Refrain, Then Amplify: A Curriculum Framework for Sequencing Generative AI to Form Professional Judgement

Synthesis: Torres-Sahli et al. (2026) propose a program-level curriculum framework — "refrain, then amplify" — for sequencing Generative AI to form professional judgment. On this design, a program withholds a generative tool while a capacity is forming, then restores it to amplify that capacity once the student can direct it, judge what it returns, and answer for it. Devices are allowed where they support engaged work and excluded where they drain attention, all governed by a forming-versus-offloading criterion: whether a stretch of work forms a capacity or merely passes it through the tool. The program fixes outcomes and integrity, reserves teaching method to the instructor, and places a hard-to-fake checkpoint at each refrain-to-amplify hinge.

The Underlying Question

Universities hold three separate conversations about the same question: one about student devices, another about generative AI in assessed work, and a third, unspoken, that leaves instructors to decide both at their own professional risk. The question underneath: does the technology serve the formation of a capacity, or do the work in the student's place? This paper holds the three as one, at the level where curricular design happens — the program.

The framework echoes an ancient concern: Socrates' complaint against writing in Plato's Phaedrus — that those who trust external marks will cease to remember for themselves. With Generative AI, the danger is no longer only that we lose capacities we stop exercising; as one commentator puts it, "you cannot atrophy a muscle that was never built." A student who lets the tool do the work before the capacity has formed has no baseline to return to.

Refrain, Then Amplify

On the refrain-then-amplify design:

1. Refrain — a program withholds a generative tool while a capacity is forming, so the learner builds the foundational capability without outsourcing it.
2. Amplify — once the student can direct the tool, judge what it returns, and answer for it, the tool is restored to amplify that capacity.

Devices are governed by a forming-versus-offloading criterion: whether a stretch of work forms a capacity or puts it through the tool. Devices are allowed where they support engaged work and excluded where they drain attention. This connects directly to concerns about Cognitive Offloading and the risk of never-skilling when a tool does the work of a capacity still forming.

Program-Level Design

The framework operates at the program level, sitting as a single floor beneath every course:

1. The program fixes outcomes and integrity, reserving teaching method to the instructor.
2. A hard-to-fake checkpoint sits at each refrain-to-amplify hinge, providing a verified transition between the phases.
3. Each element has published precedent — the withholding, the taught restoration, the test — but none of that prior work owns the movement at program level with a verified transition.

This positions sequencing as a Curriculum Design and AI Governance decision rather than a per-assignment or per-instructor choice, giving programs a principled way to decide when and how generative AI enters the curriculum in service of forming professional judgment.

What this means for practice

• Curriculum designers. Settle at the program level, not per assignment or per instructor, when a generative tool is withheld and when it is restored, deciding each step with the forming-versus-offloading criterion.
• Curriculum designers. Place a hard-to-fake checkpoint at each refrain-to-amplify hinge — an unseen variant, an adversarial defense under live questioning, repair of a flawed output — because a finished product alone cannot attest that a capacity has formed.
• Curriculum designers. Keep the checkpoints recurring rather than one-off, since a single test designed against an earlier generation of the tool can be outflanked as the tool advances.
• Administrators. Fund the checkpoint as the core of the design: the recurring, student-by-student reading is its most expensive element in time, and a program that automates the reading away has automated away its own evidence.
• Faculty developers. Support instructors in designing with the tools themselves over time, since teaching well with a new tool is a situated form of knowledge that readiness cannot be mandated into being. Leave room to decline personal tool use, because the floor mandates invariants rather than any individual's adoption of a tool.

Limitations

• The framework is a conceptual proposal, and the assumption that a capacity formed unaided and then amplified under supervision leaves a graduate able to judge when to refrain in work no program observes is a claim about transfer for which the authors state they have no direct evidence.
• The experimental support it cites for keeping the student's own thinking in the loop comes from single sessions, with ownership self-reported.
• The recurring checkpoint that makes the floor enforceable is expensive in time, and recurring probes invite coaching — an acute problem for the adversarial-defense case in writing — with the mitigations left to each program to design.
• The design asks instructors to teach a practice many are still learning themselves, and the authors warn that a floor is easy to read as a ceiling and an invariant easy to convert into a form to be filled in, leaving a program with the apparatus but not the design.

Citation

Torres-Sahli, M., Blake, J., Novoa-Echaurren, Á., & Pavez, I. (2026). Refrain, Then Amplify: A Curriculum Framework for Sequencing Generative AI to Form Professional Judgement. EdArXiv preprint.