The DOI matches its Crossref record; 29 of the 29 references it cites were found in a registry or an index; 13 of the 49 sources its claims cite were checked for support.
A Tripartite Feedback Framework for AI-Assisted Assessment of Complex Reports in Higher Education
Open accessAbstract
Assessment feedback on complex written reports remains one of the most persistent and resource-intensive challenges in higher education. However, no principled framework exists for deciding which feedback tasks might appropriately involve artificial intelligence and which must remain human responsibilities. This paper addresses that gap by proposing a tripartite feedback framework that distinguishes three analytically distinct levels: low-level structural and presentational feedback, intermediate-level factual content validation, and high-level critical evaluation and synthesis. Grounded in established feedback theory, including Hattie and Timperley’s feedback model and Boud and Molloy’s sustainable feedback design principles, the framework provides pedagogically justified criteria for allocating tasks between AI systems and human assessors, rather than automating whatever technology can technically perform. Five non-negotiable boundary principles govern any AI involvement at the intermediate level, preserving human oversight, academic accountability, and assessment integrity. This paper examines current technological capabilities and limitations at each level, proposes a phased implementation pathway with explicit human-in-the-loop requirements, and addresses implications for feedback literacy, student agency, equity, and security. A comprehensive mixed-methods evaluation design specifying the evidence required for empirical validation is also presented. The framework’s contribution lies not in prescriptive solutions but in providing structured categories, explicit boundary conditions, and validation criteria to guide context-sensitive institutional decision-making about AI integration in assessment.
Overview
CoChat Summary
TL;DR
This conceptual paper proposes a tripartite feedback framework that sorts feedback on complex written reports into low-level (structural/presentational), intermediate-level (factual content validation), and high-level (critical evaluation and synthesis) categories, and argues that only the first two may appropriately involve AI, with the intermediate level permitted only under five non-negotiable boundary principles. No empirical validation is presented; the paper instead specifies a mixed-methods evaluation design and research agenda.
Key findings
- The framework distinguishes three analytically distinct feedback levels and argues they are dimensions rather than a hierarchy: a calculation can be correctly presented (low-level) while arithmetically wrong (intermediate-level) and methodologically inappropriate (high-level).
- Five non-negotiable boundary principles govern AI involvement at the intermediate level, and "institutions considering AI involvement at the intermediate level must demonstrate that all five conditions are satisfied simultaneously; satisfying only four is insufficient": (1) assessor-curated knowledge bases, (2) human oversight and final authority, (3) explicit limitation to bounded verification tasks, (4) transparency and attribution, and (5) security and integrity by design.
- Principle 3 requires all four bounded-verification criteria to hold at once: unambiguous documented criteria, direct comparison against established knowledge without interpretive judgment, no assessment of alternative valid approaches, and a single correct answer or pre-specified acceptable alternatives.
- The paper claims three contributions: an operational taxonomy complementing existing feedback theories, five concrete boundary principles, and a mixed-methods evaluation design; it positions itself as complementary to Hattie and Timperley (2007) and Boud and Molloy (2013), which it says do not address the epistemic status of feedback tasks or human–AI task allocation.
Methods
- Conceptual/theoretical paper: "This is a conceptual paper that proposes a theoretical framework; it does not present empirical validation or proven effectiveness."
- Framework developed through iterative analysis of three intersecting domains (feedback theory, AI capabilities evidence, report assessment practice) across five interconnected stages: literature review, mapping report assessment practices, examining AI capabilities, identifying boundary conditions, and integrating security/equity as design constraints.
- No data or sample; the paper proposes (but does not conduct) a mixed-methods evaluation design combining quantitative metrics (turnaround time, staff time per submission, inter-marker reliability, AI error frequency, student satisfaction, learning outcomes) with qualitative methods (student focus groups, staff interviews, process evaluation) and controlled comparison against matched control modules, with RCTs noted as strongest but often infeasible.
Limitations
- Stated: the intermediate/high boundary "may prove contextually ambiguous in practice"; the five principles "have not been tested for simultaneous feasibility," with a risk that conditions are "so demanding that intermediate-level AI support proves impractical in most real institutional contexts"; the framework was developed for STEM and social science report genres and should not be assumed to transfer to essays, portfolios, dissertations, or reflective writing; pedagogical assumptions about feedback literacy are "an assumption rather than a demonstrated finding"; security countermeasures are "theoretically reasonable but empirically unverified"; there is "the absence of an implemented pilot," leaving open which LLM, prompting, and RAG configuration would be used; and equity problems "cannot [be] pre-specify[ied]."
- Stated: the framework may "shift staff effort rather than reduce it," so efficiency gains are an open empirical question; a bounded framework "could still ease future expansion of AI's role in assessment."
- Evident from the text: all claims about benefits are theoretical possibilities, and the paper reports no empirical results, effect sizes, or pilot data of its own; the evaluation design is a proposal, not a completed study.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
9 Valid · 0 Not valid · 23 Not checked · 85 Own results and interpretations
None of the sources cited for this claim have been checked yet.
In the paper
“Academic staff struggle to deliver feedback that is simultaneously timely, comprehensive, and actionable, particularly for complex written assignments such as reports that require judgments across multiple dimensions of performance (Parkinson, 2017; Morris et al., 2021).”
What the sources say
- Not checked yetcites Parkinson, 2017 source
- Could not verifycites Morris et al., 2021 sourceExperimental — accuracy not yet measured
Our copy of Morris et al. (2021) is a systematic review of quantitative effect studies; across six searches (staff difficulty/workload, timely-specific-actionable feedback, written assignments/reports, staff perceptions and resource constraints, recommendations for practice) no passage states that academic staff struggle to deliver feedback that is simultaneously timely, comprehensive and actionable, and reports or multi-dimensional written tasks are never discussed. The nearest content is the 'Evidence-informed formative assessment and feedback practice' section, which reports that practitioners 'are frequently expected to spend extensive amounts of time providing detailed, written responses' with little evidence of effectiveness — but that is credited to Elliott et al. (2016) at compulsory school level, and the DISCUSSION notes only that assessment and feedback receive lower scores from students (OfS, 2019a). The 'Timing and spacing' section finds a 'fairly unclear picture' on immediate versus delayed feedback, which concerns effects on students rather than staff delivery difficulty. (LLM, conf=0.70)
plain- 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.
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Secondly, AI feedback lacks depth and disciplinary specificity, tending toward generic comments rather than targeted observations (Banihashem et al., 2024).”
What the sources say
- Partly supportedcites Banihashem et al., 2024 sourceExperimental — accuracy not yet measured
Our copy supports only part of the attribution: the RQ1 results and 'Discussion on the results of RQ1' report that peers gave higher-quality feedback than ChatGPT, mainly because ChatGPT produced more extensive descriptive feedback (summaries of the essay) while students were better at pinpointing and identifying problems — i.e., fewer targeted observations. However, the paper never assesses 'depth' or 'disciplinary/domain specificity' of the AI feedback; the domain-specific-expertise point in the Introduction is made about peers, not AI, and Table 1 shows no significant differences in justification or constructive features. Indeed, the Discussion states ChatGPT feedback 'comprehensively encompassed all essential components characterizing high-quality feedback' and calls its descriptiveness a strength, so the flat claim that AI feedback 'lacks depth and disciplinary specificity' overstates what this single-course study found. (LLM, conf=0.78)
plain- 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
“Evaluation assesses the quality of a student's reasoning, the appropriateness of their methodological choices, and the sophistication of their disciplinary thinking, a process that requires the situated professional judgment that current AI systems fundamentally cannot exercise (Banihashem et al., 2024; Deeva et al., 2021).”
What the sources say
- Partly supportedcites Banihashem et al., 2024 sourceExperimental — accuracy not yet measured
Our copy of Banihashem et al. (2024) supports only a narrower, hedged version of the point: in the Discussion on RQ1 the authors attribute peers' better problem-identification to 'humans' potential in cognitive skills, critical thinking abilities, and contextual understanding', saying students 'may be better equipped to identify issues within the essays that ChatGPT may overlook'. Nothing in the paper asserts that current AI systems 'fundamentally cannot exercise' the professional judgment needed for evaluation, nor does it assess reasoning quality, methodological appropriateness or disciplinary sophistication as evaluative dimensions — the study is one institution, one course, argumentative-essay feedback (Limitations section). Indeed the same Discussion states ChatGPT's feedback 'comprehensively encompassed all essential components characterizing high-quality feedback, including affective, cognitive, and constructive dimensions' and could 'serve as a viable source of feedback', and the Limitations/implications section advocates integrating ChatGPT alongside peer feedback, so the strong incapacity claim overstates what this source says. (LLM, conf=0.72)
- Not checked yetcites Deeva et al., 2021
plain- 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
Every source the paper cites for this claim was checked and supports it.
In the paper
“Feedback has limited value if students cannot interpret or act upon it.”
What the sources say
- Supportedcites Carless & Boud, 2018 source
The Introduction of Carless and Boud (2018) repeatedly makes this point in its own voice: 'Without the skills to interpret comments received, few students are successful in acting on feedback' and 'Unless students see themselves as agents of their own change ... they may be unable to make productive use of comments about their work' (Introduction, p. 2-3), and 'Information becomes feedback only when students act on it to improve work or learning strategies' plus 'Approaches that emphasise feedback as telling are insufficient because students are often not equipped to decode or act on statements satisfactorily, so key messages remain invisible' (Introduction, pedagogical implications and 'Appreciating feedback processes'). The claim is a broad paraphrase of this framing, so it is supported even though the underlying empirical statements are credited to Robinson, Pope and Holyoak (2013) and Sadler (2010). (LLM, conf=0.88)
plainstates 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
“The first is Carless and Boud's (2018) feedback literacy framework, which holds that students' capacity to interpret and act on feedback depends on understanding where it comes from, how it was generated, and how much weight to give it.”
What the sources say
- Partly supportedcites Carless and Boud's (2018) sourceExperimental — accuracy not yet measured
Our copy of Carless and Boud (2018) proposes a four-feature feedback literacy framework — appreciating feedback, making judgments, managing affect, taking action — and under 'Appreciating feedback' states that feedback literate students 'understand and appreciate the role of feedback in improving work and the active learner role in these processes' and 'recognise that feedback information comes in different forms and from different sources' (Features of student feedback literacy). That supports the claim's general point that interpreting and acting on feedback rests on appreciating where feedback comes from and one's own role in its processes. However, the specific gloss 'how it was generated' and especially 'how much weight to give it' (calibrating the credibility or relative worth of feedback sources) is not stated anywhere in our copy; the paper's adjacent features concern evaluative judgment about the quality of one's own and others' work and managing affect/defensiveness, which the citing sentence drops. The claim therefore over-narrows and re-frames the framework as source appraisal. (LLM, conf=0.75)
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
28 more claims · 4 with their cited sources checked
Create a free account to see all 117 claims and to verify the 36 references still open.
- Probably not a substantial claim. 68
- It may state two separate claims as one. 48
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
§ 3.2.2. Principle 2: Human Oversight and Final Authority
“This means that assessors must review all AI-generated outputs before they reach students;”
§ 5.1.2. Stage 2: AI-Supported Intermediate-Level Checking Under Human Oversight
“Crucially, this stage begins with high-frequency human oversight of all AI outputs, scaling back only as evidence demonstrates acceptable reliability and security.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 3.2.5. Principle 5: Security and Integrity by Design
“a system that satisfies Principles 1 through 4 but concentrates authority in AI outputs rather than human assessors violates the accountability commitments that underpin the entire framework.”
§ 3.2.2. Principle 2: Human Oversight and Final Authority
“Human assessors must retain full oversight and unambiguous final authority at every stage of the assessment process in which AI is involved.”
Citations
Cites
Hattie, J., & Timperley, H. (2007) · Review of Educational Research · doi:10.3102/003465430298487 · Cited by 12,905 · cited 5×
“The transformative potential of feedback in higher education is well established (Hattie & Timperley, 2007), yet its effective provision at scale remains persistently elusive.”
InfluentialZawacki-Richter, O., Marín, V. I., Bond, M., & Gouverneur, F. (2019) · International Journal of Educational Technology in Higher Education · doi:10.1186/s41239-019-0171-0 · Cited by 7,242 · cited 1×
“Critics, however, point to substantial limitations, including factual unreliability, an inability to exercise genuine judgment, and risks of undermining the essentially human dimensions of teaching and learning (Zawacki-Richter et al., 2019; Banihashem et al., 2024).”
Nicol, D. J., & Macfarlane‐Dick, D. (2006) · Studies in Higher Education · doi:10.1080/03075070600572090 · Cited by 5,846
Artificial Intelligence in Education: A Review
Chen, L., Chen, P., & Lin, Z. (2020) · IEEE Access · doi:10.1109/access.2020.2988510 · Cited by 4,133 · cited 4×
“…and several scholars have argued that the central question is not whether to use AI but which tasks are appropriate and under what conditions (Chen et al., 2020; Crompton & Burke, 2023 The feedback literacy literature (Carless & Boud, 2018; Boud & Dawson, 2023) has established that effective…”
InfluentialThe development of student feedback literacy: Enabling uptake of feedback
Carless, D., & Boud, D. (2018) · Assessment & Evaluation in Higher Education · doi:10.1080/02602938.2018.1463354 · Cited by 2,428 · cited 5×
“…to use AI but which tasks are appropriate and under what conditions (Chen et al., 2020; Crompton & Burke, 2023 The feedback literacy literature (Carless & Boud, 2018; Boud & Dawson, 2023) has established that effective feedback provision requires more than accurate information delivery: it…”
InfluentialArtificial intelligence in higher education: the state of the field
Crompton, H., & Burke, D. (2023) · International Journal of Educational Technology in Higher Education · doi:10.1186/s41239-023-00392-8 · Cited by 2,010 · cited 3×
“…have argued that the central question is not whether to use AI but which tasks are appropriate and under what conditions (Chen et al., 2020; Crompton & Burke, 2023 The feedback literacy literature (Carless & Boud, 2018; Boud & Dawson, 2023) has established that effective feedback provision…”
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 5×
“Boud and Molloy (2013) distinguish two feedback paradigms: Feedback Mark 1 treats feedback as information transmission, i.e., teachers pointing out errors and providing solutions.”
InfluentialMaking Sense of Assessment Feedback in Higher Education
Evans, C. (2013) · Review of Educational Research · doi:10.3102/0034654312474350 · Cited by 1,143
The role of laboratory work in university chemistry
Reid, N., & Shah, I. (2007) · Chem. Educ. Res. Pract. · doi:10.1039/b5rp90026c · Cited by 471
Formative assessment and feedback for learning in higher education: A systematic review
Morris, R., Perry, T., & Wardle, L. (2021) · Review of Education · doi:10.1002/rev3.3292 · Cited by 358
Research works printed but not linked: 2
- Gruenhagen, 2024, The rapid rise of generative AI and its implications for academic integrity: Students’ perceptions and use of chatbots for assistance with assessments, Computers and Education: Artificial Intelligence, 7, 100273
- Sahoo, D., Majhi, V., Neekhra, A., Sinha, Y., Mandal, M., and Kumar, D. (2025). How to trick your AI TA: A systematic study of academic jailbreaking in LLM code evaluation. arXiv.
Metadata
- Authors
- Venetsanos, D. T.
- Affiliations
- Imperial College London (GB)
- Corresponding author
- Venetsanos, Demetrios T.
- Venue
- AI in Education
- Publisher
- MDPI AG
- Published
- 2026-08-01
- Volume
- 2
- Issue
- 3
- Pages
- 26
- ISSN
- 3042-8130 (ISSN-L)
- OpenAlex
- W7172316169
- Cited by
- 0
- References
- 28
- FWCI
- 0.00
- Type
- journal article
- Language
- en
- Open access
- diamond
- doi.org
- Licence
- Locations
- AI in Education (publishedVersion, cc-by)
- Funders
- Imperial College London (G74038)
- Topics
- Social Sciences › Education › Student Assessment and Feedback
- Social Sciences › Education › Artificial Intelligence in Education
- Physical Sciences › Computer Science › Computer Science Applications › Online Learning and Analytics
- Keywords
- AI-assisted assessment, assessment feedback, written reports, human-in-the-loop, Hattie and Timperley feedback model, human oversight, academic integrity, feedback literacy, student agency, assessment fairness, assessment security, higher education
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 2026-10-11, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-11 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (qwen3.8-omni-flash).
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: A Tripartite Feedback Framework for AI-Assisted Assessment of Complex Reports in Higher Education. https://app.cochat.ai/papers/10.3390/aieduc2030026
- The paper
- Venetsanos, D. T. (2026). A Tripartite Feedback Framework for AI-Assisted Assessment of Complex Reports in Higher Education. AI in Education. https://doi.org/10.3390/aieduc2030026
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.