When does AI tutoring support independent learning?

Scaffolding, assessment timing and the distinction between assisted success and retained capability

This retrospective review considers source versions available by 19 January 2026. The cutoff defines the evidence included; preparation and source checking took place on 21 October 2026.

Keywords: AI tutoring; scaffolding; learning; retrieval practice; independent performance

1. Assistance and learning are different outcomes

An AI tutor can help a learner finish an exercise without establishing that the learner can solve a related problem independently. It can also help the learner develop understanding through feedback that would otherwise be unavailable. The research question is which assistance arrangements improve learning, how those arrangements differ and what the resulting assessments actually measure.

This distinction is especially relevant to applied professional education. A polished marketing assignment can contain reasoning generated by the system, reasoning developed by the learner or a mixture. Assessing the submitted artifact evaluates the artifact. Assessing the person's capability requires a task in which the relevant contribution can be observed. These assessment aims can coexist, but neither should be inferred automatically from the other.

The reviewed literature includes direct generative-AI interventions and earlier research on tutoring, feedback and retrieval. Historical learning research can clarify a mechanism or an assessment design. It cannot supply the unmeasured effect of a new tutor simply because the tutor uses the language of scaffolding or personalisation. The whole instructional arrangement, including the tasks and feedback, remains the relevant intervention.

2. Narrative review method

Searches on 21 October 2026 combined AI tutoring, guardrails, independent learning, retrieval practice, self-explanation, formative feedback and cognitive tutors. 17 sources were selected from journal pages, author manuscripts, university records and an association research summary. Full text was inspected for the school-mathematics study, the college-physics study and the Tutor CoPilot preprint, as well as Chi's learning-activity framework.

The comparison extracts learner population, recipient of the AI assistance, comparator and timing of assessment. Tool access during practice is separated from tool access during the test. A direct-to-student tutor is separated from assistance delivered to a human tutor. The studies are not pooled, because the tasks and outcome scales differ. No formal systematic review, duplicate independent screening or new risk-of-bias score was undertaken.

Several foundational sources were accessible only through abstracts or research summaries. Their treatment remains at that level. Version differences are preserved: the Tutor CoPilot discussion concerns the original 2024 manuscript, while the school-mathematics source is linked to its 2025 journal publication. No finding is assigned the status of a later version that was not inspected.

3. What counts as a learning gain?

3.1 Performance during support

Practice performance records what learners can accomplish under the available assistance. It may be an important outcome if the aim is to improve current work. It combines the learner's knowledge with tool access, interface design and the feedback provided. A high score in that condition cannot isolate the learner's independent competence unless the design includes a way to observe it.

Removing the tool changes the task conditions. The subsequent test assesses what remains available to the learner without that support. The difference between the two scores is not a direct measure of dependence unless alternative explanations, such as task difficulty and test timing, are addressed. The comparison must preserve what changed and what remained constant.

3.2 Retention and transfer

An immediate test assesses a different interval from a delayed test. A problem structurally similar to practice assesses a different kind of generalisation from a new problem in another domain. Both distinctions matter when a course claims that learners retain a method or can apply it to unfamiliar professional work. The assessment must match the claim's interval and scope.

Roediger and Karpicke found that retrieval testing could improve later retention relative to further study under their experimental conditions [8]. Karpicke and Roediger's later work further examines the contribution of retrieval to learning [7]. These studies establish findings about practice arrangements in memory tasks. They support direct attention to what learners actively retrieve, without establishing an effect size for AI-supported marketing education.

3.3 Confidence and experience

Feeling engaged or believing that one understands a topic are meaningful experience measures. They differ from correctly performing a task. A tutor can receive high satisfaction ratings while leaving a misconception uncorrected. An effortful task can feel less fluent while supporting later performance. The design should retain both measures when both matter, rather than allowing one to stand in for the other.

4. A school-mathematics field experiment

Bastani and colleagues studied nearly a thousand students at a Turkish high school using GPT-4-based support during mathematics practice. They compared a standard chat interface, a tutor with safeguards and a control arrangement. The author manuscript describes four 90-minute sessions and separates assisted practice from a closed-book, closed-laptop assessment on conceptually similar material [1]. The design directly observes performance after assistance is removed.

Practice improved with both AI variants, while the standard interface was followed by worse unaided performance relative to control. The safeguarded tutor substantially mitigated that negative pattern. Its support included teacher-designed, problem-specific hints and instructions intended to avoid giving away answers [1]. The treatment therefore includes instructional design work as well as access to a language model. It is not a test of an unrestricted chatbot against another unrestricted chatbot.

The result establishes a comparison within that setting and exposure period. It does not measure permanent loss of mathematical ability or the effect of every form of AI tutoring. The important design distinction is visible: supported task success and later independent performance moved differently. A report that includes only the practice score would miss the endpoint the researchers intended to evaluate.

5. A college-physics tutoring comparison

Kestin and colleagues compared a custom AI tutor with in-class active learning. The study used a crossover design: students experienced both conditions across two physics lessons, with a pre-test and a post-test for each. Worksheets and lesson content were held constant, while the mode of interaction and feedback differed. The analysis accounted for repeated observations at student level [2].

The AI condition produced stronger immediate learning results and higher reported engagement and motivation. The intervention was a designed instructional system, rather than unrestricted access to a generic assistant. Students could also pace their work differently from the fixed classroom period. These features belong to the treatment being evaluated and should accompany any interpretation of its advantage [2].

This result and the school-mathematics finding should not be reduced to a contradiction about whether AI helps education. They involve different populations, learning materials, comparators and tutoring arrangements. The physics study concerns immediate learning in a particular designed setting. The school study explicitly contrasts supported practice with unaided performance. A general conclusion requires preserving those differences rather than counting the studies as opposing votes.

6. Assistance to the human tutor

Tutor CoPilot provides a different route. Wang and colleagues offered real-time guidance to human tutors in a US school district. In the inspected v1 manuscript, randomisation occurred at tutor level; the participant section reports 874 tutors and 1,787 students, while the abstract rounds these counts. The analysed sessions include exit-ticket outcomes that determine topic progression [3]. These units must remain distinct from the much larger number of messages.

The intention-to-treat comparison found a four-percentage-point increase in the probability of passing an exit ticket. The manuscript also examines tutoring language and reports more guiding questions and less answer-giving. Students received assistance through tutors who could decide how to use the suggestions [3]. This differs from a model tutoring students directly and from a tutor being removed entirely.

The primary outcome is topic mastery as operationalised by exit tickets, not a general measure of lifelong independent capability. Randomising access estimates the effect of making the system available in that setting. Comparing sessions in which it was actually used introduces a different analytical problem, because use responds to the session. The manuscript distinguishes those analyses; the access result is the central comparison reported here.

7. Why the assistance arrangement matters

7.1 A long-standing design question

Koedinger and Aleven discuss the assistance dilemma through experiments with cognitive tutors [5]. The question concerns when help supports learning and when it displaces activity the learner needs to perform. This predates conversational models. A new interface can change the available assistance while retaining the underlying need to decide which parts of the task should remain with the learner.

VanLehn's review compares human tutoring, intelligent tutoring systems and other tutoring systems [4]. Its classification makes the form of support and the comparison condition visible. A result against no tutoring is not equivalent to a result against an established active-learning programme. Claims that a system matches tutoring need to specify which tutoring arrangement and which learning outcome were used.

7.2 Explanation generated by the learner

Chi and colleagues studied how learners use examples and generate self-explanations when solving problems [6]. The learner's explanatory activity is different from receiving a complete explanation produced by someone else. An AI system that supplies an explanation can be informative, but that observation does not establish that it elicited the learner's own reasoning. Process evidence is needed to distinguish those cases.

Chi's later active–constructive–interactive framework classifies learning activities by what learners do, including generating information beyond the presented material and jointly developing ideas [12]. The framework does not make every dialogue interactive in the stronger learning sense. A long conversation can consist largely of requesting and receiving answers. The content of the learner's contribution remains relevant.

7.3 Cognitive load and feedback

Sweller's work examines how problem-solving demands affect learning and how instructional design can alter those demands [11]. The relevance is that reducing difficulty is not a single operation: an intervention can remove extraneous effort or remove the reasoning the learner was intended to practise. The distinction needs to be evaluated for the task, rather than inferred from whether the experience feels easier.

Shute's review concerns formative feedback intended to change the learner's thinking or behaviour [10]. Feedback is therefore defined in relation to learning, not merely the quantity of commentary. A tutor's response can identify an error, supply a hint or disclose a solution. Those actions provide different information and leave different amounts of work for the learner. Their effect depends on how they fit the instructional task.

8. Evaluation needs more than a final artifact

8.1 Define the assessed capability

An applied course may aim to teach evidence assessment, option comparison or interpretation of a research result. Each aim requires an observable performance. Producing a strategic recommendation can demonstrate several capabilities together, but it can also conceal which component caused a weakness. Assessments that separate the relevant operations make the inference more specific, while an integrated assignment can assess whether the learner can coordinate them.

The role of AI in the test must be stated. An open-tool test can evaluate competent use of assistance, including checking sources and rejecting errors. A closed-tool test can evaluate independent knowledge. Neither is inherently sufficient for every educational objective. Calling both understanding without describing the conditions makes it impossible to tell what the learner demonstrated.

8.2 Record reasoning at the relevant stage

A correct final answer can arise from several routes. The learner may understand the principle, imitate a supplied pattern or accept a model's answer. Asking for an explanation can provide additional evidence, although a model can generate that explanation too. The assessment design must make the learner's own contribution observable if independent reasoning is the intended construct.

Process records also need cautious interpretation. Time spent, number of messages and number of revisions describe activity, not automatically learning. A learner can need many messages because the system is unclear; another can ask few because the task is already familiar. These traces become informative when connected to the task and an independently assessed outcome. They should not be treated as self-validating engagement scores.

8.3 Test transfer at a defined distance

A new task can preserve the underlying structure while changing surface details, or require applying the method in a substantially different context. These are different transfer demands. A test that repeats the practice wording can assess recall or near transfer without establishing performance on unfamiliar work. The degree of similarity should accompany the result.

Professional marketing work often has several defensible answers and incomplete feedback. A mathematics problem with a clear solution and a marketing recommendation judged under uncertainty therefore differ in their answer criterion. Findings from the former can inform instructional design, but a professional programme needs its own evidence of the capability it claims to develop. No reviewed study supplies that transfer by assumption.

9. Common interpretation errors

9.1 Pooling unlike interventions

The phrase AI tutor can cover a generic chatbot, a constrained dialogue, a system supplied with teacher hints and an assistant to a human tutor. Pooling those arrangements under one treatment label makes a simple average difficult to interpret. It can conceal differences that are central to the intervention, such as who receives the suggestions and whether answers are withheld.

The comparator is equally consequential. A lecture, an active-learning class, an ordinary tutoring session and unsupported practice provide different opportunities to learn. A result against one cannot establish superiority over the others. The appropriate conclusion names both sides of the comparison, the learning material and the observed interval.

9.2 Mixing assignment and uptake

Randomising access does not force identical use. Some learners may request answers; others may use the tool for hints or clarification. An intention-to-treat estimate includes that variation because it assesses the offer of the system. An estimate conditioned on actual use asks a different question and may need additional assumptions. Usage frequency should not be substituted for random assignment in a causal claim.

9.3 Treating a score increase as a complete learning theory

A positive result identifies an outcome under the tested intervention. Explaining its mechanism requires additional evidence about what changed. A tutor can combine feedback, pacing, examples and motivation. If all change together, the experiment estimates the package. Attributing the gain entirely to personalisation or to conversational style would require a comparison that isolates that component.

9.4 Ignoring incomplete observations

A post-test can be missing because a learner was absent, stopped using the system or failed to complete the assessment. These pathways may differ between conditions. Analysis should state the observed sample and how missing outcomes were handled. Counting only completers can change the population represented by the result. It cannot be assumed to preserve the balance created by initial randomisation.

10. Implications for applied learning research

The evidence supports evaluating both assisted work and subsequent independent performance when a programme claims both productivity and capability development. The two outcomes can be measured within one design while remaining distinct. A course can legitimately teach effective tool use; it should identify that aim rather than allow a tool-supported deliverable to become an unqualified claim of independent mastery.

Dunlosky and colleagues' review evaluates widely used learning techniques, including practice testing and distributed practice [9]. It offers a broader evidence base for thinking about how study activity is arranged. It does not validate an AI programme merely because the programme uses quizzes or spacing. Implementation, content alignment and the assessed outcome still determine whether those techniques are represented in the intervention.

An applied assessment also needs a stated feedback standard. In some tasks a correct answer is verifiable; in others, quality depends on evidence, coherence and the treatment of uncertainty. A tutor trained to reward polished language can miss weak reasoning. To assess a professional method, the scoring criteria should address the actual decisions and evidence the method requires. An attractive document is only one observable product.

The practical research question can be framed without assuming a single best amount of help: which assistance arrangement improves the specified outcome for the intended learners? Answering it requires comparison, attention to prior knowledge and a follow-up interval that matches the claim. The studies reviewed here provide examples of those choices, rather than a universal template for every learner or subject.

11. Measurement over a longer interval

Immediate success can be valuable feedback while remaining incomplete evidence about retention. A later assessment can reveal whether the learner retrieves the method after the original prompts and examples are no longer visible. The interval must be described. A test the next day and a test after several months involve different opportunities for forgetting, additional study and workplace experience.

A delayed assessment also creates an attribution problem if learners receive different subsequent instruction. The observed score can reflect the original intervention together with later exposure. Recording those opportunities and preserving a comparison condition helps bound the interpretation. The aim is not to eliminate every real-world influence, but to avoid assigning all later performance to a brief initial session without evidence.

Repeated assessment can itself improve learning through retrieval, which makes the schedule part of the intervention [7], [8]. If one condition receives more testing or more explanatory feedback, the comparison includes that additional practice. A programme evaluation should therefore record the assessments that contribute to learning as well as those intended to measure it. Calling a quiz measurement does not make it instructionally inert.

Cost and scalability form a separate evaluation dimension. A system can improve an immediate score while requiring extensive expert preparation, review or supervision. Those resources belong in the implementation account when making claims about scale. They do not negate the learning result; they specify the arrangement that produced it and the resources required to reproduce it.

Risko and Gilbert review cognitive offloading, distinguishing changes in how work is performed from changes in internal cognitive demand. [13] Sparrow and colleagues study memory when information is expected to remain externally available. [14] Cognitive-forcing research examines overreliance in an assisted decision task; it supplies adjacent interface evidence rather than a direct estimate of learning. [15]

Kumar and colleagues distinguish assisted creative performance from subsequent performance without assistance. [16] Liu and colleagues examine creative output after ChatGPT is removed. These creative-task findings reinforce the need to specify tool availability at assessment, while leaving transfer to curriculum learning an empirical question. [17]

12. Limitations and conclusion

The reviewed interventions concern mathematics, physics and tutoring settings, with different populations and outcome definitions. Several foundational sources were inspected in summary form. The review does not estimate a common learning effect or validate transfer to professional marketing education. It also does not establish permanent harm or universal benefit from AI assistance.

The evidence shows that instructional design and assessment conditions matter. AI support can improve immediate learning in a designed setting, can support human tutors, and can improve assisted practice while leaving weaker unaided performance under another arrangement. These findings require the tool, recipient, comparator and timing to remain visible. A defensible claim about learning rests on what learners can subsequently demonstrate under the conditions the programme intends them to master.

Source access and appraisal register

The additional comparative sources in references 13–17 are used for the specific distinctions cited in the discussion. Access is limited to previously inspected primary-source abstracts, publication records or manuscripts. Their inclusion does not imply a fresh full-text risk-of-bias appraisal or direct validation in the application discussed here.

Sources were consulted on 21 October 2026. Access categories describe the material inspected for this narrative review. An abstract supports only its reported summary; it does not establish a check of the complete analysis or supplementary material. No independent replication or formal risk-of-bias scoring was undertaken.

References