Literature notes

Everything cited below was checked against the source on 2026-10-25 unless marked otherwise. Access method and status are recorded per source. Where a source could not be fetched directly (publisher bot-blocks), that is stated, and the claim rests only on what the publisher’s own index shows. Sources that could not be verified at all are removed or marked UNVERIFIED.

1. Assessment fundamentals

Constructive alignment — Biggs

Biggs, J. & Tang, C., Teaching for Quality Learning at University (Open University Press; multiple editions). Constructive alignment: intended learning outcomes, teaching/learning activities and assessment tasks must point at the same capabilities; assessment drives what students actually do. [Book citation; publisher page not fetched — foundational textbook, statement uncontroversial.]

Validity, and assessment for vs. of learning

• The working question for this kit: “Does this task show the learning we claim it shows?” If an AI tool can produce a passing product without the target learning, the evidence the task yields is no longer valid — not because students “cheated”, but because the task no longer measures what it claims (see stress-test rationale, §2).
• Assessment of learning (assurance of individual achievement) vs. for and as learning (building capability). The University of Sydney two-lane approach operationalises this distinction institutionally (below, §2).
• TEQSA’s assessment-reform advice uses the same distinction and proposes reform principles for AI-era assessment: Lodge, J. et al., Assessment Reform for the Age of Artificial Intelligence, TEQSA discussion paper, September 2023 — https://www.teqsa.gov.au/guides-resources/resources/corporate-publications/assessment-reform-age-artificial-intelligence (search-index verified 2026-10-25; the PDF fetch timed out, so exact principle wording was not quoted from the source).

2. AI-use level frameworks

AI Assessment Scale (AIAS) — Perkins, Furze, Roe & MacVaugh

• Original article (verified, fetched 2026-10-25): Perkins, M., Furze, L., Roe, J., & MacVaugh, J. (2024). “The Artificial Intelligence Assessment Scale (AIAS): A Framework for Ethical Integration of Generative AI in Educational Assessment.” Journal of University Teaching and Learning Practice, 21(06). https://doi.org/10.53761/q3azde36
• Latest version (verified, fetched 2026-10-25): the scale’s official resource site https://aiassessmentscale.com lists AIAS 2.1 as the current revision (first published 2023, revised through Version 2 and 2.1). AIAS 2.1’s stated change is “one clear statement per level”; the level structure is unchanged.
• The five levels, exact names (verified verbatim from aiassessmentscale.com, 2026-10-25):

Level	Exact name
1	No AI
2	AI Planning
3	AI Collaboration
4	Full AI
5	AI Exploration

• A 2025 JUTLP article “Reimagining the AIAS” (https://doi.org/10.53761/rrm4y757) is cited on the official site as the revised model’s publication. DOI resolved via the site’s own citation list; article page not fetched — cite the site for the current level names.
• This kit uses AIAS 2.1 level names on every student-facing brief (ai-use-levels). Icons are always paired with text labels.

Two-lane approach — University of Sydney

• Original post (verified, fetched 2026-10-25): Liu, D. & Bridgeman, A. (2023, 12 July). “What to do about assessments if we can’t out-design or out-run AI?” Teaching@Sydney. https://educational-innovation.sydney.edu.au/teaching@sydney/what-to-do-about-assessments-if-we-cant-out-design-or-out-run-ai/
  • Lane 1 = secured assessment of learning (in-class, viva, live simulation) to assure outcomes.
  • Lane 2 = open assessment as/for learning where AI use is assumed, scaffolded, and the documented process can carry more weight than the artefact.
  • The post explicitly recommends testing assignment prompts against AI tools, and avoiding detection software (unreliability; bias against non-native English writers).
• Implementation follow-up (verified, fetched 2026-10-25): Bridgeman, A., Weeks, R. & Liu, D. (2024, 26 November). “Aligning our assessments to the age of generative AI.” Teaching@Sydney. https://educational-innovation.sydney.edu.au/teaching@sydney/aligning-our-assessments-to-the-age-of-generative-ai/ — new assessment principles (Nov 2023) embedded the two lanes in the university’s Coursework Policy.
• This kit applies the two lanes at task level: each redesign designates a secure lane (a few weighted checks) and an open lane (practice and process work).

National guidance

• TEQSA (Australia) advice cited above. QAA (UK) guidance exists but was not verified during this build and is therefore not cited in these notes. (UNVERIFIED — deliberately not used.)

3. Why not detection

All three primary sources below were fetched on 2026-10-25.

1. Accuracy is poor and unreliable. Weber-Wulff, D., Anohina-Naumeca, A., Bjelobaba, S., Foltýnek, T., Guerrero-Dib, J., Popoola, O., Šigut, P., & Waddington, L. (2023). “Testing of detection tools for AI-generated text.” International Journal for Educational Integrity, 19, 26. https://doi.org/10.1007/s40979-023-00146-z
   • 14 tools tested (incl. Turnitin); none above 80% accuracy; more-obfuscated text and machine-translated text sharply degrade accuracy (~50% misattribution; ~20-point accuracy drop on machine-translated English).
   • Conclusion quoted from the abstract: tools “should not be used in academic settings” as sole evidence of misconduct.
2. Bias against non-native English writers. Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). “GPT detectors are biased against non-native English writers.” Patterns, 4(7), 100779. https://doi.org/10.1016/j.patter.2023.100779 (DOI resolution to the Patterns article PII S2666389923001307 verified 2026-10-25; arXiv version fetched and verified: https://arxiv.org/abs/2304.02819)
   • Detectors “consistently misclassify non-native English writing samples as AI-generated” (abstract, verified) while classifying native-English essays accurately; the authors caution against detector use in evaluative or educational settings. (The commonly cited ≈61% false-positive figure on TOEFL essays appears in the full text; the arXiv abstract page fetched here states the qualitative finding — the number is therefore reported as “over half” only in secondary summaries and not asserted from the fetched primary text.)
3. Detector vendors themselves concede it. OpenAI’s AI Classifier was launched 31 January 2023 and withdrawn on 20 July 2023: “As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy” (notice quoted in the publisher’s page, verified via search index 2026-10-25: https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/).
4. Detection is also not a design strategy: it asks nothing of learning design, invites adversarial games, and mis-targets the students most at risk of unfair accusations (Liang et al., above). This kit designs assessment so that AI use is explicit and levelled rather than policed. Practical corollary from Liu & Bridgeman (2023): assume unsecured work involves AI and design accordingly.

4. Process-based assessment

• Evidence base for staged drafts, annotated bibliographies, process logs and short oral checks (“viva-vee”): these are standard process-assessment moves, and the Sydney two-lane work and TEQSA advice both recommend interactive oral formats for the secure lane. Ratings in approaches are labelled strong / emerging / anecdotal with this caveat: strong = multiple independent studies or institutional adoptions; emerging = single-site studies or documented pilots; anecdotal = practitioner reports. No new meta-analysis was performed for this kit.
• Oral-assessment revival context: Liu & Bridgeman (2023) and Bridgeman, Weeks & Liu (2024) document institutional adoption of interactive orals as the secure lane at scale.

5. Equity and accessibility

• Detection bias against non-native English writers: Liang et al. (2023) and Weber-Wulff et al. (2023) — see §3. Any approach that invites accusations disproportionately harms multilingual students.
• Oral assessment and anxiety / accommodations: interactive orals can disadvantage students with communication disabilities or high anxiety unless a written or asynchronous alternative is built in from the start. This kit’s faculty notes require an alternative format in every case with an oral component (case A, B and C faculty notes).
• Workload at scale: redesigns that cost more grading time than the original will not be adopted; every case’s faculty-notes page includes a time estimate and a 200+ student scaling plan (TA rubrics, sampled orals, cohort-level data collection).
• Digital access: AI-use levels must not assume paid subscriptions. Every student brief states which level can be met with free-tier tools, and Full AI levels are only set where lack of access does not affect the mark (case-by-case notes).

6. Context sources (2026)

All dated 2026-10-25.

1. Alalwani, S. (2026). “Faculty perceptions of ChatGPT on academic integrity and institutional roles in higher education.” Scientific Reports, 16, 27532. https://doi.org/10.1038/s41598-026-63852-x — verified via the publisher’s citation metadata (title, author, volume, DOI). Faculty in administrative roles ask for clearer guidelines on AI in assessment and integrity procedures.
2. Balart, T., Raju, G. & Shryock, K. (2026). “Examining Generative AI Disruption: A Repeated Cross-Sectional Study of Faculty and Staff Sensemaking in Higher Education.” Algorithms, 19(8), 703. https://doi.org/10.3390/a19080703 — MDPI’s site blocks automated fetching (HTTP 403); details verified via the publisher’s search-index record. Faculty discourse is shifting from containment toward selective integration.
3. Carvajal, H. et al. (2026). “Artificial Intelligence and Academic Integrity in Virtual Higher Education.” Trends in Higher Education, 5(3), 67. https://www.mdpi.com/2813-4346/5/3/67 — MDPI fetch blocked (403); title (truncated in the publisher index) and author verified via the publisher’s search-index record; no DOI asserted. Students frame AI as academic support; faculty frame it as authorship/integrity risk.
4. Sangwa, S. (2026). “The EPIQ-AI Framework for Epistemic Integrity.” Preprint, https://www.preprints.org/manuscript/202604.0310 — site blocks automated fetching (403); title/author verified via search index only. Note: the plan that briefed this repo described this preprint as “faculty readiness” evidence; the indexed title under that ID is the EPIQ-AI framework paper. The plan’s characterisation is recorded here as UNVERIFIED and this kit does not rest any claim on it.
5. Teamed, L&D Skills Landscape 2026 and other cross-portfolio context sources are listed in the portfolio overview, not relied on in this repo.

7. What these notes are for

These notes feed the six design principles and the approach comparison. They are the research layer only — the design decisions and their reasoning live in the docs.