Cognitive
Partnership.

The semester went well. The grades are good. Now close the laptop.

• Picture a student at semester's end: every essay polished with AI, every reading summarised by AI, strong grades throughout.
• Now ask them to explain their own argument with the laptop closed.
• Sometimes the knowledge is in their head — and sometimes it was living in the chat window, and it left when the AI did.
• This study is about telling those two students apart.

Two things can be true at the same time

• Working with AI feels transformative: in a three-wave study of 912 students, handing tasks to AI positively predicted self-reported transformative learning (Wang & Zhang, 2026).
• But their outcome is self-report — and perceived learning cannot detect erosion: a student can sincerely report transformation while unaided capability declines.
• Both findings can hold at once: heavy delegation that feels like growth, and a shrinking capacity to work alone.
• Adjudicating between them requires measuring the student with the AI removed.

Handing work to a tool is usually smart — until it isn't

• Cognitive offloading: distributing mental work onto external tools — a shopping list carrying your memory.
• When people expect a computer to store information, they recall less of the content and more about where to find it (Sparrow, Liu & Wegner, 2011).
• Offloading is usually adaptive: it frees limited cognitive resources for harder work (Risko & Gilbert, 2016).
• It harms learning under one condition: when it discharges work germane to the learning goal and the freed effort is not reinvested — offload the formatting, fine; offload the understanding, that's the loss.

Cognitive residue: what stays when the tool is gone

• Salomon, Perkins & Globerson (1991) distinguish effects with technology — performance while the tool is in hand — from effects of it: what remains once it is gone.
• That remainder is cognitive residue — for educators, the outcome that matters.
• The question is not what a student can do while an AI assists, but what they can reconstruct unaided.
• Residue accrues only under mindful, effortful engagement — it is earned, not absorbed.

Cognitive partnership

Sustained intellectual work with AI as a genuine partner in a shared system of thinking.

• The idea is grounded in the distributed cognition literature (Hollan, Hutchins & Kirsh, 2000): thinking spreads across persons, tools, and settings — the functional system, not the individual, is the unit of analysis.
• When student and AI work well together, the thinking happens in that configuration, and the product belongs to it.
• An alternative position to "AI as a tool": sustained co-work, with the understanding retained by the student.

Three dimensions, growing from one precondition

Teaching cannot act on the student–AI configuration itself, but it can develop what the student brings to it, so we measure the student-side orientation — three co-equal dimensions:

Ontological openness — staying open about what AI is — encourages all three: taught and measured alongside them, but not a facet of the orientation.

Purpose: think further, not finish faster

• The usual story about AI is productivity — but universities are not in the time-saving business.
• The prize is a mind that reaches further than it could alone.
• Partnership purpose: offload the trivial, keep the core — the student can explain the whole logic without the AI in the room.
• Handing over the understanding itself is over-reliance, the failure mode we want to minimise.

"My goal with AI is to think further than I could alone."

Engagement: curiosity and imagination do the reaching

• Purpose sets the destination; curiosity and imagination do the travelling.
• Without them the partnership flattens into transactions: ask, receive, submit.
• With them, an odd answer becomes a new direction — the student ends up somewhere neither party would have gone alone.
• Trait curiosity predicts AI self-efficacy, which predicts use (Schutte & Li, 2025); whether engagement predicts lasting knowledge is what we test.

"I work with AI to come up with options I wouldn't have thought of alone."

Iteration: the partnership lives in the back-and-forth

• A single exchange with AI is retrieval, not partnership.
• Taking the first answer and stopping is over-reliance wearing a friendly face.
• The partnership lives in the loop — push back, redirect, build — and the loop is where the work stays yours.
• Iteration is also behaviourally observable: students' AI process logs let self-reports be checked against actual practice.

"My best work with AI happens over several rounds, not one."

Staying open about what AI is

• To enable the cognitive-engagement mode, a student has to see AI as a subject worth real effort — hence, cognitive partnership.
• "It's just a tool" invites the opposite: minimal effort, tasks finished and forgotten.
• So the courses treat what is AI? as a genuinely open question — because for many researchers, it genuinely is.
• Ontological openness is presented as one position among several: students can take whatever position they want.
• Ontological openness encourages the three properties of cognitive partnership: intellectual augmentation as purpose, curiosity and imagination as the main mode of co-working, and iteration as the basic working principle.

"I treat what AI is as an open question."

Week 1 and Week 14: the same measures, twice

All measures run in Weeks 1 and 14; every model adjusts for the student's own baseline.

The quiz is the study's objective outcome — the residue measure.

The no-AI test: knowledge you can't google mid-answer

• Self-report cannot separate learning from perceived learning; an objective no-AI measure can.
• The items demand reconstruction, not recall: each presents a scenario and asks which reasoning is correct, and why.
• Administered on paper, in class, both waves; open-ended answers scored independently by two raters against pre-written rubrics, with inter-rater agreement reported.
• Half the items repeat across the two sittings (the pre-post anchor) and half are new at Week 14 — on paper, never returned to students — so item security holds and item memory cannot produce the Week-14 score.
• What appears on that sheet is residue: knowledge that survives the AI's removal.

Measuring the orientation: twelve items, three per block

• Each dimension is measured by three self-developed 5-point Likert items, one reverse-coded per block; ontological openness has its own three-item block outside the composite.
• The partnership composite = the mean of the three dimension means, z-scored within course — the sole confirmatory predictor.
• The items are new, so reliability is estimated from the study data (McDonald's omega and Cronbach's alpha, both waves), with a pre-committed floor: any block below ω = .60 leaves the composite and is reported descriptively.
• One behavioural validity check: the iteration items are set against iteration markers coded from students' AI process logs — the check that does not rely on self-report.

The limitations

The design must answer its own strongest objection: the researcher built the framework, teaches it, and measures it. The answer is architectural.

Participation is voluntary and pseudonymous; consent decisions are invisible to the instructor and cannot touch grades. Consent self-selection likely tilts the sample toward engaged students — range restriction on the predictor, which makes the tests conservative.

The question isn't whether students will work with AI

• That question is settled — students already work with AI, in every discipline.
• The open question: what kind of working relationship leaves students knowing more once the AI is out of the room?
• If partnership leaves measurable residue, assessment design gains a testable target: AI-free checkpoints, with/without-AI contrasts, process logs as evidence of engagement.
• Teaching design gains a template too: creative projects where students build with AI, iterate, and publish — folding the results into a personal GitHub or LinkedIn — so co-working visibly compounds into a trajectory, not just a grade, and the get-the-essay-done mindset loses its grip.
• And if the association doesn't show up, that is worth knowing too, though with 70 students we will usually only be able to say the result is inconclusive, not that there is nothing there.