On this page

Synthesis: Wood (2026) argues that the assessment problem created by generative AI is not detection but validity: a polished product is compatible with several different production histories, so a grade can no longer show what the learner can do or whether producing the work developed the understanding the qualification certifies. His response, learning-consequential assessment design, lets students use AI to produce assessed work while making success depend on demonstrating that the understanding the work displays is their own. The proposed mechanism is an incentive effect rather than a claim that AI causes learning: if students expect that explaining consequential decisions will be required, using AI for explanation, comparison and challenge becomes the rational route. Wood specifies three distinct capability claims an assessment can warrant — independent capability, learner-directed production and accountable verification — and treats the coherence between a learner's stated reasons and their work as the core evidential test.

Key Findings

  1. The product no longer identifies the capability. An essay, a translation or a design can result from unaided analysis, skilled coordination of AI, or heavy AI shaping, and the artifact alone does not distinguish them.
  2. Making AI use visible is not enough. Disclosure, process logs and reflections record that an activity happened; they do not establish what the learner understands, and process records can themselves be produced with AI.
  3. The mechanism is an incentive effect, not an instructional one. Knowing that understanding will be consequential is what is supposed to shift AI use toward learning, and the supporting evidence is indirect.
  4. Assessment can warrant three separate claims. Independent capability, learner-directed production and accountable verification concern different aspects of attribution, so evidence for one does not establish another.
  5. Coherence is the evidential test, with corroboration as a normal feature. The learner's account of consequential decisions must show the understanding those decisions required, and independent application should be available by design rather than triggered by suspicion.
  6. Only limited claims are defensible. A single assessment cannot show when learning happened or attribute development to AI-supported activity, so the defensible claim is that the required understanding can reasonably be attributed to the learner.

Three capability claims, not one

The distinction between claims is the part most likely to be collapsed in practice. Independent capability means the learner can perform the assessed reasoning without AI doing it. Learner-directed production means the learner made the consequential decisions that shaped the AI-supported work, rather than only prompting and managing output. Accountable verification means the learner checked the work and takes responsibility for it. Each requires different evidence, and each can be present without the others: a learner may direct a project without verifying its claims, or verify output whose direction was set elsewhere.

Wood draws on Salomon, Perkins and Globerson's distinction between effects with technology, visible in performance during use, and effects of technology, the cognitive residue that remains afterwards. That distinction is what forbids inferring independent capability from a successful AI-supported performance, and it is why the framework asks what a grade is evidence for rather than only whether AI was allowed.

Learning orchestration and what evidence to collect

The capability side of the framework is Learning Orchestration: the cognitive-metacognitive, epistemic and ethical judgment through which learners coordinate AI and other resources. The cognitive-metacognitive domain covers decisions about what to delegate and what to keep, including the argument that some difficulty is productive because generation and retrieval build understanding. The epistemic domain covers calibrated judgment about AI claims, and the ethical domain covers disclosure and responsibility for the final work.

Assessment evidence can be elicited without recording an entire production process. Short narrated screencasts, targeted questions linked to selected artifacts and brief accounts of specific decisions let a marker judge whether the learner's reasoning coheres with the work. Where an account is generic or weakly tied to the artifact, further questions follow. Wood is explicit that this is not a detection regime: the questions exist to test whether understanding can be attributed to the learner, and a fabricated account must survive questions about the drafts, sources and exchanges it refers to, which raises the cost of fabrication without eliminating it.

The incentive argument rests on washback, and the evidence for it is thin in a way the author acknowledges. A cross-sectional study of 742 undergraduates (Zhang, Lu and Kanji, 2026) found learning-oriented AI use more strongly associated with student agency where assessment was perceived as supporting feedback, reflection and improvement. That cannot show that changing assessment demands changes AI use, and Wood notes that learners may optimize the product and prepare narrowly for a later performance instead.

What this means for practice

  • Assessment designers. Decide which of the three capability claims each assessment is meant to warrant, then design the evidence that claim needs rather than adding a generic AI-use statement.
  • Instructors. Ask for short accounts of a few consequential decisions tied to the actual drafts and AI exchanges, not a full process narrative that is impractical to assess and easy to fabricate.
  • Instructors. Make independent application a routine part of the design — an unseen case, a re-explanation, a judgment of a new example — so corroboration is normal rather than an accusation.
  • Administrators. Shift policy questions from how AI use is detected to what evidence shows the required understanding can be attributed to the learner.

Limitations

  • The incentive effect is not directly demonstrated. Wood describes the supporting evidence as indirect and cites a cross-sectional study of 742 undergraduates that cannot show that changing assessment changes AI use.
  • Washback is neither automatic nor uniform. Learners may optimize the product and prepare narrowly for a later performance rather than engage deeply with the supported task.
  • The framework is a proposal, not an evaluated design. No implementation study, reliability evidence or comparison with conventional assessment is reported.
  • Feasibility at scale is unaddressed. Reviewing decision accounts and follow-up questioning adds assessor workload, and the article does not test a proportionate model.

Citation

Wood, J. (2026). Learning-consequential assessment: making learning count when students use generative AI. EdArXiv preprint. https://doi.org/10.35542/osf.io/jt5uf_v2.

Embed this page

Copy the code below to embed this page in a learning management system or other website. The embedded version shows only the page content — it hides the site header, navigation, and footer.