postgrok

scaffolding-vs-substitution

Properties3
tagscognition, tools, ai
typeconcept
statusseedling

2 min read

Scaffolding vs substitution

Definition here: the split sole-2026-llm-cognitive-virus draws between two forms of external cognitive support. Scaffolding reduces immediate cognitive demand while preserving or increasing the user’s later capacity to do the task. Substitution removes the need to perform the underlying cognitive operation.

The complementary and competitive pair

Following Krakauer (2016), the paper maps the split onto complementary and competitive cognitive artifacts. A complementary artifact strengthens capacities beyond its immediate use; a competitive one improves performance while potentially weakening the skill underneath. An LLM can be a tutor whose arguments the user reconstructs, challenges, and verifies, or a substitute that does synthesis, evaluation, and composition on the user’s behalf.

The withdrawal test

The paper’s operative question is what cognitive capacity remains when the tool is withdrawn. The relevant distinction is between forms of coupling, and it is the same test skill-atrophy-ai applies when it asks what skills survive reliance.

Evidence read in the vault

  • bastani-2025-ai-without-guardrails: a preregistered RCT where unrestricted GPT-4 access raised practice scores 48% and left students 17% below never-assisted controls on the unassisted exam; the guarded tutor arm eliminated the drop and was itself no better than control. The crutch mechanism is measured, not inferred, and students did not perceive the harm.
  • zi-2025-beginners-code-comprehension: CS1 students read LLM code far worse than they read the problem descriptions (32.5% against 59.3%), and 29 of 32 assumed at least once that the code did what the prompt said, against explicit warnings that it might not.
  • Wong and Qiu, Educational Psychology Review 2026, the guided “think first, ChatGPT later” protocol: still not read in this vault. The claim is the paper’s.

Graph View

Backlinks