Scaffolding, the Zone of Proximal Development, and Using GenAI Well
Two ideas from educational psychology explain why a good literature review — and a well-used AI assistant — accelerate a project instead of hollowing it out, while a badly-used one does the opposite. Both hinge on the same distinction: support that stretches you toward what you couldn't yet do alone is different in kind from a service that does it for you.
The Zone of Proximal Development
Vygotsky distinguished what a learner can do unaided from what the same learner can do with appropriate support, and named the gap between them the zone of proximal development (ZPD) [1]. Below the zone, a task needs no help. Above it, no amount of support makes the task tractable yet — the prerequisite understanding simply isn't there. Inside the zone is where support actually does something: a task you couldn't complete alone but can complete with the right prompt, structure, or explanation, and — this is the part that matters for a project — a task that, having done it once with support, you can increasingly do again without it.
Scaffolding is the practical term for support that operates in that zone: temporary structure, calibrated to what the learner needs right now, deliberately withdrawn as competence grows [2]. A good scaffold has three properties an answer machine doesn't: it's contingent (responsive to where you actually are, not a fixed script), it's temporary (it fades as you improve, rather than staying load-bearing forever), and it transfers capability to you, rather than substituting for it [3].
The Literature as Scaffolding
This reframes what a literature review is actually for, beyond the mechanical role it plays in Project Navigation & the Art of Triangulation. A paper that solved a nearby problem isn't just a citation to include — it's scaffolding: it shows you a method, a proof technique, an evaluation design, or a failure mode you couldn't yet have derived alone, calibrated to a problem close enough to yours that the transfer is real. Read well, it doesn't just tell you an answer — it changes what you're capable of doing unaided on the next problem. Read badly — copied without understanding why it worked — it's scaffolding that never gets internalised, structure you're still leaning on in month eight because nothing was ever transferred to you in month two.
The test for whether a source functioned as scaffolding rather than a shortcut is the same test Learning as a Feedback Loop proposes for AI-generated code: could you now explain, unaided, why the approach works — not just that it does?
GenAI as a Scaffold, Not an Answer Machine
Everything above applies directly to a generative AI assistant, because the same two failure modes are available and the second one is far easier to fall into with a model that responds instantly and confidently. Used as a scaffold, an AI assistant meets the three properties above: it can be genuinely contingent (it responds to your specific confusion, not a fixed script), genuinely temporary (you stop needing to ask once you understand), and it can transfer real capability. Used as an answer machine, it satisfies none of them — it produces a working artefact without changing what you can do without it, which is precisely the "open-loop" failure mode described in Learning as a Feedback Loop (and the AI Partner): the loop still spins, but inside the model rather than inside you.
| Situation | Answer-machine use | Scaffold use |
|---|---|---|
| Stuck on a design decision | "Design this component for me" | "Here's my design and reasoning — what am I not considering?" |
| A paper's method is unclear | "Just tell me what to implement" | "Explain why this precondition matters before I decide whether it holds for my data" |
| Code doesn't work | Paste the error, paste the fix, move on | Ask what class of bug this is, then find and fix it yourself; use the model to check your diagnosis afterwards |
| Writing the literature review | Ask for a summary of a paper you haven't read | Read the paper, draft your own synthesis, then ask the model to probe for gaps or unclear claims in your draft |
| Signal you're in the wrong column | You can't predict what the next output will say before you see it | You could have written a rough version yourself; the model refined or checked it |
Notice the pattern across every scaffold-use row: you produce a first attempt, and the model's role is to react to your attempt — probing, checking, explaining — rather than originating the output you submit. That ordering is what keeps you inside the zone of proximal development instead of outside the loop entirely. It's also exactly the discipline the "prompting for growth" patterns in Learning as a Feedback Loop describe for coding specifically; this page is the same argument applied to reading, designing, and writing.
A Practical Check
Before accepting AI output into your project — a paragraph, a design, a piece of code — ask which of these is true:
- Fading: if you needed this kind of help again next week on a similar problem, would you need less of it? If the answer is no, it isn't functioning as a scaffold — you've built a dependency, not a capability.
- Contingency: did the response actually engage with your specific attempt and reasoning, or would the same answer have come back regardless of what you'd tried first? A generic answer to a specific question is a sign you asked for the answer rather than for help getting there.
- Transfer: could you now do a similar task from a blank page, without the transcript open? If not, nothing was scaffolded — something was outsourced.
None of this is an argument against using AI tools in a project — used this way, they're a genuinely powerful scaffold, available on demand in a way a supervisor or a co-author isn't. It's an argument for noticing which of the two modes you're in, because only one of them leaves you more capable at the end of the project than you were at the start — and that capability, not the artefact, is most of what the project is actually for. See Process Over Product for the wider argument about why that capability outlasts the deliverable.
References
- Vygotsky, L. S. (1978). Mind in Society: The Development of Higher Psychological Processes. Harvard University Press.
- Wood, D., Bruner, J. S., & Ross, G. (1976). The role of tutoring in problem solving. Journal of Child Psychology and Psychiatry, 17(2), 89–100. https://doi.org/10.1111/j.1469-7610.1976.tb00381.x
- Puntambekar, S., & Hübscher, R. (2005). Tools for scaffolding students in a complex learning environment: What have we gained and what have we missed? Educational Psychologist, 40(1), 1–12. https://doi.org/10.1207/s15326985ep4001_1