FRAMEWORK
The Etholum Framework
Generative AI does not automatically weaken human reasoning. It weakens reasoning under specific conditions — when the tool supplies a complete answer before the learner has attempted the task, when its output is trusted without question, and when support never fades. We call this pattern conditional erosion. The Framework is built to prevent it.
Grounded in established theory
Every element of the Framework traces to a body of research that predates generative AI.
Germane cognitive load
Sweller (1988)
Scaffolding and the zone of proximal development
Vygotsky (1978); Wood, Bruner & Ross (1976)
Self-regulated learning
Zimmerman (2002)
Metacognitive monitoring
Flavell (1979)
Automation complacency and automation bias
Parasuraman & Manzey (2010)
Four Pillars
A design model in four pillars, grounded in learning theory established before generative AI.
Pillar I
Curriculum: Preserving Germane Load
Keep the useful difficulty that builds understanding.
Not all mental effort is waste. Germane load is the effort through which learners organise information and form durable knowledge. A curriculum should remove confusion, not remove the thinking the task was designed to develop. Learners make an independent attempt before seeing any AI-generated answer.
Pillar II
Pedagogy: The Friction Cycle
Four steps that keep the learner responsible for the reasoning.
Friction does not mean needless difficulty. It means preserving enough resistance to require thought. The cycle protects the stages of self-regulated learning and metacognitive monitoring that disappear when AI is used without structure.
Pillar III
Assessment: Rewarding the Reasoning Process
Assess how the answer was reached, not only the answer.
A polished final product no longer shows how much thinking a learner did, and AI-detection tools are unreliable. Assessment includes drafts, oral defences, in-class reasoning, argument maps and reflections. Learners disclose what AI help they used, what they accepted, what they rejected, and how they checked it.
Pillar IV
Technology: Fading, Not Replacing, the Scaffold
Support should shrink as capability grows.
A tool that gives the same level of help regardless of progress creates dependence. Good educational AI starts with a question, then a small hint, and only later a fuller explanation. It shows uncertainty honestly instead of presenting every answer with the same confidence.
The Friction Cycle
01
Attempt
The learner works on the task before asking AI for anything. They write an initial answer, identify the problem, or explain what they already know.
02
Consult
The learner asks AI for an explanation, feedback, another perspective, or help spotting a weakness. The AI enters only after an attempt has been made.
03
Interrogate
The learner questions the AI’s response instead of accepting it. They check evidence, look for assumptions, ask for a counterargument.
04
Reflect
The learner compares their original attempt with the final result. They state what they accepted, what they rejected, what changed, and how the tool affected their thinking.
A scaffold that never fades is no longer a scaffold. It becomes a substitute for the capability it was supposed to develop.
The Learning Progression
Stage 1 · High AI support
Understand the task and recognise key information. AI support may be relatively high.
Stage 2 · Guided support
Apply that knowledge to a guided task. AI offers examples and feedback; the learner makes the important decisions.
Stage 3 · Question-led support
Analyse, compare, and identify weaknesses and assumptions. AI support narrows and shifts toward questions rather than answers.
Stage 4 · Minimal support
Evaluate evidence, challenge AI output, and justify a position. The learner can disagree with the AI and explain why.
Stage 5 · Independent reasoning
Produce original work and complete a comparable task independently. AI is no longer necessary for the main reasoning.
The goal is not permanent AI assistance. The goal is independence.
The Etholum Design Test
Nine questions to ask before adopting any curriculum, assessment, learning activity or educational AI tool.
✓ Does the learner make an independent attempt before receiving an AI-generated answer?
✓ Does the activity protect useful mental effort, or does the AI do the main thinking?
✓ Is the learner required to question, check or challenge the AI response?
✓ Does the activity include reflection on how AI influenced the learner’s thinking?
✓ Can the teacher see and assess the reasoning process, not only the final product?
✓ Does the amount of AI support decrease as the learner becomes more capable?
✓ Does the tool ask for justification and counterarguments instead of immediately giving a polished answer?
✓ Does the system show uncertainty and encourage the learner to verify important information?
✓ Can the learner eventually complete a similar task without AI assistance?
If the answers are mostly no, the activity may improve short-term performance while weakening long-term learning.
Three Commitments
The principles behind the Framework, set out in full in our manifesto.
Cultivate Judgment
Learning should strengthen the capacity to reason under uncertainty.
Preserve Agency
AI should extend human capability without displacing responsibility.
Build Institutions
Durable learning requires practices, norms and institutions that outlast tools.
What our programmes develop
The Framework describes how we design learning. These are the capabilities our programmes are built to develop.
Critical Thinking
Separating signal from noise and reasoning carefully when the answer is not clear.
AI Collaboration
Working with AI as a collaborator, and deciding when to rely on it, question it, or think independently.
Ethical Judgment
Weighing consequences and treating ethics as a practical skill.
Leadership
Guiding people through change with clarity and trust.
Continuous Learning
Maintaining habits of inquiry, reflection and growth.