Human involvement map · Research and analysis
When a student challenges an AI mark
Follow an objection through reasons, authority, correction and the work of responding.
Public research edition · 15 November 2026 · Human by choice — a Policy Window advocacy programme
A student disputes an assessment. A new mark arrives. What would make that a fair response? The answer depends on how the objection was heard, why the decision changed or stood, and whether the person answering had the authority and time to act.
HBC’s starting question is which human role is worth protecting. In assessment, that might be an accountable person able to examine an objection, explain a judgment and correct an error. Requiring a signature or insisting that every preparatory task be done by hand would need its own justification. Human reviewers can also be mistaken or overburdened.
What the studies actually measured
A request for review, a changed mark and a justified decision are three different things. These studies examine different steps; their figures cannot rank appeal systems across settings.
Real courses; randomized grading assignment
Requests for a human regrade
- Method
- In four political science classes in 2023–24, 271 students produced 3,080 short answers. Answers were assigned to human or AI-assisted grading; students could request a human regrade, which could lower the mark.
- Reported findings
- Human-graded answers drew more regrade requests than AI-assisted answers.
- Limits
- Requests count what students challenged, not all errors. The risk of a lower mark may discourage requests. This comparison does not establish which appeal procedure is best.
- HBC question
- How would an institution identify consequential errors that students did not challenge?
One programming course; observational preprint
A changed mark after an objection
- Method
- In the programming-course study, students’ objections, answers, marking criteria and earlier evaluations were sent to an AI language model. Its replies were shared with students.
- Reported findings
- The paper reports 137 grade changes among 185 appeals (74%). These are appeal records, not 185 different students.
- Limits
- A changed grade is not an independently verified correction. The reported process does not establish independent human adjudication of every dispute or the correctness of every revision.
- HBC question
- Who can check a reconsidered answer and change it again if the new reasoning is wrong?
Prototype; model-simulated challenges
Testing whether a model should change its answer
- Method
- The prototype tested 500 essays and one round of challenges, with ChatGPT playing the person disputing a grade. Four human raters supplied reference grades and later rated the generated feedback.
- Reported findings
- Using the raters’ grades as the benchmark, it reported better accuracy after a challenge than three comparison models. Tests distinguished keeping a correct grade from correcting a wrong one.
- Limits
- Simulated challenges and coder ratings are not observed student appeals or examination-board decisions. The study does not establish classroom effects or a resourced human response.
- HBC question
- Can a proposed process accept a well-founded objection while resisting an unfounded one?
A fair response can uphold a mark
A useful evaluation should follow the objection as well as the score. Record what the student questioned, what information the reviewer considered, the reason for changing or upholding the decision, and whether the result was communicated and implemented. A reasoned decision to keep a correct mark can be a good outcome. An increase obtained through poor reasoning can be a bad one.
This is a proposed evidence requirement, not a claim that HBC has audited such cases. Define how correctness will be judged, allow uncertainty and disagreement, and check a sample of decisions that were never challenged. Otherwise an institution could report quick closures while overlooking students who cannot identify an error or make their case.
Compare workable arrangements
For one kind of assessment, hold initial marking, the marking criteria and the stakes fixed, then compare the response routes below. Ask who answers first, who can change the mark, when someone else takes over, and what time and support each route needs. These are research options within existing assessment rules, not ranked recommendations or new rights. A practice comment and a qualification decision may need different processes.
The earlier choice of marking with or without AI needs a separate comparison. Include feasible teacher-led assessment without AI as an alternative for the whole process; do not assume unlimited staff attention. Specify all differences if comparing complete packages rather than isolating a response route.
Review by the original assessor
- Who can act
- The person responsible for the assessment examines the objection, can change the mark and gives reasons.
- Work to count
- Time to revisit the work, check the grounds, explain the result and put any change into effect.
- Risk to examine
- The reviewer may repeat or defend an earlier mistake. A human response is not automatically an independent one.
Review by a second qualified assessor
- Who can act
- Another assessor can examine the work and uphold or change the result under the institution’s rules.
- Work to count
- A second assessor’s paid capacity, access to the work and criteria, and time for reasons and communication.
- Risk to examine
- Extra work may cause delay. Shared assumptions or an unclear remit can reproduce the original problem.
Automated reconsideration with defined human escalation
- Who can act
- A tool responds first; a specified person can take over a consequential or unresolved dispute and change the result.
- Work to count
- Automated response plus escalation capacity, independent checks and help making an objection.
- Risk to examine
- The trigger may miss students who cannot articulate a problem. Repeated automated replies may postpone an effective response.
Pay for the whole response
Count staff time for preparation, checking, reconsideration, explanations, further escalation and unresolved cases. Include the student’s effort and the time spent waiting. Specify who pays, what work is displaced and how people can obtain assistance. A teacher who is accountable for a decision needs access to the work, relevant competence and practical power to reject or revise the tool’s contribution.
Apply HBC’s existing allocation question: who controls resources and who receives the benefit? A proposed pilot could log paid response capacity, student waiting time and spending decisions separately. Name the body deciding how any verified saving is used. Test whether reviewers receive paid time or must absorb extra work. Faster generation alone cannot settle that allocation.
Access matters too. A universal manual check could consume time needed for teaching or delay urgent responses. A process available only to confident complainants could miss others. Compare funded human escalation, targeted independent checks and accessible ways to explain an objection, rather than assuming that either a person or a tool automatically provides a fair hearing.
Turn this into an institutional decision
Start with one assessment and one kind of objection. Ask who can change the mark, what evidence they need, when the student will receive reasons, and how the decision will be put into effect. Compare response routes while holding initial marking fixed, counting staff time, student effort and the cost of mistakes or delay. Check existing duties and appeal routes first.
Human by choice could develop this into a short design brief, examining staff and student interests separately. This desk comparison claims no participating institution, student consultation or agreed mandate. A compulsory human-review proposal would need evidence of benefits, costs and a workable route to adoption.
What would change the case for protection?
A resourced human role becomes more compelling if it corrects consequential errors, gives usable reasons or secures a form of participation people value. Change the design if human sign-off is routine, adds delay without useful gains or pushes work onto unsupported staff. Better automated reconsideration could weaken a particular checking requirement while leaving a separate case for accountable human judgment in consequential disputes.
Keep the three aims separate: meaningful agency, income and economic power, and access to human services. Success on one does not establish success on the others. These studies inform the questions; they do not settle which arrangement an institution should adopt.
Sources and limits
Bounded primary-paper comparison, inspected 15 November 2026. No systematic review, original consultation, private-data reanalysis or evaluation of French guidance. Named preprint versions retain their own limits.
- Source 1
Heinrich et al. — AI-assisted grading and personalized feedback in large political science classes ↗
- Locator
- Research design; Table 1; Regrade requests, PDF pp.6–7
- Edition
- PLOS ONE 20(8), e0328041, 19 August 2025
- Source 2
Aytutuldu et al. — Integrating LLMs for Grading and Appeal Resolution in Computer Science Education ↗
- Locator
- Workflow and methods, pp.4–6; appeal results, p.9, Table 7
- Edition
- arXiv v1, 18 April 2025
- Source 3
Hong et al. — “My Grade is Wrong!”: A Contestable AI Framework for Interactive Feedback in Evaluating Student Essays ↗
- Locator
- Evaluation Metrics; Experiment Results; Human Evaluation Result
- Edition
- arXiv v1, 11 September 2024