Our education and training systems were built for an era in which information and cognitive labour were scarce. That era has ended.
Everything those systems measure follows from the old scarcity. A polished answer proved understanding because producing one was hard. A completed activity guaranteed capability because completing it took thinking. When any answer can be generated in seconds, the thing education has always sold - assurance that learning actually happened - quietly collapses.
"Learning, by definition, should be earned. The fact that universities now have to design deliberately for that tells us how much the conditions have changed."
Most of the sector's response has been to bolt AI tools onto existing systems. That changes the tools. It does not change the learning. The harder question is the one worth building around: how do you know capability has been built when the output of learning can be produced without the thinking behind it?
The efficiency trap
The core insight behind Earned Learning is uncomfortable: the efficiency that makes AI attractive is exactly what threatens deep learning. When AI does the cognitive work, learners do not get augmented understanding - they get cognitive offload, which the evidence links to declining critical thinking, weaker metacognition and surface-level knowledge.
Mainstream AI systems are trained to be helpful, agreeable and frictionless. They resolve uncertainty instantly, so the learner never sits with the discomfort of not knowing - which is where actual learning happens. Their output sounds authoritative, so learners mistake recognition for understanding. Their scaffolding never withdraws, so knowledge is never reconstructed independently. The learner stops asking "do I actually understand this?" because the model's confidence substitutes for their own. The hard cognitive work - structuring an argument, evaluating evidence, synthesising across sources - gets outsourced, and the learner becomes an editor of AI output rather than a thinker.
The net effect: learners feel productive, perform well on surface tasks, and learn less. It is the worst possible combination.
"AI is frictionless by nature. That is powerful for productivity, but dangerous for learning. If we remove the difficulty, we remove the part where capability is built."
Two kinds of friction
Earned Learning rests on a single distinction.
Friction that blocks - clunky systems, unclear instructions, administrative load, unnecessary complexity - serves no educational purpose. Madgwick removes it, and AI assistance is maximised for exactly these tasks.
Friction that builds - working through a problem, defending a decision, navigating ambiguity, improving through feedback - is the learning itself. Madgwick preserves it, designs for it and protects it from AI's efficiency imperative.
The aim is not to make learning artificially hard, or nostalgically manual. It is to make the right parts of learning impossible to bypass.
How the platform enforces it
Inside Madgwick Studio, AI assistance is calibrated to the cognitive demand of each task. When the task requires deep processing - generating a solution, retrieving knowledge from memory, constructing an argument - the AI steps back or turns Socratic. When the task is logistics - navigation, resources, clarification, admin - AI support is total. The learner does the thinking. The platform does the paperwork.
Unlike a standard chatbot, the scaffolding withdraws as competence grows, and the platform holds the friction in a productive band - never rescuing learners from difficulty they can handle, never letting genuine overload become disengagement. The full mechanics sit in the Calibrated Friction Architecture - the design framework behind the pedagogy, and the subject of a research programme we are just beginning.
In practice, learners attempt challenges before content unlocks, recall knowledge from memory rather than recognise it on a screen, explain their reasoning along the way, and at milestones build original work and defend it against an AI that questions and probes - never confirms or completes.
Earned Learning does not stop at the lesson boundary. The AI calibrates friction inside the Studio; in The Commons, humans do - educators and peers who challenge, question and push back.
A tale of two papers
Two studies tell the same story from opposite ends. An MIT study compared students writing essays with a generic LLM, with a search engine, or unaided. The LLM group showed the least neural engagement, the weakest sense of authorship and the poorest recall.1 A Harvard randomised controlled trial found the opposite outcome from a different design: students using a deliberately engineered AI tutor learned significantly more in less time than in-class active learning - and felt more engaged doing it.2
The difference was not the underlying model. It was the design. Generic AI erodes learning; a designed learning environment outperforms. Madgwick is the designed environment - with Earned Learning as the design.
Earned Learning is grounded in productive failure theory, desirable difficulties, cognitive load theory, self-regulated learning and critical pedagogy - formalised in the Calibrated Friction Architecture. The research programme behind it is just beginning: led by Associate Professor Aaron Driver, UNE's Chief AI Officer, and shaped by academic staff across UNE, inside a working university.
Standard AI makes learning feel easy. Madgwick makes learning feel earned.
See how Earned Learning runs inside the platform →
- Kosmyna et al. (2025), MIT Media Lab - essay writing with LLM, search engine and unaided conditions.
- Kestin et al. (2025), Scientific Reports (Nature portfolio) - randomised controlled trial, AI tutor vs in-class active learning, Harvard.