Trade-offs in assessment and feedback design: a trilemma-based model

Synthesis: Pedlow, Fischer and Williams (2026) argue that the problem in assessment and feedback design is not a shortage of principles but the absence of a way to reason about competing ones. Their Dual Trilemma Framework recasts design as the deliberate making of trade-offs: authenticity, observability and feasibility in assessment, and usability, sources and sustainability in feedback. Generative AI is the occasion rather than the subject — it intensified longstanding tensions around authorship, the evidentiary value of student-produced work and the visibility of student thinking. The proposed response is not to pick poles but to distribute them across a coherent assessment system over time, so that different activities make complementary contributions.

Key Findings

1. The framework is a meta-lens, explicitly not a replacement. The authors position it as complementing rather than supplanting existing assessment and feedback frameworks: those say what good practice looks like, while this one addresses how to reason when those qualities compete. It is also deliberately not prescriptive — it does not claim any balance between priorities is optimal.
2. Two trilemmas, six design priorities. The assessment trilemma names authenticity, observability and feasibility; the feedback trilemma names usability, sources and sustainability. The decision to stop at three is justified as conceptual clarity and practical utility: a framework accommodating every desirable characteristic "risks becoming another comprehensive inventory of good practice".
3. Observability is the framework's genuinely new term. The authors distinguish it from surveillance, centering it instead on evidence: what does this assessment actually generate, and what claims can reasonably be made from it? Where opportunities to observe reasoning, judgment or decision-making are limited, claims about learning become harder to sustain regardless of whether AI was used.
4. Feasibility is argued up from an implementation detail to an educational priority. It covers assessment volume, moderation, timeliness, technological infrastructure, institutional policy and educator capability — and is placed alongside authenticity and observability rather than treated as the constraint that erodes them.
5. Feedback's third pole is the one most often assumed rather than designed. Usability is defined beyond clarity or readability as learners recognizing feedback as valuable, making sense of it against standards, managing the emotional response, and translating it into action; sources extends beyond educator comments to peers, self-assessment, exemplars, workplace supervisors, tasks themselves and AI-enabled technologies; sustainability requires that feedback not depend on individual academic labor.
6. Position comes from design, not from method. Tables 2 and 3 work through common approaches — written examinations, reports, vivas, individual educator comments, peer feedback, automated or AI-enabled feedback — and the framework's claim is that no method carries a fixed position. The same approach privileges different dimensions depending on how it is structured and what evidence it generates.
7. The unit of design is the system, not the task. Because no single activity can maximize all six priorities, the authors argue that trade-offs should be distributed across a sequence, with early formative activities carrying exploration and feedback while later points build observability and draw on diverse feedback sources.

Why a trilemma, and why only three

The term is borrowed from economics, where a trilemma describes three desirable objectives that cannot be simultaneously maximized (the authors cite Obstfeld, Shambaugh and Taylor's 2005 application). The authors import the logic, not the determinism: they state plainly that assessment and feedback rarely involve absolute constraints, that the dimensions exist in dynamic tension, and that strengthening one shifts what is achievable elsewhere depending on discipline, cohort and institutional conditions. The three-dimension limit is a design choice against comprehensiveness — an inventory of every desirable property is exactly what the field already has, and what leaves educators negotiating priorities implicitly.

What each pole asks of a design

Authenticity is defined broadly, as the extent to which assessment invites meaningful problems, professional or disciplinary judgment and capabilities beyond reproducing decontextualised knowledge — explicitly not as workplace replication, and not as employability alone. It is what stops assessment collapsing into compliance and surveillance, and the authors note the GenAI irony that security-focused responses can privilege controlled formats at the expense of meaningful learning. Observability addresses the defensibility of claims about learning; richer evidence through orals, live demonstrations or iterative checkpoints may raise workload or constrain authentic practice. Feasibility acknowledges finite time, resources, expertise and institutional capacity, and the page's framing of a "highly feasible" design simplifying complex capabilities is the sharpest statement in the paper of what is lost when feasibility alone governs.

Designing across time

The framework's practical payoff is temporal. Assessment and feedback are described as increasingly entangled: assessment generates evidence, feedback determines whether it contributes to future learning. Different activities are then allowed to privilege different dimensions — early work on authenticity and usability, later points on observability, final progression decisions integrating accumulated evidence rather than any single performance. The authors link this to programmatic assessment's aggregation of multiple observations, and to the layered-safeguards logic of the Swiss Cheese Model (Reason 1990; applied to academic integrity by Rundle, Curtis and Clare 2019, and to assessment validity by Dawson et al. 2024): confidence comes from the coherence of the whole rather than the perfection of any component. They also sketch how synchronous and asynchronous modes can be allocated — early guidance, rubrics and exemplars before assessment, scalable asynchronous self- and peer-assessment, and selective synchronous encounters where dialogue or responsive judgment adds value.

Where GenAI sits in the argument

GenAI is treated less as a new problem than as an accelerant that makes existing vulnerabilities hard to ignore: product-based evidence, individual authorship and retrospective detection of misconduct. Seen through the framework, AI-enabled practices can support authenticity, broaden feedback sources and create scalable opportunities for learner engagement, while simultaneously reshaping observability, feedback usability and sustainability. The authors' conclusion is that these are interconnected design considerations distributed across a system, not a property of the tool.

What this means for practice

• Instructors. Name the trade-off you are making instead of defending the task. If a take-home report bought authenticity and feasibility, say so and plan where the reasoning behind the product will actually become visible — an oral defense, a checkpoint, a revision log — rather than leaving that evidence unexamined.
• Assessment and curriculum designers. Use the framework to audit a sequence, not a task. Ask which dimension each activity carries, whether observability accumulates toward the points where high-stakes claims are made, and whether feedback sources are redundant or genuinely complementary before adding another one.
• Faculty developers. The framework is most useful as a deliberation tool: teams can place an existing design on both trilemmas, disagree about where it sits, and treat that disagreement as the training content rather than resolving it for them.
• Institutions. Feasibility claims should be budgeted as educational decisions. Workload, moderation volume, timeliness and educator capability determine what designs survive, so resourcing decisions are design decisions whether or not they are made explicitly.

Limitations

• This is a conceptual paper: the framework is derived from the literature and has no empirical validation. The authors state this and propose studies of how educators actually use it, whether it supports explicit deliberation about trade-offs, and how its dimensions interact across contexts.
• The framework ranks nothing and calibrates nothing. It gives no guidance for weighing one pole against another, no thresholds and no measurement approach, so two teams using it can reach different and equally defensible designs — which the authors treat as a feature, but it leaves the framework unable to adjudicate a contested decision.
• Accessibility and learner variability are named as unresolved: inclusion is identified as an important area needing further investigation, not as a dimension the trilemmas currently represent.
• The illustrative tables are reasoned examples rather than evidence. No study tests whether the predicted trade-offs hold, so the claim that a given approach shifts position under redesign is an argument, not a finding.
• The framework's own provenance is a single institution, and the authors disclose that generative AI (ChatGPT 5.6 Sol) was used to support language refinement, with intellectual content attributed to the authors.

Citation

Pedlow, M., Fischer, J., & Williams, A. (2026). Trade-offs in assessment and feedback design: a trilemma-based model. Assessment & Evaluation in Higher Education, 1–19.