assessments from undergraduate and graduate courses, completed by ChatGPT and Gemini, then marked by the instructors who wrote them.
Understand the differences
Regional focus groups on how instructors read, apply and work around their institutions' AI policies and what support they say they need.
Hand back something usable
This toolkit: guidelines, worked examples, revision moves and policy guidance, free for any educator to use or adapt.
The project team
Funded through the Worldwide Universities Network Research Development Fund and led from the University of Alberta.
Okan Bulut
Measurement, Evaluation and Data Science
University of Alberta, Canada
Principal investigator, responsible for coordinating data collection across partner institutions and for the statistical analysis of how assessment format, discipline, degree level and AI platform interact. Research interests span educational measurement, AI-powered automated scoring and feedback generation, learning analytics, and the responsible use of AI in assessment.
Gavin Brown
Educational assessment · New Zealand
Ryan Nefdt
Philosophy · South Africa
Timothy Teo
Educational psychology · Hong Kong
Judith Kleine Staarman
Education and digital technology · United Kingdom
Fred Kofi Boateng
Higher education · Ghana
Petch Sajjacholapunt
Computer science · Thailand
Alex Okello
Physics · Uganda
Andrea Barrett
Teacher education · United States
Stephen Woodcock
Mathematics · Australia
Marlit Annalena Lindner
External partner · Germany
Lisa Kim
External partner · Australia
Graduate research assistants
Doctoral and master's students at the University of Alberta who carried out the assessment scoring, interview coding and analysis behind these findings.
Joyce Liu
PhD student in Measurement, Evaluation, and Data Science · University of Alberta
Eunji Kim
Master's student in Measurement, Evaluation, and Data Science · University of Alberta
Andrea Gonzalez Garcia
Master's student in Measurement, Evaluation, and Data Science · University of Alberta
What we found
We ran two phases. In the first, we gave real university assessments to two frontier AI systems and scored their answers against the instructors' own keys and criteria. In the second, we interviewed faculty across five world regions about what they are doing in response.
AI performance by assessment format
Both systems performed strongly across every format we could score. The differences between formats are small, and they do not run in the direction commonly assumed: constructed-response tasks were not more resistant than multiple-choice ones.
Mean percentage scores. Based on 47 assessments comprising 821 items. Multiple choice n = 31, constructed response n = 8, mixed format n = 8; each assessment weighted equally.
SD 10.20
SD 9.67
ChatGPT scored full marks on 15 assessments and Gemini on 16 — roughly a third of the sample in each case.
No assessment in the sample defeated either system outright. The weakest result was a pass.
For the formats we could score, AI performance is not the differentiator — the design of the task is. Which is why the rest of this toolkit is about design, not detection.
What instructors told us
Thirty faculty members across North America, Europe, Africa, Asia and Australia, interviewed about how their assessment practice is changing. Six patterns ran through the interviews.
Instructors in every region described assignments, essays and problem sets that stopped functioning as evidence of learning almost overnight — problem descriptions could simply be pasted in and solved.
Many are reintroducing in-person, handwritten and supervised assessment. Several described it as going "full circle" after years of moving toward flexible and online formats — pragmatic rather than welcome.
Oral examinations, vivas, recorded explanations and live demonstrations emerged as the preferred alternative, since students must articulate reasoning in the moment. Scalability in large courses is the open question.
Reflective writing, learning journals and process records were widely seen as resistant, because they require students to narrate their own experience — and were valued pedagogically in their own right.
Some instructors design to prevent AI use; others consider that a losing battle and design to teach students to work with it well. A third group holds the question open. Region and discipline shape which view dominates.
Faculty found themselves revisiting long-standing assumptions: whether exams measure what matters, which knowledge and skills are now worth developing, and what a degree certifies.
Institutional AI policies at partner universities View policies
Publicly available AI policies and guidance from institutions involved in this project, gathered so that instructors and teaching leads can compare how different universities have framed the same problem. Each entry records the date we consulted it.
Institutional guidance changes often, and some of these documents sit behind staff logins. Always check your own institution's current policy before relying on any of them. The policy for Makerere University is still being confirmed.
University of Alberta
Canada
Cited 1 October 2026
University of Rochester
United States
Cited 1 October 2026
Leibniz Institute for Science and Mathematics Education
Germany
Cited 1 October 2026
Mahidol University
Thailand
Cited 1 October 2026
University of Ghana
Ghana
Cited 1 October 2026
University of Technology Sydney
Australia
Cited 1 October 2026
University of Sydney
Australia
Cited 1 October 2026
University of Auckland
New Zealand
Cited 1 October 2026
University of Exeter
United Kingdom
Cited 1 October 2026
University of Cape Town
South Africa
Cited 1 October 2026
The Chinese University of Hong Kong
Hong Kong
Cited 2 October 2026
What we found about institutional policy
Policy existed at most participating universities, but it was written and implemented inconsistently — and the majority of instructors told us they wanted more specificity, particularly about when students may and may not use AI.
Instructors described the reasons for that vagueness sympathetically: AI changes too fast for fixed wording, universities are reluctant to be seen policing students, and the prior question — what knowledge and skills matter now — has not been settled. Where no policy existed, instructors were left to write their own, or to wait.
Latitude intended as academic freedom was frequently experienced as abandonment: one participant described a policy with so much latitude that it left everything to the individual instructor to decide.
Four approaches we observed in practice
Scale back
Restrict the format — no take-home tests, pen-and-paper only.
Permitted with guidance
Students may use AI, with explicit instructions on how and where.
Permitted with citation
Full use allowed, provided AI-generated portions are cited.
Open use
Assessments revised proactively on the assumption students will use AI.
These approaches co-existed not only across institutions but often within a single university — and sometimes within a single instructor's own courses.
For teaching and learning leads
What follows is what we would advise an institution to do with the policy findings above — offer instructors real choices, revisit the wording as the technology moves, and treat professional learning as a multi-year commitment rather than a workshop.
What we recommend
Give instructors a menu, not a mandate
Offer a set of sanctioned options instructors can choose between, based on their course objectives, their own readiness to work with AI, and the design of the assessment.
Review the policy annually
Revise it as the technology changes and as feedback arrives from students and instructors. A policy written once will be wrong within a year.
Plan professional learning over several years
Instructors consistently named time as the binding constraint. Make the format flexible and collaborative, and evaluate it as often as needs shift.
Questions worth building training around
Drawn from what instructors in our study said they still needed answered.
How do learning objectives change in the age of AI?
Which types of assessment are most resilient to AI use?
Which types of assessment are most compatible with AI use?
What AI-supported tasks best measure student understanding?
How can colleagues and institutions collaborate to share practice?
Turning the findings into practice
Everything below is the practical half of this project: check an assessment you already run, find the revision moves for its format, read the design guidelines, and browse worked examples from partner institutions. All of it is free to use and adapt.
Six design guidelines
Held in balance deliberately: three reduce what AI can do for a student unseen, three make AI use legitimate and visible. Neither half works alone.
Grade drafts, decisions, discarded approaches and revision notes alongside the final artefact. A process record is cheap for an honest student to produce and expensive to fabricate.
Data your class collected, a case discussed in week seven, a reading not available online, a site students visited. Generality is what makes a prompt easy to answer well.
Five minutes of structured questions on a submission tells you more than any detector. Build it in as a small graded component rather than an investigation triggered by suspicion.
A short statement of which tools were used, for what, and what the student changed. Assess the judgement shown in that account, not merely its presence.
Give students a model-generated answer and ask them to find where it is wrong, shallow or unsupported. The task becomes disciplinary judgement, which is what you wanted to assess.
"Permitted", "required", "not for this one" — written on the assessment itself. Most confusion we heard from instructors and students came from blanket policies applied to mixed workloads.
Redesign moves by assessment type
Pick the kind of task you're revising. Each one keeps the learning outcome intact and changes only what the task asks for.
Effort and disruption vary; most departments find two of the three moves enough to change the exposure profile.
Worked examples
Six assessments redesigned for reuse, each mapped to the design guidelines it applies. Each states what AI may be used for, because that is part of the design. The growing library holds the rest, including redesigns sent in by instructors using this toolkit.
Have you redesigned an assessment?
Instructors in our study asked for examples from colleagues far more often than they asked for policy. If you have revised a task and would let others learn from it, send it to us and we will add it to the library with credit to you and your institution — a short description of the before and after is enough to start.
Work with us
The funded period has concluded, but the work continues! This toolkit will grow with what people send us. We are interested in hearing from instructors, institutions and teaching centres in four areas in particular.
Assessment redesigns for the library — send us a task you have revised, with a line on what changed, and we will publish it with credit
Replicating the first-phase evaluation in new disciplines, languages or national systems
Longitudinal work on whether redesigned assessments hold their resilience over several cohorts
Joint applications to SSHRC, NSF, UKRI and equivalent national funders
Share a redesign, or tell us what you're working on
This composes a message in your own mail client — nothing is sent from this page and nothing is stored.