The Short Answer

Assessment is being redesigned because generative AI has changed what students can produce, while marking has largely remained built around scripts, recall, similarity checks and uniform written answers. A student who can ask a model to draft an essay, solve a problem, generate code or explain a difficult concept can often submit work that resembles a conventional assignment without demonstrating that they learned the intended skill. Institutions have therefore begun changing tasks, assessment evidence and supervision, but marking systems have been slower to change because they must serve larger classes, remain consistent, protect academic integrity and produce usable feedback.

Also worth reading: How Can AI Extension Risk Assessment Protect Your Browser Workflows? · How Can an AI Readiness Assessment Checklist Guide AI-Driven Tutorials? · What Makes an Enterprise AI Skills Assessment Framework Effective for AI Adoption?

That does not mean every traditional assignment has become meaningless. Essays, examinations, multiple-choice tests and practical work still assess valuable knowledge when students must work under controlled conditions, explain their reasoning orally or produce evidence connected to a process they can defend. The deeper issue is alignment: an assessment should not ask students to demonstrate a capability while leaving a low-cost generative route to the same output. Redesignment therefore begins with identifying what each task is intended to prove, not with banning a particular tool.

As of 1 October 2026, there is no universal rule saying that AI must be permitted or forbidden in every assessment. The more defensible position is to specify permitted uses, require disclosure where relevant, assess the learner's own contribution and use varied evidence. Marking should become partly process-aware and partly judgment-based, with human review directed toward reasoning, authenticity and the criteria that matter most.

Why Assessment Changes Faster Than Marking

The pressure comes from the mismatch between rapidly changing student capabilities and slow institutional processes. Generative AI can produce a polished answer in seconds, revise it for a specified audience and imitate several academic styles. It can also explain mistakes, create alternative methods and solve structured problems. The result is that a submission may look correct while revealing little about how the student would perform unaided. Research reported by Times Higher Education and by engineering faculty cited in Frontiers describes a growing gap between assignments designed for earlier learning conditions and students' actual use of these systems.

Marking is difficult to automate responsibly because grades are not only measurements of output. They are institutional decisions involving academic judgement, appeals, accreditation, comparability and sometimes financial or professional consequences. A rubric may define quality reasonably well, but models can game surface features such as structure, vocabulary and formulaic balance without possessing the understanding a marker expects. Institutions also fear false accusations: a detector or similarity checker may flag a student's own writing, while an AI-generated answer may escape notice. These uncertainties make blanket detector-based enforcement especially weak.

There is also a labour problem. When an entire course moves online or enrolments expand, a marker may face several hundred submissions. Asking each student to provide drafts, process notes, an oral defence or a live demonstration can improve evidence, but it takes time. Institutions have not consistently supplied the staffing, training or software budget to redesign every task at once. This is why some programmes first introduce a small number of redesigned assessments, retain existing marking practices where they remain defensible and review outcomes after one or two teaching cycles.

What AI-Era Assessment Actually Measures

The strongest redesign is not an attempt to defeat AI with clever prompts. It is an attempt to assess a broader and more relevant set of capabilities. A useful AI-era task can ask students to verify an AI-generated answer, find errors, select reliable sources, redesign a flawed solution, compare competing approaches or justify a decision. The student still uses AI, but the grade depends on critical judgment rather than whether the tool can produce fluent prose.

Assessment evidence should consequently be distributed across more than one moment. A short initial attempt can establish what a student knows before assistance; a prepared artefact can show applied ability; and a live conversation, demonstration or reflection can reveal whether the student understands the result. This approach does not require every assessment to become an oral exam. It requires designers to state the purpose of each evidence source and make sure that the total workload is reasonable. For example, a 2,000-word essay might be replaced by a 1,200-word technical argument, an annotated source record, a five-minute defence and a short post-submission correction exercise.

The key word is alignment. If the learning outcome is independent problem-solving, the assessment must include an unaided component. If it is ethical use of AI, the student should be assessed on documentation, verification and disclosure. If it is programming, the student may be allowed to use an assistant but must pass hidden tests, explain the code and modify it after requirements change. Redesign fails when educators add several new tasks without removing old workload, because students then experience assessment inflation rather than better assessment.

FeatureTraditional marked submissionAI-era redesigned assessment
Main evidenceFinal essay, exam or codeFinal work plus drafts, checks, explanation or demonstration
AI positionOften unspecified or treated as cheatingDefined by task: allowed, limited, disclosed or prohibited
Marking emphasisSurface polish and correctnessReasoning, judgment, verification, application and authenticity
Moderation roleCompare final scores and resolve disagreementCheck rubric consistency and learner contribution
Main weaknessOutput may be difficult to attributeCan consume more staff time unless carefully simplified
Typical safeguardSimilarity or AI-detection softwareTask design, disclosure, oral checks and authentic context
## Practical Steps for Redesigning a Course

Start with an inventory of assessments and their intended learning outcomes. For each task, record whether it is meant to test recall, analysis, creation, communication, practical skill, ethics or independent performance. Markers should identify exactly where AI could replace the learning process or merely provide low-value scaffolding. An assignment that tests recall of definitions may be better converted into a short, controlled knowledge check; an assignment testing engineering design may require a defensible decision record and a modified design under exam-like conditions.

Next, create a clear AI-use policy for the course. State what students may use, what they must disclose and how the permitted use will affect marking. A practical threshold is to require a brief declaration whenever AI materially generated, edited or interpreted content, naming the tool, date, purpose and significant prompts or transformations. This threshold is more workable than requiring students to disclose every spelling correction or harmless paraphrase. Educators should also avoid pretending that a detector produces a reliable probability of authorship. Treat detector scores as a prompt for review, not as proof, and avoid disciplinary action based on one result.

Then redesign the marking workflow. Rubrics should distinguish factual accuracy, method, evidence, originality, explanation of uncertainty and the quality of the learner's contribution. Human markers should spend more time on the parts that require subject knowledge and less time on grammar that software or students can reasonably correct. One practical sample is to require students to submit an initial plan or solution, the final artefact and a 150-word reflection identifying one limitation and one verification step. This is not foolproof, but it produces useful evidence and keeps assessment proportionate.

Pilot before scaling. Run the redesigned task with one cohort, compare workload and student learning, and revise the rubric before adopting it across multiple sections. A useful review interval is one teaching cycle, such as a semester, rather than announcing a permanent system based on a single experiment. Collect student feedback, check whether the task is accessible to students using assistive technology, and ask whether an experienced marker can apply the criteria consistently. Assessment redesign should be an evidence-based process, not a technology purchase followed by compulsory policy language.

Alternatives and Different Teaching Contexts

Not every course should adopt the same model. In foundational mathematics, structured data analysis or introductory programming, controlled assessments can test whether students can perform a procedure without a tool. In writing, translation, design and open-ended research, forcing artificial restrictions may be less realistic than requiring source checks, process evidence and an oral explanation. In professional programmes, students may need practice using AI while also proving that they can supervise it and accept responsibility for the result.

ContextRecommended assessment approachWhy it fits
Large introductory lectureShort in-class evidence plus a low-stakes assignmentMakes independent understanding visible without creating a large oral-marking load
Laboratory or engineering coursePractical demonstration, hidden test cases and design defenceSeparates working code or equipment operation from an AI-generated explanation
Creative writingWorkshop, revision history and critical commentaryAssesses voice and deliberate revision rather than only a final manuscript
Professional or workplace simulationAI-supported case analysis with disclosure and judgement rubricRecognizes realistic tool use while testing responsible decisions
Small advanced seminarOral presentation, peer critique and reflective defenceSupports close supervision and deeper conversation
Open or distance educationAuthenticated tasks, staged submissions and periodic identity checksAddresses weaker supervision without assuming every learner cheats
These alternatives are not automatically superior. Oral assessments can disadvantage nervous students, students with certain disabilities or learners whose first language is not the language of instruction, although reasonable adjustments and suitable formats can reduce those effects. Authentic projects can privilege students with personal technology, time or prior professional access. A course team should therefore combine methods instead of selecting the most dramatic intervention, and should preserve an accessible route to demonstrate the same outcome.

The phrase “AI-era assessment redesign” should not imply that technology alone determines quality. A poorly designed project can be completed by copying, and a tightly timed exam can still measure memorisation. The relevant test is whether the evidence gives a fair account of the capability named in the learning outcome. That principle also applies to older technologies: calculators changed mathematics assessment, search engines changed research, and online collaboration changed group work. AI increases the need for explicit evidence, but it does not remove the need for educational design.

Common Mistakes and Cost Considerations

The most common mistake is equating AI detection with assessment redesign. Detection tools have limitations, including false positives, false negatives, language bias and disagreement between vendors. They may also encourage academic surveillance without improving learning. A second mistake is relying on strange wording or personal anecdotes as a supposed secret that AI cannot detect. Models and students adapt, and such tricks shift effort away from learning rather than producing a durable standard.

A third mistake is adding an oral defence to every assignment without removing the original essay. This raises workload and may create equity concerns. A fourth is stating a broad prohibition without explaining acceptable use: students cannot judge whether grammar assistance, code completion, literature searching or idea brainstorming is permitted. Institutions should publish course-level examples and revise them after each term. Excessive policy changes create confusion, and inconsistent enforcement damages trust more than a candid, well-explained rule.

Costs vary widely. Basic redesign can be free if instructors revise prompts, reuse existing learning-management-system functions and assign extra short explanations. A small pilot might require approximately 20 to 40 staff hours for rubric design, training, review and analysis, although the real figure depends on cohort size and discipline. Learning-management-system plug-ins, oral-assessment platforms or identity-verification services may cost from a few dollars per student per month to several hundred or several thousand dollars annually for an institution, depending on licensing and scale. Commercial AI detectors should not be treated as mandatory purchases without evidence that they improve decisions.

Budget time and professional development before buying software. Train markers in constructive feedback, sampling work, oral questioning and bias awareness. Give departments a shared set of exemplar submissions and model answers, then fund a small number of pilots. The return is not simply fewer suspected violations; it is better evidence of learning, more useful feedback and less time spent policing uncertain outputs. Institutions should also budget for accessibility testing and periodic review because new tools can reproduce old forms of exclusion.

When Institutions Should Act

Action is warranted when the cost of continuing an unchanged assessment is no longer acceptable. Warning signs include widespread confusion about AI rules, student reports that ordinary assignments are no longer meaningful, rubric scores that no longer distinguish understanding, and markers spending most of their time checking suspected authorship. A practical trigger is not a particular percentage of students using AI, because estimates vary by discipline and survey method. It is a demonstrated mismatch between an assessment outcome and the evidence currently being graded.

Begin at course level rather than imposing an institution-wide switch. The first phase can last one semester and cover one high-enrolment, high-risk assignment. Establish a baseline by recording the workload, score distribution, appeals, student perceptions and examples of weak or strong work. After redesign, repeat those measures and ask whether the new assessment better predicts later performance, such as laboratory work, clinical reasoning, design review or final-project quality. If it does not, revise the task again.

Leadership must act on equity and governance as well as technology. Students should know how their work is assessed, what evidence they must provide and what happens when a tool fails. Appeals routes should remain available, and staff should not be expected to investigate every suspected breach without support. Faculty should have authority to choose approaches appropriate to their subject, while programme-level teams ensure that comparable outcomes are assessed across sections. This balance is preferable to a central policy that appears simple but is difficult to apply in practice.

The date matters because generative-AI capability and student expectations are changing faster than many institutional policies. By October 2026, a course designed only around a presumed pre-AI student workflow deserves review. However, urgency should not become a reason to claim that every assessment must be replaced immediately. Review the highest-risk tasks, test alternatives, publish the evidence and scale only what demonstrably improves learning. Assessment is successful when students can do the intended work, markers can explain the grade, and institutions can defend the decision.

A Sustainable Marking Model

A sustainable model combines clear task design with human accountability. Before submission, students receive the rubric, AI expectations and examples of acceptable evidence. During production, they may receive ordinary academic support, including feedback on ideas and clarity. After submission, markers assess the intended capability, ask targeted questions when authorship or reasoning is unclear, and record a proportionate response. The student may be asked to correct the work, explain a choice, provide missing evidence or redo a section under controlled conditions.

This model does not require assuming that every polished submission is fraudulent. It does require recognising that suspicion is not the same as evidence. If a marker cannot explain why a judgement was made, the institution should not turn an uncertain detector score into a disciplinary conclusion. Conversely, if a student refuses to discuss a submission or cannot explain a supposedly independent result, that may justify review under the institution's ordinary academic-integrity procedure. The response should follow established rules and individual facts rather than public reaction to a new AI product.

The most useful long-term measure is not “AI use” as a binary category. It is the quality of learning evidence. Can the student solve a novel problem, justify a conclusion, recognise an error, communicate to a real audience and revise after feedback? These capabilities remain valuable even as the tools used to produce artefacts change. Institutions that make those outcomes explicit will find it easier to revise prompts, rubrics and marking procedures as models evolve. Those that do not will continue debating whether a detector caught the latest version of cheating.