Assessment practice · Further and higher education
Designing AI-resilient assignments: rethinking assessment in the age of generative AI
You are marking two assignments. Both are well structured, convincingly argued and meet the criteria. The finished documents may look equally accomplished. The learning behind them could be very different.
Watch the webinar
Learner A
Spent two weeks comparing sources, testing ideas and revising their conclusions.
Learner B
Entered your brief into an AI tool, edited the response and submitted it.
For teachers, lecturers and assessors, this creates a pressing question: how can we judge what a learner knows, understands and can do when AI can produce a convincing submission?
That question shaped TeacherMatic’s webinar, Designing Resilient Assignments: Rethinking assessment in the age of generative AI. Esam Baboukhan explored the educational challenge before Richard Goddard demonstrated how TeacherMatic’s redesigned Assignment Generator can help educators review and strengthen their assignments.
What does a polished assignment tell us?
Assessment involves drawing conclusions from evidence. A report, practical demonstration or professional discussion helps us judge whether a learner has achieved an intended outcome.
Generative AI puts particular pressure on assignments where the finished document carries almost all that evidential weight. A fluent explanation may demonstrate understanding, but it may also conceal how little thinking the learner undertook.
This is a question of assessment validity: does the evidence support the judgement we want to make?
If we want to assess critical thinking, we need evidence of how a learner evaluates information and justifies conclusions. If we want to assess practical competence, an account of a procedure may need to sit alongside evidence of the learner performing it.
Authentic tasks can help by giving learners meaningful problems to address. Yet a realistic workplace scenario does not automatically establish individual understanding. AI can produce plausible business reports and emergency plans too. The task needs to require learners to engage with specific evidence and explain their decisions.
Academic integrity adds another consideration: learners must follow the rules, acknowledge assistance appropriately and submit work that meets the requirements for independent achievement.
These questions are connected, but distinct. A learner may disclose permitted AI assistance and still provide insufficient evidence of an outcome. Equally, a realistic task may remain vulnerable if a generic response can satisfy its criteria.
Clear expectations about AI are essential
AI already forms part of many students’ study routines. The HEPI and Kortext Student Generative AI Survey 2026 found that 94% of respondents used generative AI to help with assessed work.
of surveyed UK undergraduates used generative AI to help with assessed work.
Assistance with assessment is not a cheating rate. Survey of 1,054 full-time UK undergraduates.A learner might ask AI to explain an unfamiliar concept, suggest practice questions or provide feedback on a draft. Whether a particular use is acceptable depends on the assessment and its rules.
A general instruction to “use AI responsibly” leaves too much open to interpretation. Learners need to know:
- What AI may help them do.
- Which activities they must complete independently.
- How they should acknowledge assistance and retain any required records.
Institutions are developing ways to communicate these boundaries. UCL’s framework distinguishes between assessments where AI is prohibited, assistive or integral to the task. Its guidance makes clear that the specific assessment instructions take precedence. Explore UCL’s assessment categories.
In FE, Jisc reported that Coleg Sir Gâr and Coleg Ceredigion adapted an AI Assessment Scale and introduced it with examples and guidance. Early pilot feedback suggested that over 90% of learners found it helped them understand appropriate AI use. That is encouraging feedback about clarity, rather than evidence that the approach eliminated misuse or improved attainment. Read Jisc’s account.
For qualification assessments, instructions must remain consistent with the relevant awarding organisation’s requirements. JCQ guidance emphasises that learners must demonstrate their own independent work. Acknowledging generated content does not, by itself, demonstrate independent achievement of the marking criteria. Read the JCQ guidance.
Detection cannot demonstrate understanding
AI detectors can make errors in both directions. A false positive flags human writing as generated; a false negative misses generated content.
Research findings also need careful interpretation. Weber-Wulff and colleagues found substantial reliability limitations in the systems they evaluated in 2023. Liang and colleagues identified unfair misclassification of non-native English writing in their study. A 2026 study in Czech found no systematic bias against non-native speakers across the detector families examined. Results depend on the tools, texts and conditions tested. See Weber-Wulff et al., Liang et al. and Al Ali et al.
A detector result cannot establish what a learner understands. An unflagged submission does not demonstrate learning, while a flagged submission requires careful investigation through the institution’s procedures.
Assessment design provides a more direct route to the educational question: give learners opportunities to demonstrate their reasoning, apply their knowledge and explain the choices they make.
Five principles for stronger assignments
The webinar proposed five principles for designing assignments that remain useful as AI capabilities develop.
01 Start with the learning outcome
Before changing the format, identify what learners must demonstrate.
If the outcome concerns evaluating evidence, what would successful evaluation look like? Learners might need to compare explanations, identify a limitation in a dataset or justify why one source deserves greater weight.
Check that the task elicits those actions and that the criteria reward them. A lengthy descriptive report may provide limited evidence of an evaluative skill.
Where practical performance is central, the evidence must allow an assessor to judge that performance. Written explanation can support the assessment, but cannot automatically replace demonstration.
02 Make the task specific and meaningful
Generic briefs invite generic responses. Supplied data, a defined client need or a particular workplace constraint gives learners something concrete to investigate.
The key is to make those details matter. Ask how the evidence affects the recommendation, which constraint limits the options or what would change under different conditions.
For example, a research methods assignment could ask students to justify a sampling decision for a supplied study scenario. A vocational task might require learners to explain how a particular workplace condition affects their proposed procedure.
Specificity supports application and judgement. It does not make a task immune to AI.
03 Make thinking visible as the work develops
A short outline, decision log or explanation of a revised recommendation can reveal thinking that disappears from a polished final submission.
These activities can also support formative assessment. An early checkpoint might ask learners to identify their strongest evidence and one unresolved question. Feedback can then help them improve their approach before submission.
Learners could briefly explain which feedback they acted on and why. This encourages them to evaluate advice and take responsibility for their decisions.
Process evidence still needs interpretation. AI can generate reflections and decision logs too. Their value increases when they connect to work the tutor has seen, a specific change or a focused follow-up question: “Which piece of evidence led you to revise this recommendation?”
The aim is to capture meaningful decisions through proportionate evidence.
04 Explain permitted AI use within the brief
AI guidance should relate directly to the task.
Where the rules allow it, learners might use AI to suggest initial research questions while taking responsibility for verifying sources and developing their analysis. Another assessment may prohibit AI during production because independent performance is central to the outcome.
Explain what learners must declare and what records they need to retain. Discuss examples before submission, while clarification can still influence their choices.
05 Preserve the standard and plan accessible methods
Stronger evidence must also be fair evidence.
A discussion can help an assessor explore understanding. However, a compulsory oral format may introduce barriers when speaking is not the competence being assessed.
The distinction between a competence standard and an assessment method matters. The standard defines the required level of achievement. The method is how the learner demonstrates it.
The webinar referred to University of Bristol v Abrahart, following the High Court’s 2024 decision. Natasha Abrahart was a physics student who experienced severe social anxiety. The case concerned the university’s failure to make reasonable adjustments to assessments, including oral assessments. The EHRC’s advice highlights the need to distinguish assessment methods from competence standards and plan adjustments in advance. Read the EHRC advice.
Where appropriate, a written explanation or another accessible response may allow an assessor to examine the same reasoning. Alternatives must still demonstrate the relevant competence and satisfy the applicable qualification requirements.
A Level 3 Business assignment, redesigned
The webinar illustrated these principles with a familiar brief:
“Write a 2,000-word report analysing the marketing strategies of a business.”
An assessor receives a finished report, but may have limited insight into how the learner developed its conclusions.
The redesigned brief
You are a marketing assistant at a local business.
Using the sales and social media data provided:
- Identify a problem shown in the evidence.
- Recommend a strategy and justify it using the data.
- Keep a short log explaining two decisions you changed and why.
- Explain your reasoning through a five-minute discussion, or an appropriate written or recorded alternative.
- Declare any permitted AI assistance with research and ideas.
The assessor can now examine how the recommendation follows from the evidence. A focused question can explore whether the learner understands a limitation or can explain an alternative.
The decision log also provides an opportunity to examine development. Did the learner revise an assumption because the evidence challenged it? Can they explain why the revised approach is stronger?
The intended outcomes and standard remain central. Any redesign would still need checking against the qualification’s assessment requirements.
Workload matters too. Replace some less informative work with evidence that serves a clearer purpose. A short log and focused explanation should not simply accumulate on top of every existing requirement.
How TeacherMatic’s Assignment Generator supports the process
Turning these principles into a coherent brief takes time. Richard Goddard, TeacherMatic’s AI Architect, demonstrated how the redesigned Assignment Generator can support both new assignment design and the review of existing briefs.
His example involved a Level 3 travel and tourism task: creating an airport emergency response plan. Richard uploaded a qualification specification, identified the relevant unit and learning outcome, and specified individual work. He requested a report and an additional summary handout for airport staff.
The workflow also offered choices about assessment purpose, submission format and a draft marking rubric. These structured inputs prompt educators to establish the assignment’s requirements before reviewing the result.
Supply the specification, unit and learning outcome.
Set the context, submission format and additional evidence.
Request analysis and suggestions to strengthen the brief.
Check the draft using your professional judgement.
Richard selected four options to support AI resilience:
- Highlight parts likely to invite generic AI responses.
- Suggest alternative assessment strategies.
- Incorporate features that strengthen the assignment.
- Include a statement about permitted AI use.
In his example, AI could assist with research but could not draft the submission text.
The generated material suggested location-specific airport details, a process log and opportunities for learners to justify decisions. It also highlighted broad explanations of emergency planning and general descriptions of roles as areas where generic responses could satisfy a loosely worded brief.
This gives educators material to evaluate alongside a draft assignment. They can consider where evidence may be weak and which suggested changes would make it more useful.
Richard also explained that educators can submit an existing assignment for redrafting. That offers a manageable starting point: review a brief already in use, identify weaknesses and consider targeted improvements.
During the webinar, participants described using the Assignment Generator to develop tasks that encourage deeper critical thinking and adapting its output for assignments in Microsoft Teams. These examples suggest ways educators can incorporate the tool into existing practice while retaining responsibility for assessment quality.
The potential workload benefit comes from bringing several design decisions into one structured workflow. The webinar did not measure time savings, but it demonstrated a route to a draft that educators can review and adapt.
The educator makes the final decision
Richard emphasised that TeacherMatic provides a first draft. Professional review remains essential.
Generated wording may introduce unnecessary complexity, conflict with the intended purpose or require correction. In the demonstration, Richard selected summative assessment, while part of the generated guidance later described a formative task. That mismatch illustrates why educators need to read the output closely.
A generated rubric also needs checking against the actual criteria and approved grading arrangements. Uploading a specification does not establish that every detail of the result complies with it.
The most useful review questions are educational ones. Does the task elicit the intended learning? Are the instructions clear? Is the evidence sufficient and accessible? Can the assessment be delivered and marked within the available time?
This preserves professional autonomy. TeacherMatic supports the drafting and review process; educators decide what is suitable for their learners and qualification.
Richard and Esam also invited feedback on the redesigned generator. That co-design approach gives educators a way to shape development around practical assessment needs.
Start with one assignment
You do not need to redesign every assessment at once. Choose one existing brief and ask:
- What must learners demonstrate? Check that the task and criteria address the intended outcome.
- Where will I see their thinking? Look for a meaningful decision, application or explanation.
- Could a small checkpoint help? Consider whether brief feedback during development would improve learning and evidence.
- Are the AI rules clear? Specify permissions, restrictions and acknowledgement requirements.
- Can learners demonstrate the standard fairly? Review accessibility alongside learner and staff workload.
- Does the redesign meet the requirements? Check the relevant qualification, awarding organisation and institutional rules.
Make the smallest changes that materially improve the evidence.
Return to the video, or try the Assignment Generator with one of your existing briefs. Review the suggestions and share your experience.
What will this assessment allow me to judge confidently about what the learner knows, understands and can do?
