Designing for Reach: Seven Levers and the Student Alone with AI

Synthesis: Halani (2026) proposes a course-design framework of seven levers through which a teacher shapes the environment in which students decide what to do with a machine that answers. The key innovation is Reach: which lever settings remain in force in the unsupervised moment — the student alone at night and stuck, with AI one tab away — where nearly all free choice about AI is exercised. None of the levers are new or AI-specific; what the framework adds is asking each lever how far it reaches when the teacher is absent. Three classroom configurations (think-first, direct-instruction, Harkness) illustrate coherent lever settings, and a mapping to Schoenfeld's Teaching for Robust Understanding dimensions shows the levers describe environments for thinking with or without the machine.

Generative AI can supply the distilled product of reasoning without the thinking. The framework addresses a design question at the heart of AI in education: what, in the course environment, still makes the thinking worth doing when the student is alone? It sits inside Schoenfeld's (2022) "learning environment" circle, asking what a teacher can actually adjust given a machine available to every student at every hour.

The seven levers

A lever is a recurring, configurable feature of the course environment that may influence what students do with or without supervision. Each lever has a diagnostic question:

Structure — what pays: What Feedback is given, what do assessments look like, what earns the grade, and what evidence does the grade require? The grading architecture is the certifying lever: whatever the grade certifies becomes what students take the subject to be. A grade that certifies replication invites replication; one whose top requires retention and transfer invites something else. Reach comes through what students believe will count.
Material — what they work on and when: Does the task ask students to think their way through content, or can the result be produced without that thinking? Whether a task is a problem or an exercise depends on the solver's relationship to it, not its text (Schoenfeld 2022). A discovery task may be emptied by a result produced without thinking.
Practice — how the room runs: Who is doing the thinking, where, and does every student think actively about each task? Grouping, workspaces, and the ordering of modes (demonstration-before-attempt vs. attempt-first).
Response — what gets said back: Who responds to student work (teacher, peers, a tool, or the learner), and does the response return the work or complete it? The key distinction is help that keeps the solver in charge versus help that takes the problem over. This lever names the fourth agent (the delegated tool) alongside Wiliam & Thompson's (2007) teacher, peer, and learner.
Norms — what is normal: What is normal to do, admit, claim, ask; and what propagates student-to-student without the teacher — including, now, ways of prompting AI.
Frame — what they understand about why: What students believe about why the course is built this way and what the subject is for. Frame is credible only where the other levers corroborate it — telling students process matters while the gradebook rewards product asks them to believe words over architecture.
Relationship — what is safe: Does the student know the teacher wants to know them; is confusion safe to show? Relationship makes the risks requested by the other levers tolerable (presenting a wrong attempt, admitting confusion). As Sal Khan noted on Khanmigo's low adoption, the biggest lever is "investing in the human systems."

Three general observations: every classroom has a setting on all seven levers (deliberate or default); who holds each lever varies by institution (a department may set Material, a grading policy part of Structure); and class size, contact hours, and tracking are conditions of the cultural surround, not levers — they determine what the levers operate on.

Reach

Reach is the organizing question: which levers are still set at the unsupervised moment when free choice about AI is exercised almost entirely outside the classroom? Reach is a property of a lever's setting, not the lever itself — the same lever can reach far under one setting and not at all under another, and how far often depends on the other levers' settings.

To value process over product, one could tell students process matters (Frame), but a larger-reach move is making the product less valuable (Structure).
To keep students out of a pitfall, one could warn them (Frame), but a higher-reach move is a task the pitfall does not work on (Material).

The gap between what a design produces under supervision and what it produces once the tool is gone has been measured directly (Rismanchian et al. 2026; Wong & Qiu 2026): in a "think first, ChatGPT later" study, the free-ChatGPT group produced the most creative work on the assisted task but fell back to baseline on a later no-ChatGPT task, while the think-first group produced more creative responses.

Three illustrative configurations

The same levers can be set coherently in pursuit of different aims (all three are classrooms the author has taught):

Think-first classroom (aims: curiosity, perseverance, collaboration, ownership) — the configuration built explicitly with reach in mind. Standards-based Structure with no grades until the end; in-class tasks where the thinking cannot be skipped; students work on vertical surfaces; responses return the work; norms make not-knowing sayable; declining the tool is not the goal — success is when a student, having located where their reasoning stalled, brings the machine that difficulty rather than the whole problem.
Direct-instruction classroom (aims: confidence, fluency, content mastery) — points accumulate steadily; worked examples before practice; "I do, we do, you do"; immediate correction. High reach on Structure and Material (points and near-transfer tests persist at 10 p.m.), but that reach protects an aim the machine can now reach too: where assessment items resemble practiced ones, a solution with different numbers can be produced without the reasoning.
Harkness classroom (aims: communication, ownership, meaning-making) — the first thinking happens alone at home; students present attempts around the table; a spiraled problem set. Its classroom Practice reaches backward into the home moment only indirectly, through the student's anticipation of presenting and explaining — which depends on how fully Norms, Frame, Structure, and Relationship have traveled.

Reach sorts the three more sharply than exposure: all are exposed in different places, but only the think-first design is set intentionally so the levers reach the unsupervised moment.

Relationship to existing research

The framework sits in a lineage of enumerating classrooms' adjustable dimensions — Ames's (1992) six Motivation structures, van den Akker's (2003) curriculum components — distributing those among the levers. What it adds is scale (levers are set and read at the scale of a course) and the Reach question. Its nearest neighbor is the assessment-redesign literature: Corbin, Dawson & Liu (2025) argue discursive changes cannot secure Assessment in an age of Generative AI and that structural changes are needed; Reach asks that question of full course design rather than assessment alone.

The levers are domain-general and not AI-specific: mapping them to Schoenfeld's Teaching for Robust Understanding dimensions (rich content, cognitive demand, equitable access, agency/ownership, formative assessment) shows the same levers describe environments for thinking with or without the machine.

Implications for AI in education

Design the unsupervised moment, not just the supervised one: course design should ask how far each lever reaches when the student is alone — that is where free choice about AI is exercised.
Structural changes beat discursive ones: telling students process matters (Frame) is weakly reaching; changing what pays (Structure) or what cannot be done without thinking (Material) reaches further.
Norms propagate to AI: a student who has internalized "take ideas you can reconstruct, not answers you can copy" applies it to both a classmate's notebook and a chatbot — the norm is about what help is for, not the source.
The machine changes the calculus of "reaching" aims: a direct-instruction design that once built fluency can now be short-circuited by the very tool, because its reach protects an aim the machine can reach too.
Relationship is a lever, not a disposition: investing in the human systems (as with Khanmigo) is treated here as designable, not fixed.

Connected Concepts

Connected Articles

Think First Chatgpt Later 2026 — Think First, ChatGPT Later: Independent Human Creativity
Cognitive Offloading LLM Synthesis Writing — Cognitive offloading in LLM-assisted writing
Cognitive Offloading Speedup Illusion — The speedup illusion of cognitive offloading
Generative Refusal AI Tools For Thought — Generative refusal: AI tools for thought
AI Availability Student Motivation — How AI availability shapes student motivation
Correct Answer Trap AI Tutor — The correct-answer trap in AI tutoring
Rethinking Scaffolding LLM Tutors — Rethinking scaffolding in LLM tutors
Learning By Chatting GenAI Impact — Learning by chatting? GenAI impact
Student Cheat Sheets Make Or Take — Make or take: student cheat sheets
Beyond Detection Authentic Assessment AI 2025 — Beyond detection: authentic assessment
AI Assessment Scale Reform — The AI Assessment Scale and assessment reform

Citation

Halani, A. (2026). Designing for Reach: Seven Levers and the Student Alone with AI. EdArXiv. https://osf.io/preprints/edarxiv/t4cd2_v1

View Source View Changes Edit Chat