Research Article
AI-integrated programmatic assessment: a proposed framework for safeguarding cognitive development in AI-augmented learning environments
Synthesis: Ahmed and Shehata argue that the integrity-and-detection response to generative AI answers the wrong question. What a submitted essay no longer establishes is that the student can do the thinking, and the risk they name is developmental: repeated delegation of retrieval, elaboration, inference and error correction may erode the capacity those activities build, or prevent a novice from ever building it. Their proposed framework, AI-integrated programmatic assessment (AIPA), keeps the longitudinal, feedback-first, triangulated commitments of programmatic assessment and of assessment practice more broadly, and adds a formal treatment of AI use, built on three evidence streams — unaided cognition, AI-mediated performance, and orchestration quality — sampled through matched task pairs and read as developmental signals rather than compliance flags.
Key Findings
- The inference from product to competence has broken. When a fluent essay can be produced with little independent effort, assessment can no longer treat the submitted artifact as evidence of what the student can do.
- The named risk is developmental, not just procedural. Cognitive erosion means the gradual degradation of independent generative cognition — retrieval, elaboration, inference, error-based revision — when those activities are repeatedly delegated.
- Erosion and non-development look identical in a snapshot. Both produce a learner who performs better with AI than without it; only a documented earlier level of independent capability distinguishes decline from never having been built.
- AIPA adds three evidence streams. Unaided Cognition Evidence, AI-Mediated Performance Evidence and Orchestration Quality Evidence are interpreted together, so no single domain decides a high-stakes judgment.
- Matched task pairs make the comparison fair. Two tasks targeting the same construct at the same difficulty, one completed without AI and one with it, with a short structured viva attached to each.
- Orchestration is treated as a meta-competence. The framework defines it as selecting, governing and integrating AI contributions while retaining epistemic responsibility, and expects a developmental arc from uncritical reliance to accountable use.
- Credit moves from polish to reasoning. Because judgments accrue to demonstrated reasoning rather than output quality, students have a reason to show their thinking rather than minimize it.
The architecture in practice
The model is layered. An expanded evidence base feeds a judgment system of feedback cycles, triangulation, assessor moderation, discrepancy interpretation and competence committee review, with bounded AI-assisted analytics supporting but not deciding. A meta-competence strand runs across the layers, tracking five orchestration capacities from task–tool judgment to epistemic responsibility, and judgments feed back to learners as targeted sampling rather than as a final score.
The developmental arc runs from independence through hybrid performance to accountable orchestration, with the emphasis early in a program on building retrieval, explanation, problem formulation and reasoned justification, and a later shift toward selective, governed AI use. Worked examples in the paper pair a clinical case analysis written in 90 supervised minutes on a locked-down device with a parallel case written with access to an institutional AI tool, each followed by a structured viva; an engineering analogue pairs two design-scoping briefs of equivalent complexity.
The framework is explicit that orchestration must be taught before it is assessed, and that discrepancies between streams are read as developmental signals. A learner whose AI-mediated work far outperforms unaided work is not a cheat to be caught; that pattern says something about where practice is missing.
What the framework changes for institutions
The paper's governance section keeps human judgment central as a design principle rather than a residual safeguard, and treats data governance, privacy and equity as conditions of the model rather than afterthoughts. It also names faculty workload and institutional culture as feasibility constraints, which is where most assessment reforms fail: three evidence streams, matched task pairs and vivas multiply assessor time.
The research agenda follows from the framework's own logic. If erosion is a claim about change, it needs longitudinal designs with a documented baseline; if orchestration is a meta-competence, it needs instruments that can detect growth. Neither exists yet, and the paper presents AIPA as a testable proposal rather than an established effect.
For a program already running programmatic assessment, the practical claim is that generative AI does not require a new assessment philosophy. It requires one more evidence stream, a matched-task design, and a judgment rule that treats mismatch between streams as information.
What this means for practice
- Assessment designers: stop buying detection. Add an unaided task and an AI-available task on the same construct, at the same difficulty, and read the difference rather than either score alone.
- Instructors: attach a short oral defense to AI-permitted work. The viva is what separates orchestration quality from output quality and it is the cheapest version of the third evidence stream.
- Program directors: teach orchestration before assessing it. A framework that grades AI-mediated judgment without a curriculum that built it penalizes students for a capability nobody taught.
- Administrators: budget assessor time honestly. Three evidence streams and matched tasks are a staffing decision, not a software purchase.
- Researchers: treat cognitive erosion as a hypothesis requiring a documented baseline; the cross-sectional evidence the paper reviews cannot separate decline from non-development.
Limitations
- This is a conceptual analysis: no data are collected, and every claim about erosion or orchestration growth is a proposal awaiting empirical test.
- The erosion construct requires a documented earlier level of independent capability under comparable conditions, which no study in the reviewed literature supplies.
- Three evidence streams and matched task pairs multiply assessment workload, and the paper names faculty capacity as a feasibility constraint without estimating the cost.
- The worked cases are illustrative designs rather than evaluated implementations, so the model's reliability, fairness and defensibility are untested.
- The worked cases and much of the assessment literature it builds on come from medical and health-professions education, where programmatic assessment is established; transfer to undergraduate general education is asserted rather than argued.
Citation
Ahmed, S. A., & Shehata, M. H. (2026). AI-integrated programmatic assessment: A proposed framework for safeguarding cognitive development in AI-augmented learning environments. Frontiers in Education, 11.