Learning environments typically share one structural feature: support fades as the learner's capability grows. The training wheels eventually come off. Worked examples give way to independent problems. A clinical supervisor observes from further back; a master craftsman stops correcting every cut. The whole arc from novice to expert depends, at some point, on the removal of help.
AI doesn't do that. It's available at full capacity on day one of a course and on day one hundred. It doesn't sense that a learner is ready to struggle productively. It doesn't withdraw. When left to its own logic, AI is a learning scaffold that never comes down.
This design problem can't be solved by a single, well-crafted assignment. You can build excellent AI-integrated activities - thoughtful feedback loops, calibrated reflection prompts, honest integrity frameworks - and still produce learners who can't perform when the tool isn't there. The individual activities are sound but no one designed a sequence connecting them.
How do you build the novice-to-expert arc when AI is present at every stage? That question belongs above the activity level. It requires a framework for sequencing, and a deliberate answer to something AI never settles on its own: when does the support stop?
Expertise is a quality of mind: an expert recognizes patterns faster, exercises judgment under ambiguity, transfers principles across unfamiliar contexts, and knows when the rules apply and when they don't. Learners build that capacity over years of prolonged, effortful practice with structured feedback aimed at specific weaknesses.1
Without any of that practice, a student using AI can write a sophisticated case analysis, generate a plausible differential diagnosis, or produce clean, commented code — and still lack the judgment that makes any of those outputs trustworthy in a real context. Expertise exists in the reasoning process behind an output. AI produces the appearance of expertise, but the reasoning never happened. Prior posts in this series have documented this in detail (see Brains on Borrowed Time).
Cognitive apprenticeship, an instructional model designed to make invisible mental processes visible, addresses this problem. The CA framework holds that expertise develops through a structured sequence of observation, guided practice, and progressively independent performance.2
Cognitive apprenticeship begins with three phases: modeling, coaching, and scaffolding. Each phase has a different learning goal, and AI's role should differ from phase to phase.
Image alt. text: Six numbered cards in a row labeled Modeling, Coaching, Scaffolding, Articulation, Reflection, and Exploration.
In the modeling phase, the learner's job is to observe expert reasoning. The expert makes their thinking visible: how they approach a problem, what they notice first, what they're still unsure of, and what they rule out and why.
AI is useful here. It can generate worked examples, produce contrasting solutions at different quality levels, or simulate expert reasoning for learners to analyze and interrogate. A first-year medical student learning clinical reasoning might watch AI generate two competing diagnostic rationales for the same case presentation, then be asked: what assumptions is each one making? Where do they diverge? What would you need to know to choose between them?
The learner is building a conceptual model of what clinical reasoning looks like, before attempting it themselves. AI can carry the work here, because the learner's job in this phase is to watch, not produce.
In the coaching phase, the learner attempts the task. The learner does the cognitive work while AI provides targeted prompts, surfaces gaps, and responds to drafts without completing them. This maps directly onto the AI feedback loop model described in Beyond the Red Pen, which works when a designer positions it correctly in the sequence: after modeling has established what good reasoning looks like, and before fading begins to remove support.
Our medical student is now generating their own differential diagnosis. AI asks: what findings support your top diagnosis? What would rule it out? AI holds the frame while the learner fills it in.
Scaffolding provides support calibrated to what the learner can't yet do alone. Fading, or the deliberate removal of that support as the learner's capacity grows, is central to the whole cognitive apprenticeship model.2 AI won't do this fading on its own. A learning designer must engineer it explicitly, through structured moments where AI access is reduced or removed entirely.
Defensibility is the test at this phase (see Trust but Verify): can the learner explain and defend their reasoning when the tool isn't present? If not, the support was withdrawn before the learner had a model to work from — or it was never really withdrawn.
Our medical student now completes a clinical reasoning exercise with AI available only for factual lookup. The following week, they complete one with no AI access at all.
After scaffolding, cognitive apprenticeship adds articulation and reflection, which give learners conscious access to their own reasoning. Exploration comes last, once the designer has withdrawn support from problem setting as well as problem solving.2 In Learning Environment Modeling™ (LEM) terms, this is the Evidence moment: summative performance that proves the learner owns the skill and can defend it. Learners should arrive here able to evaluate their own reasoning in real time (see Metacognition 2.0).
AI may be technically available — a working clinician will have reference tools on hand — but the learner still has to explain reasoning that is demonstrably their own and defend it under questioning.
Image alt. text: Diagram of the cognitive apprenticeship sequence showing four phases from left to right: Modeling, Coaching, Scaffolding to Fading, and Independent Performance. AI support decreases across the sequence from active to absent, while the learner's role increases from observation to independent performance.
Our medical student presents a full clinical case analysis, presents their diagnostic reasoning under faculty questioning, and identifies where their thinking was uncertain and why.
Expertise requires more than a collection of well-designed moments. Good learning design moves the learner in a specific direction, with support structures that shift as the learner's capacity grows. When AI is present, someone has to deliberately architect that progression.
In the LX Studio design process, the Learner Snapshot Canvas asks what prior experience and capability learners are actually bringing in. The Learning Strategy Board maps what they need to be able to do by the end. The Focus Board defines what genuine evidence of that capability looks like. The LEM Blueprint sequences the activities that build toward it. AI's role should be defined at that strategy level — mapped across the full arc of the sequence, not bolted onto individual activities after the design is already set. Click to view the full sequence.
Some of this reaches past the classroom. If entry-level work continues to automate, fewer new hires will get the structured on-the-job exposure that used to build expertise, and formal learning environments will have to carry more of that development.
An AI integration conversation that starts with a tool or an assignment has already started one level too low. These questions work at the sequence level. Use them to evaluate whether your program, course, or credential is deliberately designing AI's role across the full novice-to-expert arc.
|
Phase |
Design Question |
|
Modeling |
Where do learners observe expert reasoning before they're asked to produce it — and is AI generating useful examples for analysis? |
|
Coaching |
Where do learners get iterative feedback while they're still doing the thinking — and is AI supporting the cycle without replacing it? |
|
Scaffolding |
Where does AI support deliberately decrease — and is that fade visible in your design, or left to chance? |
|
Independent Performance |
Where must learners perform without AI — and can they explain and defend their reasoning? |
If the fade isn't visible in your design, it isn't happening.
➡ Download the Expertise Sequence Design Guide — a planning tool for mapping AI roles across a program arc, built to work alongside the Learning Strategy Board and Focus Board.
Sequence design is harder than activity design, and it's the conversation we rarely see teams having before the activities are already built. If you're working through what the novice-to-expert arc looks like for your learners, that's where we do our best work.
The rest of the series addresses the design challenges found in each phase of the sequence.
Subscribe to the LX Studio newsletter for research-grounded insights on designing learning environments that perform in the age of AI.
This post is part of a series on designing learning environments for the age of AI:
View our resource library that inspired The Human Side of AI in Learning Series.