Worked Assessment Designs for AI Use
These three designs connect a learning outcome, assignment rule, submission, rubric, and independent check. They are SAGE teaching examples, not validated assessments or reports of implementation. All local facts and figures below are fictional. Adapt the workload, disciplinary expectations, and scoring with your course team before use.
Start with assessment design and publish the relevant assignment AI rules. The levels below use SAGE’s planning labels.
The examples include educator notes and answer guidance. Prepare a separate student brief and a fresh independent task before teaching. Allow the usual approved accommodations in every part, including independent checks.
Example 1: an evidence-based argumentative essay
Original task: “Write an essay about whether universities should extend library opening hours.” This leaves the evidence and intended reasoning vague.
Outcome: weigh conflicting evidence and defend a proportionate recommendation.
Student brief and source pack
Write a 700-word recommendation to a fictional university using these supplied records. Cite their identifiers and distinguish an observation from an inference.
| Record | Fictional evidence supplied to all students |
|---|---|
| L1: attendance log | During a four-week pilot, the library recorded 120 entries after 22:00. Entries are visits, not distinct people. No comparison period was recorded. |
| L2: invitation survey | Of 80 respondents to an open invitation, 60 wanted later hours. The university has 4,000 students; respondents were not randomly sampled. |
| L3: staffing estimate | An eight-week extension would cost €8,000. The same budget would fund eight weeks of additional daytime study-space support. Neither option has a measured effect on learning. |
Recommend an extension, daytime support, or a further bounded pilot. Address the strongest objection to your choice and explain what additional evidence would change it. Several recommendations can be defensible.
AI-limited rule: you may request objections to an outline and feedback on clarity using this source pack in an approved tool. Do not use AI to write the submitted essay, invent evidence, or replace your evaluation. If you use AI, submit a short disclosure. Peer feedback using the same questions is available without an AI account.
Submit: the essay, a three-row claim-to-record table, and a 150-word note justifying one accepted and one rejected suggestion from AI or peer feedback. Suggested preparation time is three hours, including reading and revision.
Rubric and anchor responses
Score each criterion from 0 to 2, for a total of 6. Review the descriptors with students and calibrate using short anchor responses.
| Criterion | 0: not yet demonstrated | 1: partly demonstrated | 2: demonstrated |
|---|---|---|---|
| Evidence interpretation | Makes a material population or causal inference the evidence does not support | Uses records accurately but leaves an important limitation unexplained | Uses records accurately and explains how sampling or measurement limits the inference |
| Weighing alternatives | Ignores a competing option | Names an alternative without weighing its evidence | Compares a credible alternative and explains the trade-off affecting the choice |
| Justified conclusion | Recommendation lacks support in the pack | Recommendation is supported but its uncertainty is unclear | Recommendation follows from the comparison and names evidence that could change it |
Partial anchor: “Most students want later hours because 60 out of 80 said yes.” The fraction describes respondents, but the inference to most students is unsupported. Stronger anchor: “The respondents favour later hours, but the open invitation cannot establish university-wide demand. I recommend a bounded pilot that counts distinct users before committing recurring funds.” This is stronger evidence interpretation; a full essay must still weigh the alternative.
Independent check and workload
In a ten-minute supervised response, give students a fresh survey with a different sampling method. Ask which inference is warranted and how it affects a decision. Use the evidence-interpretation criterion; publish its contribution to the grade beforehand. Offer an equivalent accessible response format. Budget about twelve minutes per essay for marking, then pilot that estimate. A revision note supports reflection; it does not authenticate authorship.
Example 2: a quantitative analysis audit
Original task: “Calculate the overall conversion rate.” A correct number alone reveals little about denominator choice or checking.
Outcome: calculate an aggregate rate and explain why unequal group sizes matter. Use the complete synthetic dataset and intentionally flawed report in AI-assisted data analysis.
Student brief and permitted assistance
Audit the claim “the combined conversion rate is 50%”. Submit a one-page audit with the original claim, your formula and substituted values, a reproducible calculation, the corrected result, and a justified interpretation. Explain why the simple average does or does not answer the question.
AI-integrated rule: evaluate the supplied AI-style report. Live AI use is optional; the shared report is sufficient. If used, an approved assistant may suggest checks or explain an error after your first attempt. You must select the denominator and reproduce every calculation yourself. Do not upload real customer data. A calculator or accessible spreadsheet is available to everyone.
Submit: the audit and a short record of a consequential check, rather than a full chat. Allow 45 minutes of practice and preparation.
Rubric and answer guidance
| Criterion | 0 | 1 | 2 |
|---|---|---|---|
| Quantity and denominator | Does not identify the relevant quantity | Identifies purchases per visit but aggregates inconsistently | Uses total purchases divided by total visits and states the unit |
| Reproduction | Provides an unsupported result | Shows a method with an arithmetic or traceability gap | Reproduces 18 / 100 = 18% from the supplied rows |
| Interpretation | Accepts 50% without checking | Rejects 50% without explaining the weighting | Explains unequal group sizes and avoids a causal claim about channels |
Weak anchor: “AI says 50%, so that is the rate.” Partial anchor: “The answer is 18%,” with no working. Stronger anchor: “There are 18 purchases in 100 visits, so the pooled rate is 18%. Averaging 10% and 90% gives each channel equal weight even though their visit counts differ.” Score the actual evidence under each criterion, not these labels alone.
For an eight-minute independent check, supply two new groups with 20 and 80 visits and 4 and 8 purchases. The pooled rate is 12%; the unweighted mean is 15%. Ask students to explain the difference. Budget five minutes per audit for review; verify both the result and the reasoning. Use later practice if retention matters.
Example 3: a professional decision case
Original task: “Write a business proposal for improving study spaces.” This can reward presentation while leaving professional judgment unobserved.
Outcome: recommend a feasible action under budget and access constraints, and explain how new information changes the decision.
Student brief and case facts
You advise a fictional campus with €6,000 available for a six-week pilot. Option A provides 12 additional quiet desks for €5,000. Option B improves acoustic separation around 24 existing desks for €6,000. A facilities note confirms both are feasible and preserve accessible routes. There is no demand survey or measured noise comparison. Existing desks remain available under both options.
Prepare a 500-word recommendation with a cost check, comparison, one material uncertainty, and a feasible way to investigate it. Do not invent utilisation, student preferences, access needs, or an effect on learning.
AI-limited rule: AI may role-play a committee member challenging your draft using only these facts. It may not draft the submitted recommendation or decide which option is best. A peer may use the same challenge cards. Disclose AI use briefly when applicable.
Submit: the recommendation and two annotated decisions explaining what you changed or retained after challenge. Allow 90 minutes for preparation. During a five-minute structured follow-up, consider a fresh condition: A now costs €6,500. Explain how this affects feasibility without assuming extra funding is available.
Rubric and educator notes
| Criterion | 0 | 1 | 2 |
|---|---|---|---|
| Feasibility | Ignores budget or access constraints | Checks some constraints but leaves a material gap | Checks the proposed action against all supplied constraints |
| Decision quality | Relies on invented demand or unsupported benefits | Uses case facts but gives a limited comparison | Weighs both options and names uncertainty relevant to the choice |
| Response to change | Retains an infeasible recommendation without explanation | Recognises the new cost but gives an incomplete response | Revises or pauses the recommendation and explains what must be resolved |
Partial anchor: “B serves twice as many students.” B concerns twice as many desks; neither use nor distinct student reach is known. Stronger anchor: “B changes more existing desks, but occupancy and acoustic benefit are unknown. I would measure both before selecting the pilot.” For the follow-up, A exceeds the budget by €500. B remains within budget, but declining or revising the pilot can also be justified. There is no single preferred option to reward.
Publish the follow-up question types and scoring in advance. Provide an equivalent structured written or supported format where needed; assess the same reasoning rather than conversational fluency. Five minutes per student already requires five hours for a cohort of 60, before scheduling and marking. Pilot whether that burden is feasible or integrate the check into an existing session.
Adapt and evaluate the designs
Use the workflow evaluation record to inspect learning, review time, access, and failures. These examples draw on SAGE’s existing writing, rubric, and simulation guidance. They illustrate design decisions; they do not demonstrate that AI improves the resulting assessment.