Skip to content

Technology should augment, not supplant, learning processes in student-centered environments

Principle · All principles
Evidence · 152 claims (98 for, 44 mixed, 10 against) · 120 studies (30 theoretical, 27 qualitative, 23 review, 19 causal, 12 design, 9 associational), q1–q4 · 8 of 120 report an effect size · 148 claims rest on one study

Conditional relationship

This page owns the division of a learning process between the learner and a tool that could do it for them: a generative AI assistant that can produce the answer, the draft or the judgement, a simulation that can produce clean data, a virtual-reality module that can stand in for a hands-on procedure. For a learner who is meant to become able to do the process unaided (solve the problem, write the argument, run the measurement, carry out the procedure), using the tool so that the learner still does the process (the learner attempts first; the tool gives hints, questions, comparison or feedback on that attempt, or shows something the learner cannot otherwise perceive) is expected to produce better performance once the tool is removed (immediate-performance without the tool, delayed-retention, near-transfer) than using the tool to take the process over. Taking it over often raises performance while the tool is present; that gain is not the outcome this page is about. For simulation and virtual reality the same relationship reads: the virtual version is expected to add what the physical task cannot show, and to sit beside the physical practice that carries what the virtual version lacks (touch, real measurement error, the feel of the workplace), not to replace it.

The relationship is conditional on the learner's prior knowledge of the task (novice learners may gain from tool output that more advanced learners are hindered by), on whether the outcome is measured with the tool removed, on what the tool is configured to do (give solutions, or give hints and questions), on whether anyone checks what the tool says, and on the horizon.

What tests it in the wiki. For generative AI, several claims compare ways of using the tool on an unaided outcome: a randomised experiment in which adults using an AI assistant solved fewer fraction problems once it was removed, a quasi-experiment in which bounded AI support gave the best no-AI essays at week 8 while open collaboration gave the best supported essays, a multisite experiment in which reflective and hybrid feedback beat direct AI feedback on a delayed AI-free task, and a 9-study table on whether using AI for solutions rather than hints costs independent performance. They are mostly single studies with adults and horizons of days to weeks, and three experiments limit the model: in one, access to AI raised unaided test scores; in another, it did not significantly lower them; in a third, the preregistered tests found no effect of prior AI use on later unaided reasoning, and only an exploratory estimate found a cost, when participants had no incentive. For simulation and virtual reality against hands-on work, no claim in the wiki compares "complement" with "replace": the evidence is one within-subject comparison of 20 students' self-ratings, one reported quasi-experiment, and design reports. The default design below marks which steps rest on which.

How the converted siblings sit beside this page: Digital Learning decides whether a digital tool earns its place at all and credits gains to the method the tool carries; this page takes a tool that is already in the task and decides which part of the process stays with the learner. Scaffolding holds support versus no support, and its row on AI assistants that can produce the answer is a special case of this page; Scaffolding and Fading holds how support is withdrawn; Self-Regulated Learning holds the monitoring a learner needs to judge an AI answer; Guided Practice holds the practice phase in which an AI tutor would act. Process-Based Writing and Feedback Loops hold the drafting cycle and feedback design that AI tools enter. The Blended Learning pattern divides a course between online and in-person work.

Default design

Steps 2, 3, 4 and 6 rest on claims that compare ways of using a generative AI tool on an unaided outcome; the rest are untested proposals, kept from the earlier page and the two pages folded into it. Each step says what stands behind it. Quantities are proposals to set locally and record.

  1. Name the process the learner must own, and what the tool may do. Before the tool enters, write down the capability to be shown unaided (for example: write an argued paragraph; apply conservation of momentum to measured velocities; administer a radiopharmaceutical safely) and list which sub-processes the tool may perform (spelling, computing, rendering an invisible quantity) and which it may not (generating the argument, choosing the solution path, inventing data). Untested proposal; from the earlier page ("technology support must leave core cognitive processes in the hands of learners"). Learners can draw the same line themselves, so ask them where they would draw it: in one focus-group study of postgraduate research students, generating research ideas, original arguments or whole papers was consistently rejected as violating authorship norms, while surface-level support such as language polishing was accepted [+W], and AI-assisted ideation was acceptable while the researcher kept decision-making authority and problematic once it displaced their conceptual labour [+W] (one qualitative study, stated norms, no outcome measured). Where ownership of a written product matters, keep drafting with the learner: in two contrasts from one writing experiment, AI support during drafting lowered ownership more than support during planning or revision, and planning support preserved more than revision support (AI support during drafting lowers ownership substantially more than support during planning or revision, and planning support preserves more ownership than revision support. [~M]; an ownership rating, not a learning outcome). In an essay study comparing LLM, search-engine and no-tool groups, LLM-group participants reported low ownership of their essays and fell behind in their ability to quote from essays written minutes earlier [+W] (interviews, no test statistics).
  2. Learner attempts first, tool after. Give an independent attempt before the tool opens: a complete first draft, a solution attempt with working shown, or a set of predictions before the simulation runs. As a default, at least 10 to 15 minutes, or one full attempt, before access (proposal). Evidence: a review reports that students who worked on a task independently before consulting an LLM produced stronger outputs (Sequencing independent task engagement before LLM consultation is associated with stronger student outputs [+W]; second-hand, an output measure, no effect size). One way to enforce it in software: a gate that releases guidance only after the learner has given a reasoned answer; in its validation, the deterministic reflection gate coached filler answers and unlocked guidance only for substantive ones [+W] (a design check of the gate, no learner data).
  3. Configure the tool to give hints, questions and comparison, not solutions. Set the assistant's instructions (or choose a tool) so it asks what the learner has tried, gives the next hint rather than the answer, and compares the learner's answer with alternatives. Evidence: across the studies gathered under How genAI is used, not who uses it, predicts whether independent performance declines or improves [+M], learners who used AI for direct solutions lost independent performance while those who used it for hints improved; in a randomised experiment on fraction problems, participants who used an AI assistant solved fewer unassisted test problems and skipped more [+M], and in the same research most participants given the assistant reported using it mainly to get answers directly (61%), a report their chat logs bore out [+W], so an unconfigured tool that can give answers is mostly used for them; telling learners about the risk is not a substitute for configuring the tool: in one physics course, after a lesson on LLM use, the share of students using LLMs for solutions more often than they believed was ideal for learning rose from 53% to 75% [+W] (pre/post survey, descriptive percentages, no comparison group); in a writing quasi-experiment, bounded AI support gave the highest independent no-AI week-8 performance while open collaboration gave the highest supported quality [+M]; and, as reported second-hand, a guarded version of the same model that withheld answers erased the harm an answer-giving version did to unaided exam scores [+M] (high-school mathematics); the same experiment is reported by another review as improving problem-solving during learning but harming durable independent learning once the AI was removed [+M] (second-hand, no effect size printed; one study, not a second one).
  4. Make AI feedback ask for the learner's revision rather than supply it. When AI gives feedback on drafts, use a reflective form (questions about the learner's reasoning, the learner revises) or a hybrid of AI and teacher or peer feedback, not AI-written corrections. Evidence: Reflective and hybrid feedback outperform direct GenAI feedback on delayed AI-free transfer [+M] (higher education, scientific argumentation, multisite experiment, small effects). In EFL writing, AI-peer integrated feedback produced higher posttest cognitive engagement than peer feedback alone [~W] (one quasi-experiment; self-reported engagement, not an unaided outcome).
  5. Make checking the tool part of the task. Before learners accept an AI answer, ask them to say how confident they are and how they would check it; have them compare their own analysis with the tool's; when the tool reports a value, ask where it came from. In a lab, the learners record measurements, never the tool. Evidence: a brief metacognitive reflection prompt reduced acceptance of incorrect AI advice [+M] (college students, decision tasks); in a text-linguistics course where students analysed texts by hand first, comparing their own analyses with LLM outputs raised their sense of competence [+W] (qualitative, reflections); experimental work reported second-hand finds that AI assistance can reduce accuracy when the AI is faulty and inflate confidence even after errors [~W], and a workshop paper reports two more studies in which AI support raised task performance but left users overconfident about their accuracy, and physicians given an LLM that outperformed them did not reason better diagnostically [+W] (second-hand, no effect size); and the limit that makes this step necessary: under student pressure an AI lab facilitator produced mass values no one had measured [-M] (three student pairs, exploratory).
  6. Show learners their own pattern of use. At the end of a session, show each learner how often they took the answer from the tool rather than working it, and what that predicts. Evidence: in a preregistered online experiment, metacognitive feedback on learners' own LLM use reduced answer offloading and improved subsequent unaided test performance [+M] (the confidence interval for the unaided gain includes no effect); in the same study more answer offloading during learning went with lower unaided performance [+M] (exploratory, observational).
  7. Test without the tool, on a schedule. End each session with a short unaided check (3 to 5 items, or one paragraph, in under 10 minutes), and give an unaided task in the form of the target every two to three weeks. Read unaided capability from these checks, not from assisted work. Untested proposal as a schedule; every comparison in steps 3 to 6 found its effect only because the outcome was measured with the tool removed, and a physics framework's authors name the same test as the decisive open question for their own design: whether students retain representational competence when AI support is withdrawn [~W] (argument, no data). The unaided check is added beside the brief's own assessment, never in place of it. If the brief's rubric scores appropriate tool use, the learner's justification of it or the quality of the assisted product, keep those criteria exactly as given and grade them; the evidence on unaided performance says to include an unaided check, not to drop assisted assessment. A simple way to hold both: the learner submits the assisted work with a short note of what the tool did and why they accepted or changed it (scored on the brief's rubric), plus the short unaided check (reported separately).
  8. Use simulation and virtual reality for what the physical task cannot show, and keep the physical task. Use the virtual version to make visible what cannot be seen (a radiation field, an idealised collision) and the physical version for handling, real measurement error and realism. Sequence for a procedural skill: hands-on procedural training first, then the virtual module for the concept, then back to hands-on. Evidence: in one within-subject comparison of 20 radiological technology students who already had procedural training, VR training with radiation visualisation was rated much higher for radiation understanding than conventional training [+M], conventional training was rated higher for realism and immersion [+M], and operational skills and usability ratings did not differ [~M] (self-ratings, immediate, fixed order; equivalence not tested); the same students also rated VR higher for safety awareness and radiation protection [+W] and for learning satisfaction and self-evaluation [+W]. A review states the same position as an argument: immersive technologies such as VR should reinforce, not substitute, the original educational action [+W]. From the folded collision-carts page (untested proposal): after a simulation lesson, run the same collisions on real carts, if possible with video analysis (for example Tracker), so learners meet random measurement error and apply conservation of momentum to velocities just before and just after the collision, which the simulation does not model. Where real observational data exist beside a simulation (a research-based data lab), run the two in parallel rather than one in place of the other, have learners set their own criterion for locating the feature in both, compare their value with the one the simulation calculates, and include real data sets in which the feature is absent. Evidence: one design report of an undergraduate space-physics lab used over eleven years (instructor observations, no comparison group) finds that students locate the boundary well in magnetic-field data but struggle with the messier particle data, and are unsettled by data sets that contain no crossing [~W]; that their fits agree with each other but often differ from the simulation's own calculated location, through systematic differences in how they define it [~W]; and that overall they reach reasonable agreement with model predictions [+W]. The real data carry what the clean simulation lacks; the gap between the learners' value and the tool's is the thing to discuss, not to hide.
  9. Make the virtual activity ask for thinking, not copying. Simulation worksheets should ask for a prediction before each run and an explanation after, not values to copy. Evidence: after a guided-inquiry simulation lesson, students' self-reported knowledge rose and most rated it valuable [+M], but some found the worksheet did not generate much thinking, only values to copy [~W] (one lesson, self-report); students in the same lessons said the virtual lab made them figure out concepts themselves [+W]. Where the simulation is AI-generated, have learners check its output against the governing law before using it; its authors argue that validating and refining AI-generated simulations promotes physics understanding [+W] (argument, no data).
  10. Keep the teacher deciding. The teacher reads the tool's output or scores and decides what to re-teach, to whom, and when the tool is used; the tool does not set the lesson. Evidence (qualitative, one study): teachers who positioned an AI writing tool as a partner remained instructional leaders and used its score data to decide what to re-teach [+M], and the focus of teacher–student talk was learning with partner teachers, the score with grading-assistant teachers and completion with substitute teachers [+M].
  11. Withdraw support across sessions, not only within one. A default arc for a term (proposal): sessions 1 to 3, tool limited to hints after an independent attempt, a reflection prompt before any AI answer is accepted, and an unaided check at the end of each session; sessions 4 to 6, learners use the tool to critique or compare with their own work and see their own use pattern; later sessions, learners choose when to use it, with unaided target tasks every two to three weeks. Move a learner to freer use once their unaided checks on two consecutive occasions are close to their assisted work; move a learner back to hint-only use if unaided results fall while assisted results rise. For simulation and VR: hands-on first, virtual for the invisible or the impossible, hands-on again before assessment; in a data lab, simulation and real data side by side throughout, with real data sets carrying more of the work in later sessions (proposal). How support is withdrawn is modelled in Scaffolding and Fading.

When a condition is not met: if the tool cannot be configured (a public chatbot learners use at home), move the independent attempt and the unaided check into class time, report them beside whatever the brief already grades, and ask learners to note how they used the tool at home; if no physical equipment exists, say plainly what the simulation omits (measurement error, handling) and add a recorded real-world video to analyse.

Fitting the design to a situation

The three facts that most change the decision:

  • What the learner can already do on the task without the tool. Tool output that helps a learner with little prior knowledge can hinder one who could do it. If unstated, ask: can the learner do one representative item now, unaided, and explain it?
  • Whether the outcome that counts is measured with the tool removed, and when. If the target is the product made with the tool, the trade-offs on this page matter less; if it is unaided performance or retention, they decide the design. If unstated, ask: what must the learner do without the tool, and how long after?
  • What the tool will do when asked, and who checks it. An assistant that gives full solutions, invents data or writes feedback the learner simply accepts takes the process over; one limited to hints, or checked by a person, does not. If unstated, ask: can the tool be limited, and who reads its output before the learner relies on it?
If the brief says… Then change… Basis
Learners are novice on this task Allow tool explanations after an attempt, but keep the unaided check; watch for dependence Descriptive user-study patterns that the authors read as GPT output possibly scaffolding IVDA students and reducing CV students' performance, an interpretation by prior knowledge that the study did not measure [~M] (computer science students, descriptive, multiple-choice)
Learners are advanced or already competent on the task Restrict the tool to critique of the learner's own finished work; where AI text is used, require the learner to verify and revise it Same claim [~M]; Active engagement with AI outputs (cross-verifying, refining prompts, revising) supports productive learning and depends on learners' prior knowledge [~W] (reported second-hand: more experienced writers revised AI text more thoroughly)
Learners are child (early years or primary) Keep hands-on exploration first; use digital tools for practice and for what the classroom cannot provide; do not give open generative AI for the core task Digital tools complemented hands-on investigations by providing opportunities not possible in the classroom and letting children easily practice what they learned in hands-on activities. [+W] (pilot report, early science); The impact of AI on learning, motivation, and meaning differs by life stage: AI functions as a supplement for those with established skills but risks bypassing foundational experiences for younger learners [~W] (theoretical)
Learners are adolescent (secondary school) Use a guarded tutor (hints, no answers) for practice and an unaided exam; never an open chatbot for graded practice Unguarded answer-giving AI harmed unaided exam performance while a guarded version of the same model erased the harm (Bastani et al., 2025, as reported) [+M] (high-school mathematics, reported second-hand)
Goal is writing, and the learner's ownership of the text matters Keep drafting with the learner; allow AI for planning ideas or for questions on a finished draft AI support during drafting lowers ownership substantially more than support during planning or revision, and planning support preserves more ownership than revision support. [~M] (adults; ownership ratings, not learning)
Goal is a clinical or practical procedure (skill-performance) in workplace-clinical training Teach the procedure hands-on first; add VR for what cannot be seen (radiation, hazards); assess on the real procedure VR-based training with radiation field visualization received significantly higher ratings for radiation understanding and visualization than conventional training (r=0.878) [+M]; Conventional training received significantly higher ratings for perceived realism and immersion than VR-based training (r=0.774) [~M] (20 students with prior procedural training, self-ratings)
Goal is conceptual-understanding in a science lab, and simulation is proposed in place of the lab Keep at least one physical run of each key experiment; check understanding at a later exam, not only right after AI-generated and prebuilt simulations produced higher immediate conceptual-assessment scores than physical laboratory work, but the advantage did not persist in later examinations [~M] (one quasi-experiment, reported in a review)
A science lab where real observational or measured data can sit beside a simulation (research data, a field data set) Run simulation and real data in parallel; learners define their own criterion for the feature and apply it to both; compare their value with the simulation's calculated one and discuss the gap; include at least one real data set without the feature Student-defined boundary locations agree internally but often differ from the simulation's calculated values, through how students define the boundary [~W]; students find crossings well in field data, struggle with particle data, and are unsettled by data sets with no crossing [~W] (one undergraduate physics lab, instructor observations over eleven years, no comparison; carried to other sciences untested)
The tool reports facts or data the learner cannot check (a chatbot in a lab, AI feedback on data analysis) Learners record their own measurements; require a source or a check for every AI-supplied value AI facilitation of ISLE inquiry is fragile: under student pressure the Gem crossed from scaffolding reasoning to inventing data [-M]; In a small case study, students given ChatGPT feedback valued its immediacy and alternative analytical perspectives, but one distrusted its answers as 'often made up' for complex data analysis [~W]; in a study reported by a scoping review, only 6 of 23 students learned a set of criteria correctly from ChatGPT self-study, against all 23 from internet search, and the model had fabricated a similarly named syndrome [~W] (medical students, second-hand)
online-self-paced, no teacher present Build in the reflection prompt before accepting AI answers and an automated report of each learner's own offloading; end each module with an unaided item In one preregistered experiment, metacognitive feedback on learners' own LLM use lowered the odds of answer offloading and raised the odds of a correct unaided answer on one-sided tests, with 95% intervals reaching no effect [+M] (online experiment); A brief metacognitive reflection prompt reduces acceptance of incorrect AI advice compared with open ChatGPT support alone [+M]
classroom with one teacher and many learners, teacher-facing AI alerts or dashboards Limit alerts to a few the teacher acts on; the teacher keeps their own observation round; the teacher decides re-teaching from the data Real-time AI alerts can expand teachers' situational awareness but in some cases overload teachers' attention or draw focus away from their own observation flow [~W] (K-12 review, reported second-hand); Teachers who positioned AI as a partner remained instructional leaders, weaving the tool into rich instruction and using score data to decide what to re-teach [+M] (qualitative)
Scarce facilitator time (one tutor for many drafts, volunteers, a busy shift) Use rubric-aligned AI comments for routine points and spend the teacher's minutes on one higher-order comment per learner; keep the learner revising Teachers reported that rubric-aligned AI feedback saved time on repetitive comments, enabling focus on higher-order instruction [+W] (five teachers' self-reports); Reflective and hybrid feedback outperform direct GenAI feedback on delayed AI-free transfer [+M]
Learners are under-resourced, have low confidence with the tool or low literacy Teach how to question and summarise AI output in a scaffolded group task; do not rely on technical training or unguided use In unguided GenAI collaboration, ethnic minority preparatory students' strategy talk co-occurred with passive sharing and uncritical acceptance, forming a plan-and-paste loop [+W]; After scaffolding, strategy talk reconfigured around summarizing, critical evaluation, and co-construction, indicating repurposing toward critical scrutiny of AI outputs [+W] (one design study, groups of preparatory students over two weeks; carried to low literacy untested)
Adults in work settings or postgraduate researchers who will use the tool in their job Teach the tool as part of the work, with explicit human checks at each stage; agree with the learners which parts the tool may do (language polishing, searching, summarising) and which stay theirs (the idea, the argument, the judgement), starting from the line they already draw; still keep one unaided task for the core judgement Nearly all postgraduate research students in the study used GenAI across multiple research stages, positioning it as a supportive assistant with human oversight retained [+W] (focus groups, 27 of 28 students; reported use, not outcomes); in the same study, students rejected AI generation of research ideas, original arguments or whole papers while accepting surface-level support [+W] and found AI ideation acceptable only while they kept decision-making authority [+W] (one qualitative study, stated norms); the unaided task is an untested proposal
The brief relies on telling learners about AI risks (a lesson, a policy, an honour statement) to keep them doing the work Keep the lesson or policy, but do not count on it: also configure the tool for hints (step 3), show each learner their own use pattern (step 6) and keep the unaided check After a lesson on LLM use, the share of physics students using LLMs for solutions more often than they believed ideal rose from 53% to 75% [+W] (one course, pre/post self-report, no comparison group); most participants given an AI assistant reported using it mainly to get answers, matching their chat logs [+W] (online adults)
Time is one session only Independent attempt (10 minutes), tool in hint mode (20 minutes), unaided check (5 to 10 minutes); report only the unaided check, and claim no retention Untested proposal as a plan; the single-session effect of AI use on an unaided test is in the fraction-problem experiment [+M]
A fixed short time budget (for example 30 minutes a week) Spend it on the independent attempt and the unaided check; let the tool be used outside it, and do not cut the check to make room for tool use Untested proposal
Stakes are high, or the final assessment is unaided Make sure the unaided form is practised before the assessment (unaided checks in each session); keep any assisted work the brief grades, and record tool use on it Untested proposal; Greater answer offloading during learning is associated with lower subsequent unaided performance (exploratory) [+M]
A whole-task performance the brief wants kept whole (a full essay, a full procedure, a full investigation) Keep the task whole; place the tool after the learner's full attempt, as critique, rather than splitting the task into AI-assisted parts Untested proposal
Mixed modes: some learners remote, some in the room, or phone-only Give everyone the same independent attempt and unaided check in a form that works on a phone (short typed answer, photo of handwritten work); keep tool settings the same for all Untested proposal
A product or performance for a real audience Make the authorship rule public (which parts the tool may do) and have learners state how they used it; keep drafting with the learner Untested proposal; ownership evidence as in the writing row (drafting support [~M])
The brief names its own routine (its AI-use policy, rubric, prompt template or check-in) Keep it; change only what the evidence says matters: whether the tool gives solutions or hints, and whether there is an unaided check Untested proposal
The brief gives its own categories, criteria or rubric Use them as the tool's feedback criteria and as the unaided check's criteria; do not substitute this page's Untested proposal
The brief's rubric scores tool use itself (appropriate use, the learner's justification of it, the assisted product) Keep those criteria as given and grade them; add a short unaided check beside them, reported separately; ask for a brief note with each assisted submission of what the tool did and why the learner accepted or changed it Untested proposal as an assessment design; the reason for the added unaided check is Greater answer offloading during learning is associated with lower subsequent unaided performance (exploratory) [+M], which says nothing against assessing tool use
Offline or no connection (a prison, a remote site, a phone with no data), so no live AI or simulation Do not plan around a live tool. Use offline materials: worked examples and hint cards the learner opens only after an attempt (paper, or files preloaded on the device), answer keys with explanations to check their own work, a printed self-check card for any pre-generated tool output; keep the independent attempt and the unaided check exactly as in the default; if a tool is available only at set times (a supervised session, a sync point), use it then for critique of work already done offline Untested proposal; the hint-not-solution reasoning is carried from How genAI is used, not who uses it, predicts whether independent performance declines or improves [+M] (online settings, not offline)
Intermittent attendance (learners miss sessions, shift work, release dates, irregular study on a mobile device) Make each session self-contained: one independent attempt, hints after, one unaided check, a few minutes each; open each return with a short unaided item on the last topic before new work; let a learner move to freer tool use only after two unaided checks in a row are close to their assisted work, however far apart the sessions Untested proposal
Private or sensitive material (personal writing, patient data, learner records) The learner decides what goes into the tool; use local or institution-approved tools; never require uploading private work Untested proposal
No physical equipment or hands-on practice is available Say what the simulation omits; add recorded real-world video to analyse; schedule hands-on practice before any assessment of the procedure Untested proposal; from the folded collision-carts page

Twenty of the twenty-eight rows cite a claim; five of those (one session, high stakes, a real audience, a rubric scoring tool use, offline settings) are untested proposals that cite a claim only for a neighbouring finding, and the rest are untested proposals. Most of these claims are single studies of adults in higher education, carried here to other ages and settings; the child and adolescent rows rest on a pilot report, a theoretical argument and a second-hand report. The VR row rests on one within-subject comparison of self-ratings, which its authors describe as a check of the system's intended function rather than evidence of educational superiority; the data-lab row rests on one design report's instructor observations.

Observation, state and explanation

What can be observed: the learner's work with and without the tool, the tool's log (what was asked, what was returned, what was copied), the time between the tool's answer and the learner's submission, and the learner's own account of how they used it. "Has offloaded the thinking", "trusts the tool too much" and "owns the work" are inferred states. Assisted success shows what learner and tool did together. The learner's own account is worth asking for: in one experiment, participants' self-reports of using the AI mainly for answers matched their chat logs [+W]. The log alone is not enough, because the same request and the same returned text can be read actively or passively: a theoretical article on medical training argues that within a task where AI help is appropriate, a learner passively taking in AI output looks the same as one engaging actively, and plausible errors pass unnoticed when the learner's knowledge is incomplete [~W] (argument, no data), so it proposes that an expert checks how the learner engaged, which is what the probes below do.

Response Plausible explanations Observation that could change the interpretation
Strong assisted work, weak unaided work The tool did the process; the unaided task differs from the assisted one; anxiety or time pressure without the tool Give a matched unaided item without time limit; read the tool log for copied solutions; ask the learner to explain one step of their assisted work.
Accepts every AI answer without change, including wrong ones Lacks the knowledge to judge; believes the tool is reliable; time pressure makes checking costly Ask for a confidence rating and a one-line check before the learner accepts the tool's answer; review together one real case where the tool was wrong (from the learner's own log or a documented example), rather than planting wrong answers in learning materials, which is an untested and questionable probe.
Refuses to use the tool at all Already competent; distrust; a rule or norm against it; does not know how to ask Compare unaided performance with peers; ask what they think the tool would do; offer a hint-only mode.
High VR or simulation ratings, errors on the real equipment The virtual version taught the concept, not the handling; no hands-on practice since; the virtual interface differs from the real one Observe one real procedure after the virtual session; ask the learner to point out the invisible quantity on the real setup.
Low sense of ownership of a good product The tool drafted it; the learner revised little; the learner is unsure where the ideas came from Ask which ideas were theirs; compare their first independent draft with the final.

These rows are proposals, untested as diagnostics. Each probe changes what it measures (asking for a check before acceptance changes how learners use the tool), so record each one.

Evidence and qualifications

The claims below are the core of the model. They are mostly single studies, from 2026 preprints and journals, with adults; two print an effect size.

Claims that limit or qualify the model:

Learners, tasks and outcomes differ across these claims (online adults on fractions, students writing essays, radiological technology students rating VR), so do not rank them by effect label. None follows learners beyond a term, and none compares a complementary simulation or VR design with a replacement design.

The evidence by question

Further evidence, not yet read against this model

Claims this page cited that the model above does not use. Each keeps its old marker; the label gives its evidence as recorded on the claim page. They were read against the model on 2026-10-15 and left here because none changes a step, a row or the relationship: each group's label says how far it bears. Nineteen claims from this list were worked into the model at that reading.

The source framework this page began from (Land and Hannafin, 1996). Conceptual; it states the augment-not-supplant assumption but does not test it.

Simulation, virtual reality and hands-on work. Read 2026-10-15: one studio design report (GenAI for early ideation, hands-on work for refinement) and one second-hand meta-analysis on handwriting versus typing notes, a transcription question rather than a tool doing the learner's process; neither compares complement with replacement.

Measured outcomes of AI use that supports or substitutes for the learner's work. Read 2026-10-15: survey models, interaction-log analyses, self-reports and stated intentions; none measures performance with the tool removed, so they bear on the diagnosis rows (what offloading looks like in a log or a survey) rather than on the relationship.

Reviews of AI in education. Read 2026-10-15: summaries stating the scaffold-versus-substitute distinction or listing overreliance as a risk, with no effect sizes and no unaided-outcome comparison printed; they agree with the model without adding a test, and the studies behind them are not traced.

How learners and teachers describe the tool's role. Read 2026-10-15: interviews, surveys, essays and design reports; they bear on how the division of work is perceived and on the learner-valued goal (for example, students preferring hybrid AI and instructor feedback, teachers bounding chatbot answer-giving), not on whether it changes unaided performance.

Arguments, frameworks and design theses. Read 2026-10-15: theoretical, no new data; they restate the relationship or a mechanism for it (offloading, cognitive debt, automation bias, substitutive versus duplicative offloading), not tests.

Objective and learner-valued goal

The designer's objective is usually unaided capability: the essay the learner can write, the problem they can solve, the procedure they can carry out without the tool, at a stated horizon. The learner may value something else: finishing the assignment, a better grade on assisted work, saving time, or learning to use the tool well because their job will require it. These diverge sharply here, because the tool can deliver the learner's goal (a finished product) while bypassing the designer's (unaided capability). Elicit both, and record them.

For example, a postgraduate student may want to use generative AI efficiently across the research process, because that is how they will work; the programme's objective is that they can judge a method and an argument themselves. Both can be served: assess the judgement unaided, and teach the tool as part of the work with explicit checks. Learners often already value keeping the core work: postgraduate research students in one study rejected AI generation of ideas and arguments as violating authorship [+W], and in a survey at one university high-achieving students described using AI selectively, verifying its outputs and keeping it subordinate to their own judgement [+W] (themes included "fear of cognitive substitution" and verification "as part of the work"; open-ended answers, one author's coding). Where a learner's stated value and their observed use diverge (for example, believing solution-getting is bad for learning and doing it anyway, as in the physics course in step 3), show them the divergence rather than assume either one. A nursing student who values passing quickly may take every AI answer; one who values being safe with patients may want the hands-on procedure repeated. Agreement and divergence are both observations.

What would revise this model?

The expectation should weaken if randomised comparisons with comparable learners show that solution-giving AI use produces unaided and delayed performance as good as hint-limited use; the three experiments in which AI access did not lower (or raised) unaided scores already point that way for unrestricted access in some tasks, and the model should then say which tasks. If metacognitive feedback or reflection prompts do not raise unaided performance in replications (the one estimate's interval includes no effect), drop step 6 to a proposal. If a comparison of simulation or VR used instead of hands-on practice shows equal real-world performance at a delay, "complement, not replace" becomes a cost decision rather than a learning one. Do not protect the model by calling every failure "the wrong kind of use" after the fact; classify use before seeing the outcome.

Assisted performance, unaided immediate performance, delayed retention, transfer, ownership and the learner's own goal are separate outcomes. The present evidence establishes no safe amount of AI use, no minimum length of independent work before access, and no sequence of hands-on and virtual training that has been tested against another.

Claims

Examples

Key Sources

  • Land, Susan M.; Hannafin, Michael J. (1996). Student-Centered Learning Environments: Foundations, Assumptions, and Implications. https://eric.ed.gov/?id=ED397810
  • Loo Kang WEE. (2015). One-Dimensional Collision Carts Computer Model and its Design Ideas for Productive Experiential Learning. http://iwant2study.org/ospsg/index.php/interactive-resources/physics/02-newtonian-mechanics/02-dynamics/46-one-dimension-collision-js-model
  • Kakimoto Akihiro, Fujise Daisuke, Watanabe Shota, Otaka Yutaka, Minemoto Sakura, Yasumatsu Ayaka, Hasegawa Shin. (2026). Development and Educational Evaluation of a Radiation Visualization System in a Virtual Reality Training Module for Radiopharmaceutical Administration: Mixed Methods Study of Students’ Learning Experiences. JMIR Medical Education. https://doi.org/10.2196/104397
  • Crumley, Palczewski, and Kaster. (2007). Examining the Location of the Magnetopause in an Undergraduate Lab. https://arxiv.org/abs/0706.1985. doi:10.48550/arxiv.0706.1985
Page metadata
Field Value
type principle
id technology-augments-not-supplants-processes
aliases physical-world-augments-simulation-learning, vr-radiation-training-complementary-to-hands-on, pair-simulation-and-observation-in-research-labs
title Technology should augment, not supplant, learning processes in student-centered environments
description When a tool could do part of a task the learner must learn to do unaided, using it so the learner still does that part (attempt first; hints, questions and feedback rather than solutions; simulation or VR beside hands-on practice, not instead of it) is expected to give better performance once the tool is removed than letting it take the part over; single studies of generative AI with adults test this, three experiments limit it, and no claim tests complement against replacement for simulation or VR.
status review
canonical True
generated by claude/unspecified on 2026-10-19
trust unverified — no one has confirmed this page's content
sources 4 — land-1996, loo-kang-wee-2015, kakimoto-akihiro-2026, crumley-2007

Frontmatter is the page's machine-readable record — see the Schema & Guide.