The DOI matches its Crossref record; 18 of the 19 references it cites were found in a registry or an index; 18 of the 41 sources its claims cite were checked for support.
Designing AI-resilient assessment in higher education: a four-pillar conceptual framework
Open accessAbstract
Generative AI tools can produce polished academic text on demand, undermining the validity of assessments that treat written submissions as evidence of individual learning. Detection-based countermeasures have demonstrated variable accuracy and equity concerns. This paper does not report empirical outcomes or validation data. It proposes a framework for AI-resilient assessment that shifts evaluation from product quality to demonstrable reasoning, decision-making, and ownership of learning.
Overview
CoChat Summary
TL;DR
This conceptual paper proposes a four-pillar framework for AI-resilient higher education assessment—process-based documentation, oral defense integration, authentic task design, and transparent AI-use policies—to shift evaluation from product quality to demonstrable reasoning and ownership of learning. It offers design tools (rubric, oral-defense protocol, AI-use taxonomy, vulnerability audit) but reports no empirical outcomes or validation data.
Key findings
- The framework derives four pillars from Messick’s unified validity framework and constructive alignment, mapping each generative-AI threat to a design response: invisible process → process documentation; pre-prepared output → oral defense; generic task → authentic task design; ambiguous expectations → AI-use policy.
- Operational tools include a rubric with weight ranges (reasoning 25–35%, oral defense 25–35%, process engagement 20–30%, product quality 10–20%; worked example total 3.45/4.00), an oral-defense protocol of 8–12 min per student (warm-up 2–3 min, probing 4–6 min, extension 2–3 min), and a graduated AI-use taxonomy (Prohibited, Guided, Permitted).
- Workload estimates: 100 students at 10 min each ≈17 h direct assessment time, with a realistic range of 25–35 h including preparation and moderation; random sampling of 30–40% reduces workload but introduces a validity trade-off; a large cohort (n = 300) is estimated at ~50 additional faculty hours per assessment cycle.
- Cited detection evidence: Weber-Wulff et al. evaluated 14 tools, with GPTZero and ZeroGPT producing false-positive rates exceeding 20% on specific samples; Liang et al. demonstrated systematic misclassification of non-native English writing.
Methods
- Conceptual analysis, not an empirical study or systematic review; a focused synthesis across five domains using Scopus, Web of Science, and Google Scholar (January 2020–March 2026), with search strings combining “generative AI” or “large language model” with assessment, academic integrity, oral defense, and authentic assessment, plus backward citation tracking.
- Sources were included if they addressed assessment design, validity, or integrity in the context of generative AI or automated text production; Table 1 frameworks were selected if published 2024–2026, explicitly addressed assessment redesign, and described at least one focal dimension with operational detail.
- Table 1 components were coded present, partial, or absent by both authors independently; initial agreement was 14/16 decisions (87.5%), disagreements resolved through discussion, and the present framework was self-coded.
Limitations
- Stated: the framework has not been empirically validated; the proposed tools have not been evaluated for reliability, inter-rater consistency, or transferability; illustrative examples are drawn primarily from health sciences and engineering, so cross-disciplinary validation is needed.
- Stated: oral defense creates accessibility and equity concerns for students with speech or hearing impairments, severe anxiety or autism spectrum conditions, non-native speakers, and fully online programmes; it may introduce linguistic biases, and the “equity paradox” cannot be fully resolved within the framework alone.
- Stated and evident: the framework is AI-resilient, not AI-proof, and each pillar may be circumvented as AI capabilities evolve; random sampling has an acknowledged validity trade-off; vulnerability levels and thresholds are heuristic, not validated; Table 1 self-coding represents a potential bias; institutional, pedagogical, and cultural barriers may limit adoption.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Figures

Verification
Claims the paper makes
12 Valid · 7 Not valid · 16 Not checked · 70 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Many early institutional responses to generative AI have emphasized detection tools and misconduct procedures, often more than assessment redesign (Weber-Wulff et al., 2023).”
§ Designing AI-resilient assessment in higher education: a four-pillar conceptual framework > Introduction
What the sources say
- Partly supportedcites Weber-Wulff et al., 2023 sourceExperimental — accuracy not yet measured
The Introduction (p. 2) describes institutional bans and says restrictions on AI-generated content led to a desire for detection tools. The Conclusion (p. 27) urges educators to focus on prevention and rethink assessment rather than focus on detection. The paper does not establish that many institutions emphasized detection tools and misconduct procedures more than assessment redesign. (LLM, conf=0.82)
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“In professional education, a student may construct an exemplary case summary with AI assistance yet remain unable to justify a treatment decision under uncertainty (Kung et al., 2023).”
§ Designing AI-resilient assessment in higher education: a four-pillar conceptual framework > Introduction
What the sources say
- Could not verifycites Kung et al., 2023 sourceExperimental — accuracy not yet measured
The Discussion section on AI-assisted human learning evaluates ChatGPT’s explanations as a proxy for helpfulness and says studies with students are still needed. Our copy does not address whether a student can produce an AI-assisted case summary yet cannot justify a treatment decision under uncertainty; searches for those terms found no such passage. (LLM, conf=0.75)
hedgedstates a possible cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Could not verify
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“detection–evasion cycles may foster surveillance rather than trust”
§ Designing AI-resilient assessment in higher education: a four-pillar conceptual framework > Introduction
What the sources say
- Supportedcites Liang et al., 2023 source
The Discussion describes a potential cycle in which tighter detection leads non-native writers to use GPT more to evade false flags. It also says detector use can create a “presumption of guilt” and provoke mistrust, supporting the claim’s contrast with trust. “Surveillance” is a paraphrase of that detector-driven scrutiny, not the paper’s term. (LLM, conf=0.86)
- Could not verifycites Elkhatat et al., 2023 sourceExperimental — accuracy not yet measured
The Discussion recommends a holistic approach to AI detection and fostering a culture of honesty; the Conclusion calls for detectors to evolve alongside AI text generation. Neither section says that detection–evasion cycles may foster surveillance rather than trust. Searches for surveillance, trust, and evasion found no passage addressing that claim. (LLM, conf=0.73)
hedgedstates a possible cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“detection accuracy varies across writing styles and languages”
§ Designing AI-resilient assessment in higher education: a four-pillar conceptual framework > Introduction
What the sources say
- Partly supportedcites Liang et al., 2023 sourceExperimental — accuracy not yet measured
The Introduction reports that detection accuracy differed between native and non-native English writing and changed when word choice was enhanced or simplified. The paper does not compare detection accuracy across different languages; the writing samples discussed are in English. (LLM, conf=0.86)
- Partly supportedcites Elkhatat et al., 2023 sourceExperimental — accuracy not yet measured
The Discussion and Conclusion (pages not provided) report that detection performance varies across GPT-3.5-generated, GPT-4-generated, and human-written text. They do not report a comparison across writing styles or languages, so the claim attributes more specific sources of variation than the study establishes. (LLM, conf=0.85)
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“By requiring drafts, revision logs, and reflective checkpoints, it creates documentation that is less easily outsourced to generative AI alone (Boud and Molloy, 2013).”
§ Designing AI-resilient assessment in higher education: a four-pillar conceptual framework > A four-pillar framework for AI-resilient assessment > Pillar 1: process-based assessment
What the sources say
- Partly supportedcites Boud and Molloy, 2013 sourceExperimental — accuracy not yet measured
Feedback Mark 2 (p. 11) advocates assessment tasks with multiple stages and feedback used to improve performance; The curriculum and what it promotes (p. 16) calls for regular student self-analysis. These support the process-based aspect of the claim, but the paper does not specify drafts, revision logs, or reflective checkpoints as required documentation, or say that such documentation is harder to outsource to generative AI. (LLM, conf=0.86)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
27 more claims · 5 with their cited sources checked
Create a free account to see all 105 claims and every source verdict.
- It may state two separate claims as one. 27
- Probably not a substantial claim. 63
Citations
Cites
Kung, T. H., Cheatham, M., Medenilla, A., Sillos, C., De Leon, L., Elepaño, C., Madriaga, M., Aggabao, R., Diaz-Candido, G., Maningo, J., et al. (2023) · PLOS Digital Health · doi:10.1371/journal.pdig.0000198 · Cited by 4,115 · cited 1×
“In professional education, a student may construct an exemplary case summary with AI assistance yet remain unable to justify a treatment decision under uncertainty (Kung et al., 2023).”
Chatting and cheating: Ensuring academic integrity in the era of ChatGPT
Cotton, D. R. E., Cotton, P. A., & Shipway, J. R. (2024) · Innovations in Education and Teaching International · doi:10.1080/14703297.2023.2190148 · Cited by 3,087 · cited 1×
“More fundamentally, detection addresses unauthorized AI use without confronting whether the assessment itself measures what educators value in graduates (Bearman et al., 2023; Cotton et al., 2024).”
Assessment in medical education
Epstein, R. M. (2007) · New England Journal of Medicine · doi:10.1056/nejmra054784 · Cited by 1,872 · cited 1×
“” Such questions require real-time clinical reasoning that cannot be adequately prepared through AI-generated text alone (Epstein, 2007).”
Rethinking models of feedback for learning: the challenge of design
Boud, D., & Molloy, E. (2013) · Assessment & Evaluation in Higher Education · doi:10.1080/02602938.2012.691462 · Cited by 1,636 · cited 1×
“By requiring drafts, revision logs, and reflective checkpoints, it creates documentation that is less easily out-sourced to generative AI alone (Boud and Molloy, 2013).”
Lim, W. M., Gunasekara, A., Pallant, J. L., Pallant, J. I., & Pechenkina, E. (2023) · The International Journal of Management Education · doi:10.1016/j.ijme.2023.100790 · Cited by 1,261 · cited 1×
“Rather than blanket prohibitions, the framework advocates graduated policies aligned with intended learning outcomes (ILOs) (Lim et al., 2023).”
GPT detectors are biased against non-native English writers
Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023) · Patterns · doi:10.1016/j.patter.2023.100779 · Cited by 673 · cited 2×
“…have exposed critical limitations: detection accuracy varies across writing styles and languages; false-positive rates disproportionately affect non-native English speakers; and detection–evasion cycles may foster surveillance rather than trust (Liang et al., 2023; Elkhatat et al., 2023).”
Testing of detection tools for AI-generated text
Weber-Wulff, D., Anohina-Naumeca, A., Bjelobaba, S., Foltýnek, T., Guerrero-Dib, J., Popoola, O., Šigut, P., & Waddington, L. (2023) · International Journal for Educational Integrity · doi:10.1007/s40979-023-00146-z · Cited by 582
Authentic assessment: creating a blueprint for course design
Villarroel, V., Bloxham, S., Bruna, D., Bruna, C., & Herrera-Seda, C. (2018) · Assessment & Evaluation in Higher Education · doi:10.1080/02602938.2017.1412396 · Cited by 560 · cited 1×
“Authentic tasks require judgment under ambiguity and integration of multiple knowledge domains, making them less susceptible to generic AI outputs (Wiggins, 1990; Villarroel et al., 2018).”
Discourses of artificial intelligence in higher education: a critical literature review
Bearman, M., Ryan, J., & Ajjawi, R. (2023) · Higher Education · doi:10.1007/s10734-022-00937-2 · Cited by 545 · cited 2×
“More fundamentally, detection addresses unauthorized AI use without confronting whether the assessment itself measures what educators value in graduates (Bearman et al., 2023; Cotton et al., 2024).”
The Case for Authentic Assessment
Wiggins, G. (1990) · University of Massachusetts Amherst · doi:10.7275/ffb1-mm19 · Cited by 493 · cited 1×
“Authentic tasks require judgment under ambiguity and integration of multiple knowledge domains, making them less susceptible to generic AI outputs (Wiggins, 1990; Villarroel et al., 2018).”
Research works printed but not linked: 6
- Biggs, 2011, Teaching for Quality Learning at University
- 2025, Generative AI in Higher Education Teaching & Learning: Policy Framewor. Dublin
- Joughin, 2010, A Short Guide to Oral Assessment
- Messick, 1989, Educational Measurement, 13
- Vygotsky, 1978, Mind in Society
- Winne, 2018, Handbook of Self-Regulation, 36
Cited by
Koberda, T. M. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2609.28140 · Cited by 1
“The importance of assessment that remains informative in the presence of AI is already the subject of a substantial and rapidly developing literature; see, for example, Nikolić and Basta Nikolić [9].”
Metadata
- Authors
- Nikolić, D., & Basta Nikolić, M.
- Affiliations
- University of Novi Sad (RS)
- Faculty of Medicine, University of Novi Sad
- Corresponding author
- Nikolić, Dragan
- ORCID
- Nikolić, Dragan
- Venue
- Frontiers in Artificial Intelligence
- Publisher
- Frontiers Media SA
- Published
- 2026-07-16
- Volume
- 9
- Pages
- 1841682-1841682
- ISSN
- 2624-8212 (ISSN-L)
- PMC
- PMC13422401
- OpenAlex
- W7168776378
- Semantic Scholar
- 5a4e7c083a719eedfd4ac7672364db8b8b120405
- Cited by
- 1
- References
- 16
- Citations by year
- 2026: 1
- FWCI
- 5.63
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- Locations
- Frontiers in Artificial Intelligence (publishedVersion, cc-by)
- PubMed (publishedVersion)
- PubMed Central (publishedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
- Topics
- Social Sciences › Education › Academic integrity and plagiarism
- Social Sciences › Education › Artificial Intelligence in Education
- Social Sciences › Education › Student Assessment and Feedback
- Keywords
- generative AI, higher education assessment, academic integrity, AI detection, assessment design, oral defense, AI use policies, health sciences education, ownership of learning
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Designing AI-resilient assessment in higher education: a four-pillar conceptual framework. https://app.cochat.ai/papers/10.3389/frai.2026.1841682
- The paper
- Nikolić, D., & Basta Nikolić, M. (2026). Designing AI-resilient assessment in higher education: a four-pillar conceptual framework. Frontiers in Artificial Intelligence. https://doi.org/10.3389/frai.2026.1841682
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.