RESEARCH DOSSIER
Human Learning with Adaptive Artificial Feedback
Abstract. Human learning with adaptive artificial feedback concerns changes in a person's knowledge, skill, strategies, and self-regulation produced during interaction with computational systems that respond to evidence about the learner. This critical scientific monograph integrates the historical psychology of tutoring and formative feedback with theories of cognitive load, scaffolding, retrieval practice, metacognition, learner modeling, and computationally adaptive instruction. It evaluates original tutoring experiments, long-run intelligent-tutoring syntheses, randomized generative-AI tutoring trials, and recent meta-analyses, including positive learning results and adverse unassisted-performance findings. The synthesis distinguishes assisted task completion from durable independent competence; examines explanatory detail, feedback timing, hint fading, motivational consequences, equity, and human–artificial agency; and develops a testable assessment framework based on delayed retention, transfer, independent problem solving, and reliable feedback. Results support condition-dependent educational benefits and expose risks when support displaces the learner's cognitive work or participation in established instructional resources. The work is a critical narrative literature synthesis, not an original experiment, systematic review, or meta-analysis.
Keywords: adaptive artificial feedback; human learning; intelligent tutoring systems; generative AI tutoring; educational psychology; formative feedback; instructional scaffolding; cognitive load; self-regulated learning; metacognition; help seeking; retrieval practice; independent competence; learning versus performance; transfer of learning; retention; educational measurement; human–AI interaction; Artificial Era
Research Object, Definition, and Scope
Human learning with adaptive artificial feedback is an experience-dependent change in a person's knowledge, skills, strategies, or self-regulation that occurs during interaction with a computational system capable of selecting instructional information in response to evidence about the learner. The defining scientific outcome belongs to the human learner: durable capability under specified assessment conditions. Feedback may consist of correctness information, conceptual explanation, error diagnosis, prompts to reconsider an inference, calibrated hints, or suggestions about regulating one's own practice. Adaptation concerns how the system changes this information using observed answers, trajectories, confidence, errors, or contextual conditions. The relevant intervention is a human–system instructional arrangement, whose efficacy must be measured beyond the immediate fluency of a supported response. The formative-feedback tradition provides foundational distinctions among feedback content, delivery, and learner characteristics Shute (2008).
The research object has six distinguishable components: the learner's entering competence, the task and its demands, the observable learner response, a model of that response, a feedback-selection rule, and the resulting learning trajectory. An adaptive message can be generated by a rule-based tutor, a probabilistic learner model, or a language model; these technical implementations do not by themselves determine the educational mechanism. A fixed worked example may be more effective than a personalized dialogue in one situation, while carefully timed explanatory guidance may offer particular benefits in another. A scientific evaluation consequently compares actual instructional functions, rather than treating the name or novelty of an AI system as a treatment definition. Research on intelligent tutoring demonstrates substantial variation by comparison condition and measurement instrument Kulik & Fletcher (2016) Ma et al. (2014).
This dossier asks how assistance changes acquisition, when it supports transfer, when it replaces productive cognitive work, how learners regulate reliance on assistance, and what evidence establishes lasting independent competence. It examines school and university settings, while recognizing that vocational, informal, and adult learning require their own transportability tests. Its scope includes instructional adaptation, feedback timing and granularity, human agency, motivational consequences, and longitudinal measurement. It does not equate a system's improved performance with the learner's development or treat subjective satisfaction as achievement. The approach is a critical narrative synthesis of published primary studies, methodological research, and formal evidence syntheses. It presents no original experiment, preregistered systematic-review protocol, or pooled meta-analysis; causal claims are attributed to the designs that actually support them.
From Human Tutoring to Computational Feedback
Historical foundations arise in the psychology of instruction, tutoring, and assistance. Wood, Bruner, and Ross described the role of tutoring in problem solving in a 1976 investigation that became foundational for the language of scaffolding: a more capable partner helps a learner undertake components of an activity and supports progress toward independent control Wood et al. (1976). This idea must be operationalized in a particular learning setting; a tutor may reduce search demands, draw attention to errors, structure a problem, or gradually withdraw assistance. The broad metaphor of support becomes testable when researchers specify the learner's actions, the distribution of help across difficulty levels, and the tasks later completed without assistance. The historical relevance of scaffolding should be distinguished from a claim that every automated hint implements the same psychological process.
Educational research subsequently examined feedback as information that shapes the learner's next attempt. The experimental literature on test-like events identified how the timing and nature of feedback can alter retention and the correction of errors; simple exposure to a right answer does not fully characterize learning Bangert-Drowns et al. (1991). Butler's classroom experiment distinguished task-focused comments from evaluative grades in relation to children's performance and interest Butler (1988). Hattie and Timperley later integrated influential feedback findings into questions about goals, current progress, and subsequent actions while acknowledging that feedback can improve or impair learning depending on its characteristics Hattie & Timperley (2007). These traditions provide a causal vocabulary for studying adaptive information, including feedback specificity, the opportunities learners receive to act upon it, and the social meaning of evaluation.
Computational tutoring developed from relatively constrained representations of student skills and pedagogical rules toward increasingly interactive systems. Early intelligent tutors selected exercises or step-level guidance using records of individual responses; later conversational systems sought to approximate some instructional moves of human tutors. AutoTutor, for example, implemented mixed-initiative natural-language dialogue and studied learning outcomes under specified comparison conditions Graesser et al. (2005). Reviews of computer and human tutoring emphasize interaction granularity and measured achievement rather than an undifferentiated contrast between machines and teachers VanLehn (2011). This history establishes a continuous research problem: whether the source of guidance can respond appropriately to learner evidence while supporting the learner's own explanatory activity. Generative language models expand the space of possible responses but also introduce new concerns about validity and consistency.
Learning, Performance, Assistance, and Independence
The central inferential problem is the difference between producing a correct answer with help and acquiring a capability that persists after help is removed. Training performance captures the achievement of an assisted person–tool system at a specific moment. Independent learning outcomes require testing what the person can understand, recall, generate, or evaluate without that tool. The difference becomes especially important when an artificial tutor supplies fluent derivations or final solutions. Research on learning versus performance documents why rapid improvement during acquisition is an unreliable universal predictor of long-term learning Soderstrom & Bjork (2015). The evaluator must therefore report which questions were answered with help, which were answered without help, and how closely later tests resemble the instructional examples.
Immediate gains can arise through at least four pathways. First, feedback may genuinely strengthen a concept or skill; second, it may guide the learner to a correct response without durable reconstruction; third, it may reduce working-memory demands so that more cognitive resources remain for schema acquisition; fourth, it may encourage answer-copying or passive recognition. These pathways may yield comparable scores during practice but diverge on delayed and novel-item assessments. Cognitive-load research provides a mechanism for the value of well-designed support when novice search overwhelms limited processing resources Sweller (1988). The same theory requires researchers to monitor whether a useful temporary reduction in burden eventually prevents the learner from developing the ability to perform the required cognitive operations.
The minimal evidence architecture separates assisted-task accuracy, independent immediate performance, delayed retention, near transfer, and far transfer. An additional measure records the ability to detect erroneous AI suggestions and explain why a proposed answer is mistaken. A tutor that raises all assisted scores but leaves independent scores unchanged has demonstrated task support, which can be valuable, but has not demonstrated independent skill growth. A tutor that yields superior delayed transfer on genuinely novel problems has stronger evidence of learning. The field experiment by Bastani and colleagues illustrates why the distinction is decisive: different generative-tutor designs raised supported practice performance, while access to unrestricted assistance reduced later unassisted math performance relative to controls Bastani et al. (2025).
Theoretical Architecture of Feedback and Regulation
Feedback can be analyzed as an information-processing event embedded within a regulatory cycle. The learner interprets a goal, attempts an action, receives evidence about a discrepancy, decides what that evidence means, and chooses a revision or subsequent practice activity. Effective feedback supplies information at a level that learners can use; it may address the correctness of the current response, the conceptual procedure, or the regulation of future learning. Hattie and Timperley's account stresses the relation among goals, current performance, and the next steps required to close a gap Hattie & Timperley (2007). The model is useful for design while requiring specific operational measures of whether a feedback message was understood, accepted, ignored, misapplied, or used to revise a strategy.
The interactive tutoring feedback model treats instructional feedback as a multidimensional intervention whose effectiveness depends on the task, the learner's previous knowledge, the form of assistance, and surrounding instruction Narciss (2013). The same explanatory message can benefit one learner and overload another. A novice may need a partially completed solution or explicit correction, whereas an experienced learner may benefit from a minimal cue that stimulates retrieval or self-explanation. Timing interacts with errors and task demands: immediate corrections may prevent rehearsing a misconception, while postponed guidance may preserve independent exploration where the learner can productively persist. An adaptive mechanism therefore requires an empirically justified policy for selecting assistance, rather than simply generating more elaborate feedback whenever the student hesitates.
Self-regulated learning adds a learner-side explanation of feedback effects. Individuals plan, monitor comprehension, evaluate their progress, and alter effort or strategy; guidance matters partly through these processes. Bjork, Dunlosky, and Kornell describe how feelings of processing fluency can lead learners to misjudge durable knowledge Bjork et al. (2013). A feedback system that delivers instant polished responses can increase subjective fluency and confidence while leaving independent retrieval untested. Conversely, prompting learners to produce an explanation or estimate their confidence before displaying assistance creates opportunities for calibration. The testable question is whether adaptive feedback improves the learner's capacity to choose effective strategies without continued algorithmic direction.
Evidence Review Design and Interpretive Rules
The present synthesis integrates four streams of evidence: experimental learning and feedback psychology; early and later intelligent tutoring evaluations; research on metacognition and self-regulated learning; and recent controlled studies of generative AI tutors. Primary experiments are prioritized when a specific causal mechanism or numerical result is asserted. Narrative and systematic reviews describe the broader configuration of evidence, while meta-analyses supply aggregate effects that remain conditional on inclusion criteria, outcome definitions, and comparator types. The comparison across streams is interpretive and does not produce a new pooled estimate. The systematic review of artificial intelligence in higher education by Zawacki-Richter and colleagues supplies an earlier map of technology applications and gaps in educator participation Zawacki-Richter et al. (2019).
Evidence is assessed by examining the unit of assignment, participant characteristics, initial competence, duration, control condition, the nature of actual help received, contamination, fidelity, attrition, and whether assessments required unaided performance. A study comparing a tutor with no instructional time estimates a different treatment contrast from a study comparing it with a capable teacher, a conventional intelligent tutor, or equal-time active instruction. The type of outcome is equally decisive. An end-of-session worksheet, a standardized course examination, a delayed concept test, and a free-choice motivation measure answer different scientific questions. Meta-analytic syntheses should therefore be read in conjunction with details of their measurement and comparator architecture Ma et al. (2014) Kulik & Fletcher (2016).
Sources were selected for their bearing on specific conceptual mechanisms and documented study designs; this is a selective critical narrative review rather than an exhaustive search. The synthesis records positive, null, and adverse results and does not treat a paper's publication venue as sufficient to establish generalizability. Results observed in a particular subject and age group are assigned to that population unless independent evidence supports extension. Bibliographic entries supply stable publisher, DOI, PubMed, or institutional links where available. When only an abstract or accessible source record supports a claim, the interpretation stays within what that record establishes. This conservative provenance practice makes it possible to revise individual claims as new longitudinal and independent replication evidence becomes available.
Feedback Content, Information Value, and Explanatory Detail
Feedback is an intervention upon the information available after a response, and its instructional value depends on what uncertainty it resolves. Knowledge of results identifies whether a response meets a criterion; knowledge of correct response supplies an answer; elaborated feedback gives a reason, a discriminating feature, or a procedure for repair. Hints may constrain a search space while leaving the learner to produce the answer. A system can additionally ask for a justification, target a misconception, supply a counterexample, or request comparison of strategies. Shute's synthesis treats these formats as distinct forms of formative information and cautions that their effects depend on task and learner characteristics Shute (2008). An adaptive design should state the pedagogical reason for choosing one format over another and document which format the user actually received.
Informational detail creates a tension between explanation and cognitive substitution. A worked solution can reveal the structure of a novice task and prevent unproductive search, while a ready-to-copy answer may remove the very activity the learner needs to master. The relevant design question is the degree of completion that preserves productive learner participation. One experimental comparison should hold the target content constant and vary the level of completion: correctness feedback only, first-step cue, error-specific hint, partial solution, and full solution. Learners should then undertake equivalent novel problems without assistance. Cognitive-load theory predicts that useful guidance depends partly on prior schemas and demands on working memory Sweller (1988). The empirical goal is to identify a knowledge-sensitive point at which added detail stops improving eventual independent performance.
Feedback's credibility must also be established at the claim level. A fluent system may provide a wrong correction, assert an invalid rule, or misidentify the student's cause of error. When learners are novices, they may lack the knowledge required to challenge that explanation. Reference solutions, teacher review, external retrieval, and constrained response formats can reduce some risks, but safety claims need measured error rates and a protocol for resolving disagreements. The generative-tutor field experiment by Bastani and colleagues supplied problem-specific teacher inputs and a hint-oriented variant, making instructional design part of the evaluated intervention Bastani et al. (2025). The contrast demonstrates why model architecture alone is an inadequate treatment descriptor: what the tutor is permitted to reveal changes what students practice.
Learner Modeling, Personalization, and Adaptation Policies
An adaptive artificial feedback system must represent the learner well enough to make an instructional decision. Some tutors estimate skill mastery from success and failure histories; others infer current misconceptions, use item-response patterns, classify help-seeking behavior, or condition on conversational descriptions. Each representation is an approximation whose validity depends on the task and the observed response process. A binary correct/incorrect signal confounds lucky guessing, systematic understanding, reading difficulty, and unsupported answer copying. A more defensible learner model uses multiple indicators and records uncertainty. The model's purpose is pedagogical prediction: which next task or message will improve later independent performance, not whether an opaque latent score looks psychologically intuitive.
An adaptation policy can choose the next problem, select explanatory depth, adjust difficulty, schedule a retrieval attempt, or change whether an answer is shown. Its decisions may be rule-based, probabilistic, or learned from prior interactions. To evaluate adaptation itself, researchers should distinguish a tailored version from a nonadaptive version using equivalent content, time, and interface. A personalized system might appear superior because it gives more practice, more teacher-like encouragement, or more opportunities to correct mistakes. Factorial experiments can isolate whether individualized content selection, dynamic timing, or richer dialogue produces an incremental benefit. Intelligent tutoring meta-analyses establish that tutor-assisted learning can outperform several common alternatives but do not imply that every component of personalization is causally necessary Ma et al. (2014) Kulik & Fletcher (2016).
Personalization depends on accurate and fair information about the student's learning. Students with identical apparent errors may require different explanations because they reached the answer through different reasoning processes. Systems can solicit a step-by-step attempt, confidence rating, or explanation before committing to an inferred misconception; such probes themselves influence learning. Learner models should be recalibrated as knowledge changes and should report whether uncertainty reflects scarce evidence or conflicting observations. The value of an adaptive prediction should be demonstrated prospectively on held-out learners and tasks. When a model adapts using self-reported confidence, testing should distinguish confidence calibration from self-presentation and social desirability. Explicitly measuring model uncertainty prevents personalization from acquiring an undeserved aura of precise insight into an individual learner.
Timing, Opportunity to Respond, and Error Correction
Instructional timing affects both the content learned and the cognitive action that occurs before a correction. Immediate feedback can prevent repeated practice of a mistaken procedure and can protect a novice from prolonged confusion. Delayed feedback may allow an initial effortful response and encourage later retrieval under some circumstances. The appropriate timing depends on whether the task involves factual identification, procedural steps, conceptual reorganization, or complex reasoning. A controlled design must vary the interval while holding feedback content and practice opportunities constant. The feedback literature historically examined the relations among correction, response certainty, and opportunities to generate answers rather than merely receive them Bangert-Drowns et al. (1991). These principles apply to adaptive tutors that decide when to intervene.
A particularly consequential design variable is the amount of time a learner spends constructing an answer before feedback becomes available. A system that instantly reveals the next step may encourage recognition instead of generation; a system that requires endless persistence can leave a novice practicing ineffective strategies. Adaptive timing can use error patterns, attempt count, elapsed time, and expressed uncertainty, but those signals should be tested rather than assumed to capture productive struggle. Hattie and Timperley's framework places feedback within the regulation of progress toward learning goals Hattie & Timperley (2007). A practical experiment compares interruption thresholds while recording time-on-task, number and nature of self-generated attempts, help requests, and independent delayed outcomes.
Feedback timing interacts with assessment. If final tests resemble the immediately corrected practice items, even weak retrieval may be sufficient. Delayed transfer requires construction of knowledge under new circumstances. Butler's experiments on retrieval practice and transfer show why the criterion test should vary beyond rehearsed questions Butler (2010). Designing an adaptive tutor around the next immediate score can therefore lead to a short-horizon optimization problem: the policy may maximize the proportion of right answers during practice at the expense of later retrieval. Scientific evaluation needs separate optimization targets for real-time support and long-term human learning, with transparent decisions when the objectives conflict.
Scaffolding, Fading, and the Transition to Autonomous Skill
Scaffolding is a developmental relation between current learner capability and support that makes a more demanding action accessible. Its operational elements include diagnosing the obstacle, allocating assistance, inviting meaningful learner action, and withdrawing help as competence increases. The classic tutoring research of Wood, Bruner, and Ross provides historical grounding for assistance calibrated to problem solving Wood et al. (1976). In digital environments, hints can be ordered from general cues to specific steps, worked examples can become completion problems, and guidance can be replaced with retrieval prompts. The critical test of successful scaffolding is the learner's subsequent ability to execute the underlying process, rather than the tutor's ability to keep the learner moving through tasks.
Fading raises a causal question about when and how assistance should diminish. One policy may reduce hints after several correct answers; another may require independent explanation, successful transfer, and calibrated confidence before reducing support. Performance-triggered fading can fail when a student succeeds by copying, guessing, or receiving unrecorded assistance. A stronger policy bases withdrawal on separate unassisted probes. It can also allow reversibility: help returns after a documented change of topic or increase in difficulty. The adaptive role is thereby connected to demonstrated skill rather than to a predetermined session count. Evidence about the relation of performance and durable acquisition cautions against treating short runs of correct responses as conclusive mastery Soderstrom & Bjork (2015).
The transition to autonomy has more than a behavioral accuracy dimension. Learners need to recognize when they require assistance, articulate what they do not understand, and test an answer against external evidence. These metacognitive competencies can be taught and evaluated separately from domain knowledge. Roll and colleagues tested feedback on help-seeking errors within a geometry tutoring environment, reporting improvements in help-seeking behavior and, in one study, transfer to new domain content in the subsequent month without continued help support Roll et al. (2011). Such evidence indicates that appropriately designed feedback can target the regulation of assistance itself. Generalization to conversational generative systems requires direct replication, because their open-ended responses create different opportunities for both effective help and dependency.
Metacognition, Confidence, and Strategic Help Seeking
Metacognition shapes whether learners benefit from feedback by influencing what they notice, trust, request, and verify. A learner who overestimates understanding may reject a useful hint; a learner who underestimates competence may request unnecessary completed solutions. The research program on self-regulated learning documents systematic errors in judgments of learning, especially when ease of processing is mistaken for durable knowledge Bjork et al. (2013). A tutoring system that improves the smoothness of problem solving may inadvertently amplify these illusions if it reports success without distinguishing the student's contribution. Confidence should therefore be collected before and after feedback, and the difference compared with objective unassisted correctness.
Strategic help seeking is a learned skill. It involves recognizing a specific obstacle, choosing a level of assistance, interpreting the information received, and returning to the problem with greater independence. Meta-cognitive tutoring research analyzed feedback about students' help-seeking errors, including requests made too soon or after too little task engagement Roll et al. (2011). These patterns can motivate adaptive interfaces that ask a learner to describe an attempted strategy before revealing a hint or permit selection among conceptual explanation and direct correction. However, blanket refusal to provide answers may be unhelpful for a novice encountering entirely unfamiliar concepts. The pedagogical policy should account for the task, prior knowledge, and legitimate needs for accessibility.
Self-regulatory outcomes require their own tests. Researchers should assess the ability to choose study strategies, request proportionate assistance, monitor understanding, and detect mistakes in a previously unseen task. An AI tutor may obtain high satisfaction ratings by providing warm reassurance, yet reassurance alone supplies little evidence of accurate metacognitive calibration. Azevedo and Cromley's randomized study of self-regulated learning training with hypermedia found that instruction in learning strategies helped university students revise mental models of the circulatory system relative to a control condition Azevedo & Cromley (2004). That result supplies a precedent for experimentally changing strategic behavior, while leaving the efficacy of conversational AI feedback to be established independently. Strong adaptive feedback systems can be evaluated for whether they teach learners how to learn.
Intelligent Tutoring Systems and Cumulative Experimental Evidence
The empirical literature on intelligent tutoring systems offers a longer historical baseline than generative AI demonstrations. Conventional tutors structure tasks, provide step-level feedback, and select instruction according to models of the learner's performance. VanLehn's review compared experiments on human tutoring, intelligent tutoring, and other computer-based instruction, finding that the commonly presumed sharp hierarchy of effectiveness was not supported in the reviewed comparisons VanLehn (2011). These findings establish that interactive computer guidance can be educationally consequential, but the exact effect depends on how tutor granularity, student participation, and the comparison condition are specified. The broad term intelligent tutor groups systems with different instructional strategies, technical mechanisms, and curricula; estimates cannot be transferred indiscriminately to modern language models.
Ma and colleagues synthesized 107 effect sizes involving 14,321 participants and reported positive achievement effects for intelligent tutoring systems relative to several comparison conditions, including teacher-led large-group instruction, non-intelligent computer-based instruction, and textbooks or workbooks Ma et al. (2014). The authors reported no statistically significant difference against individualized human tutoring or small-group instruction within their analyzed evidence. This pattern matters because the practically relevant alternative to adaptive software varies by institutional setting. It also warns against the idea that introducing tutoring software into an already effective small-group environment necessarily produces the average benefits observed against conventional large-group instruction. The systematic comparison illuminates opportunities for supplementary practice and challenges inflated claims about full replacement of teachers.
Kulik and Fletcher reviewed 50 controlled intelligent-tutoring evaluations and reported a median improvement of 0.66 standard deviations over conventional teaching conditions Kulik & Fletcher (2016). The review found that locally developed tests and standardized tests produced different estimated gains, making assessment alignment central to interpretation. The contrast between classroom-scale tutoring and standardized competence therefore requires careful measurement. A computerized system trained and tested on close variants of the same exercises may excel at a local assessment without demonstrating the same improvement in broad reasoning or long-term transfer. Evidence synthesis should retain the original sample of studies, treatment comparisons, and testing conditions rather than presenting a single gain as a universal effect of artificial feedback.
Conversational Tutoring and Learner Construction of Explanations
Human tutoring research supplies a theory of why interactive support may work. Chi and colleagues analyzed tutor and student behavior to investigate whether instructional effectiveness was primarily attributable to tutor actions, student knowledge construction, or interactions between them Chi et al. (2001). This framework redirects attention from the linguistic sophistication of a feedback message to what the learner does after receiving it. A conversational tutor may pose a productive question, but learning depends on whether the student generates, revises, or integrates an explanation. A dialogue transcript should therefore be coded for substantive learner contributions, corrected misconceptions, explanatory reasoning, and opportunities to retrieve information, not merely for session length or total number of exchanges.
AutoTutor was an early effort to implement mixed-initiative natural-language tutorial dialogue. Its design illustrates the educational significance of coordinating content, conversational prompts, and learner responses rather than using a fixed list of correct answers Graesser et al. (2005). Evaluation of conversational tutoring raises distinct methodological questions: how often does the system identify the relevant misconception, how accurately does it respond, and how much substantive work is performed by the learner? The conversational surface can resemble human tutoring while differing in diagnostic accuracy and pedagogical consistency. Experimental conditions should distinguish increased engagement resulting from dialogue from increased retention resulting from opportunities to construct knowledge.
Generative dialogue expands the range of possible instructional questions and explanations, but fluent interaction does not supply a direct measure of comprehension. A student may echo a generated explanation, agree with a plausible account, or repeatedly request simpler versions without forming a transferable representation. The cognitive process of elaborating an answer must be measured independently, through explanation quality on unfamiliar prompts and delayed tasks. Research on strategy training with hypermedia demonstrates the value of observing verbal protocols and model revision instead of relying solely on correct responses Azevedo & Cromley (2004). The design implication is that systems should create traceable opportunities for human explanation and assessment should record the learner's independent reasoning. This is an empirical research requirement, not a claim that one conversational interface will work in every educational setting.
Structured Generative Tutors and Positive Randomized Findings
Recent generative-tutoring studies often compare an intentionally designed instructional interface against a defined alternative rather than evaluating a generic chatbot. Kestin and colleagues reported a randomized controlled trial in a college physics teaching context in which students assigned to a research-based AI tutor demonstrated higher measured learning in less time than those receiving an active-learning classroom lesson; participants also reported higher engagement and motivation Kestin et al. (2025). This is an important positive result because the comparison involved active pedagogy rather than complete absence of instruction. Its inference nevertheless remains tied to the particular curriculum, duration, participants, outcome measurement, and tutor design. The study establishes efficacy for an implemented configuration under the conditions evaluated, not a general superiority theorem covering every school, discipline, or AI interface.
One mechanism suggested by structured tutoring is immediate contingent practice. A learner can request clarification, try an answer, and receive another targeted prompt within a short interval, while the system preserves a consistent instructional focus. In a conventional classroom, instructor attention and scheduling constrain the frequency of such individualized interactions. However, perceived engagement can be caused by novelty or interaction convenience and must be separated from independent acquisition. Experimental reports should measure how much time learners actually spend retrieving or reconstructing knowledge, which errors are corrected, and whether gains survive after conversational support is removed. Findings from earlier intelligent tutoring research provide plausible mechanisms but should be treated as explanatory hypotheses until the newer interface is directly decomposed experimentally VanLehn (2011) Ma et al. (2014).
Structured generative systems can also be compared with peer learning. A 2026 randomized controlled study of undergraduate macroeconomics reported that study groups, AI-assisted individual study, and combined AI-plus-group arrangements outperformed a no-AI individual comparison, with the combined condition yielding the largest reported gains across the allocated course sections Babaei-Balderlou & Shakya (2026). The comparison is suggestive of complementarities between human collaboration and artificial assistance. Because course sections rather than independently sampled students were allocated to conditions, the number of clusters, instructor effects, and local curriculum must be considered when generalizing. The empirical point is narrower and more useful than a generic claim that AI alone improves learning: instructional arrangements and peer participation may jointly shape an intervention's effect.
Unrestricted Assistance, Guardrails, and the Performance Trap
The field experiment by Bastani and colleagues is central to the causal distinction between supported performance and learning. In a high-school mathematics setting with nearly one thousand students, the researchers compared a more permissive GPT-based tutor, a tutor designed to provide hints and problem-specific guidance, and a control condition Bastani et al. (2025). Both assisted conditions produced higher performance on practice problems, yet students given the unrestricted GPT Base system performed worse on a later examination taken without access to the tool than students in the control group. The paper reports a 17 percent reduction in exam grades for GPT Base relative to control, whereas the guardrailed tutor's unassisted exam result was statistically indistinguishable from control. These contrasts establish that superior assisted practice scores may coexist with inferior independent performance.
The interpretation depends on the instructional behavior produced by each interface. Students using a system that readily supplies solutions can learn to obtain completed answers instead of exercising problem-solving operations. In the reported study, interaction analyses indicated a greater tendency to seek and copy solutions in the unrestricted condition Bastani et al. (2025). The guardrailed alternative encouraged hints and independently attempted answers. This supports the hypothesis that the amount and form of cognitive work performed by the student mediates the learning consequence of artificial help. It does not establish a universal threshold for when a hint becomes too explicit, nor does the null independent-exam contrast for the guardrailed tutor establish an achievement gain. Safeguards mitigated a measured harm in this setting, a scientifically meaningful result in its own right.
The study also illustrates why some outcome labels are misleading. A reported large percentage improvement in practice is an effect on the particular assisted task, not a proportional increase in durable mathematical knowledge. Relative percentages can depend on the control-group baseline, and absolute educational significance requires additional context. The assessment conditions determine whether the result is evidence of tool-assisted productivity or independent capability. Any deployment of adaptive feedback should preserve independent evaluation blocks that reveal whether the system has educated the person or simply completed the task with the person present. Transfer tasks, delayed assessments, and independent explanation provide complementary evidence, particularly when users may not recognize a gap between apparent fluency and their own knowledge Soderstrom & Bjork (2015) Bjork et al. (2013).
Contradictory Results, New Evidence, and Meta-Analytic Limits
The accumulating research record includes adverse findings alongside positive trials. In October 2026, Liu and colleagues reported an Annenberg Institute working paper describing a randomized trial of access to a course-integrated generative AI tutor involving 2,379 undergraduates and 30 instructors at a large U.S. public university Liu et al. (2026). The abstract reported lower final course grades and substantially reduced learning-management-system participation in sections assigned tutor access. This result introduces an institutional mechanism that controlled laboratory comparisons may miss: AI assistance can alter the use of established course resources and the broader pattern of engagement. The working-paper status and specific implementation should be retained when interpreting the result; the paper is relevant evidence, not a conclusive generalization to all AI tutors.
Systematic reviews attempt to aggregate across such heterogeneous designs. Chen and Cheung's 2025 synthesis reported positive average outcomes from 57 studies of generative AI in university learning, with differences across academic achievement, language skills, motivation, and higher-order thinking; the reported metacognition effect was not statistically significant Chen & Cheung (2025). Han, Peng, and Liu's 2025 experimental-study synthesis similarly described generally positive effects with variation by educational context and outcome dimension Han et al. (2025). A more feedback-specific 2026 meta-analysis synthesized 36 experimental and quasi-experimental studies and reported a moderate average achievement benefit, emphasizing stronger effects in learner-centered instructional settings Huang et al. (2026). These studies describe aggregates of selected designs; none abolishes observed adverse effects in particular educational arrangements.
Meta-analytic results require scrutiny of effect-size dependence, publication bias, short follow-up intervals, baseline differences, discipline composition, and whether a single study contributes multiple correlated outcome measures. Measures of engagement or subjective self-efficacy are related to learning but cannot substitute for independent delayed testing. Differences in generative AI model version and instructional prompt can reduce replicability across time. The appropriate scientific conclusion is conditional: adaptive artificial feedback can improve measured learning in some carefully designed environments, and it can diminish independent performance or disrupt engagement in others. The unresolved explanatory question concerns the design and behavioral processes that predict which outcome will occur. Larger independent multi-site trials with common outcome definitions are required to reduce uncertainty about transfer and educational equity.
Retention, Retrieval Practice, and the Meaning of Transfer
An adaptive feedback intervention should be evaluated across timescales appropriate to the target competence. Immediate exercises show whether the learner can operate during instruction; delayed recall shows which information remains accessible; near transfer tests altered instances of a practiced rule; and far transfer asks whether acquired principles can be applied in substantially different contexts. These outcomes form a hierarchy of increasingly demanding claims, although no single task can completely isolate generalization. Experimental research on retrieval practice demonstrates that attempting to retrieve information can improve long-term retention more than additional exposure under the studied conditions Roediger & Karpicke (2006). Tutors that generate answers before the learner attempts retrieval may therefore interfere with a useful learning event, even while making practice feel smoother.
Transfer requires attention to what is shared between learning and assessment. A pupil who solves a new arithmetic item with the same surface pattern may have acquired a reusable procedure; a pupil who solves a structurally different word problem may demonstrate broader conceptual abstraction. Butler's experimental study found advantages of repeated testing for transfer relative to repeated studying in its tested materials Butler (2010). These results motivate assessments that systematically manipulate cues, representational forms, and problem contexts. For artificial feedback, evaluation should record whether explanations were supplied in the same representation as the test, whether the tutor saw test-like items, and whether the learner independently generated the critical intermediate steps.
Durable learning often depends on practice schedules after feedback has been delivered. Dunlosky and colleagues reviewed a range of study techniques, identifying substantial support for practice testing and distributed practice across multiple learning materials and contexts Dunlosky et al. (2013). An adaptive tutor may use estimates of past recall to schedule later attempts, but the scheduling rule itself requires comparison with fixed or randomly allocated schedules under equal instructional opportunity. A longitudinal design should investigate whether gains remain after access to the system ends and whether learners can organize their own practice. Persistence is both a psychological outcome and a criterion for distinguishing educational investment from continuous tool-dependent execution.
Motivation, Engagement, and the Experience of Feedback
Feedback conveys more than information about correctness. It can change the learner's perceived competence, expectations of success, willingness to persist, and emotional response to error. In a classroom investigation, Butler distinguished comments that directed attention to the task from evaluations that invited comparison with others, finding different consequences for performance and interest Butler (1988). The general implication for artificial systems concerns the language and meaning of evaluation, not a universal rule that one wording always improves motivation. An adaptive tutor that responds to repeated mistakes with increasingly authoritative corrections may provide useful technical guidance while simultaneously discouraging a student from taking intellectual risks.
Motivation must be measured through behavior and validated reports rather than inferred from friendly interaction alone. A learner may praise an interface, spend more time chatting, or feel supported while completing fewer demanding retrieval attempts. Deci, Koestner, and Ryan's meta-analysis of experiments concerning rewards and intrinsic motivation demonstrates how external contingencies can alter self-directed engagement depending on the reward structure and outcome Deci et al. (1999). Artificial feedback brings related questions into view: how does praise relate to actual skill, whether the system rewards speed over explanation, and whether excessive reassurance conceals errors. These are empirically distinguishable features of interaction that should be tested with measures of voluntary engagement and later independent practice.
Recent controlled trials show that subjective engagement and formal course participation can diverge. The college physics tutoring trial reported positive engagement responses alongside higher measured learning under its particular design Kestin et al. (2025). The 2026 university working paper on course-integrated tutoring reported reduced use of learning-management resources, together with lower course grades, in its allocated treatment conditions Liu et al. (2026). These results call for outcome models that separate engagement with the tutor, engagement with the broader course, and autonomous effort after the tutor is unavailable. Adaptive systems might increase one channel while displacing another. The causal significance of such displacement should be studied using logs, independently administered assessments, and instructor observations rather than one omnibus engagement score.
Developmental Differences, Prior Knowledge, and Access
Developmental level and prior knowledge alter the learner's capacity to interpret feedback. Young learners may benefit from shorter explanations, explicit examples, and opportunities for guided rehearsal; older learners may use abstract principles, error diagnoses, or strategic comparison more efficiently, depending on their expertise. The historical scaffolding model makes assistance dependent upon present task capability and the path toward independent performance Wood et al. (1976). Developmentally sensitive evaluation must distinguish chronological age, domain experience, language proficiency, and self-regulatory skill. The same adaptive rule can produce different cognitive demands when used by a child acquiring a foundational symbol system and a university student learning an advanced procedure.
Prior knowledge may modify the relationship between explanatory detail and cognitive load. A complete solution can free a novice to notice structure while being redundant for a practiced learner whose main need is retrieval or error checking Sweller (1988). An adaptive system that estimates proficiency from a few correct answers can also misclassify students who have received outside help or who guessed successfully. Better diagnostic evidence includes multiple item formats, explanations, and appropriately spaced independent probes. The research problem is to determine which features of learner history are useful for selecting feedback and which apparently predictive features are artifacts of test language, access to resources, or previous exposure.
Equity questions arise when students differ in access to devices, stable connectivity, accessible interfaces, language support, and expert human assistance. Evaluation should predefine subgroup analyses for relevant barriers while avoiding the assumption that one demographic category directly measures learning capacity. Learner-model errors may accumulate when a system treats deviations from a dominant response style as evidence of conceptual deficiency. Meta-analyses of intelligent tutoring systems encompass different educational populations but cannot establish universal equivalence of benefits across all groups Ma et al. (2014) Kulik & Fletcher (2016). Independent replication across schools, languages, regions, disability accommodations, and socioeconomic settings is needed to distinguish broadly portable teaching principles from local advantages linked to resources and implementation.
Teachers, Peers, and the Social Organization of Artificial Feedback
Artificial feedback exists within a social organization of instruction. Teachers determine curricular goals, interpret evidence about student understanding, and judge when a difficulty warrants direct explanation, peer discussion, or individual practice. A feedback algorithm can provide rapid responses while presenting only a partial record of the learner's reasoning and classroom context. Research on computer tutoring highlights the importance of identifying what type of instruction constitutes the comparison condition VanLehn (2011). For a teacher already managing effective small-group practice, the incremental value of an automated tutor may concern availability or workload rather than a dramatic difference in achievement. Educational implementation requires evidence for that local objective.
Peer collaboration changes how assistance is interpreted. Students may test an AI suggestion together, explain conflicting answers, or delegate the task to one student and one system. The 2026 macroeconomics trial of four course sections compared individual study, group study, structured AI tutoring, and their combination; the combined condition reported the largest exam gains in the local intervention Babaei-Balderlou & Shakya (2026). The design suggests a promising interaction between artificial guidance and human discussion while retaining limits arising from the small number of allocated sections. Further trials should vary group processes explicitly: whether peers independently attempt problems before asking the system, challenge one another's interpretations, or accept generated answers without verification.
Teacher oversight can also provide an external corrective mechanism when an AI-generated explanation is inaccurate or poorly matched to the learner's needs. Human educators can adjudicate ambiguous questions, connect feedback to curriculum, identify persistent misunderstandings, and notice social factors outside the tutor's records. A review of AI research in higher education found historical gaps in educator involvement and pedagogical focus Zawacki-Richter et al. (2019). These gaps argue for observing the actual division of responsibilities between teacher, learner, and tool rather than assuming that the presence of AI amounts to individualized teaching. The quality of integration depends on the tasks allocated to each participant and on independent measures of the learner's development.
Human Agency, Reliance, and the Artificial Era
Human learning mediated by artificial feedback is a relation among an individual learner, a task environment, and an artificial participant that supplies contingent information. Within the Artificial Era, this relation can be studied as a change in the distribution of cognitive activity: the system may retrieve, explain, evaluate, prompt, or propose; the learner may plan, attempt, question, revise, remember, or defer. This division of labor is an observable feature of interaction rather than proof of any particular status of the artificial system. The relevant psychological outcome is whether the person retains the ability to pursue goals, diagnose errors, and reconstruct knowledge when the configuration changes. The learning-versus-performance distinction remains foundational Soderstrom & Bjork (2015).
Reliance has both productive and costly possibilities. Delegating a routine retrieval operation can preserve effort for conceptual reasoning, while delegating the decisive reasoning step can deprive a novice of practice. Whether a particular delegation improves education depends on what the learner is expected to internalize. A mathematics student may legitimately use a computation aid when studying proof strategy, but the same aid may eliminate the target operation when arithmetic fluency is the stated objective. The Bastani experiment shows that tool permissions and prompt constraints can substantially change supported practice behavior and subsequent unaided outcomes Bastani et al. (2025). This is a concrete reason to treat interface affordances and assistance policies as causal components of the educational intervention.
Artificial feedback should also be examined in relation to epistemic authority. A polished answer can appear more trustworthy than a tentative human judgment, even when the system is mistaken. Learners need routines for evidence checking, independent explanation, and disagreement resolution; those routines become part of learning with artificial systems. The field's positive instructional trials and adverse classroom results show why educational evaluation must simultaneously assess cognitive capability and patterns of reliance Kestin et al. (2025) Liu et al. (2026). Research on human–artificial educational relations can connect agency, confidence calibration, and distributed work without claiming that the artificial participant experiences the task as a human tutor would. Its central requirement is explicit observation of who performs which cognitive operations.
Measurement Architecture for Independent Learning
The evaluation of adaptive artificial feedback begins with a transparent definition of the target learning outcome. Researchers should specify the knowledge or skill that a human participant must demonstrate, the observable response that counts as success, the permitted tools, and the time interval between instruction and assessment. The protocol should distinguish formative evidence collected during assisted practice from criterion measures administered without help. A defensible assessment battery includes baseline competence, immediate unaided probes, delayed retention, and transfer to sufficiently unfamiliar problems. Where uncertainty is integral to the task, confidence calibration and explanation quality add information beyond binary correctness. Research on learning versus performance establishes the theoretical importance of this separation Soderstrom & Bjork (2015).
Outcome measures must be sufficiently sensitive to distinguish recognition, reproduction, inference, and flexible application. An assessment drawn almost verbatim from tutor demonstrations may reward memory for the dialogue rather than mastery of the underlying concept. Matched alternate forms, blinded scoring, preregistered rubrics, and sampling across task difficulty can reduce some forms of measurement bias. Researchers should also record attempted solutions before assistance and annotate whether students later reproduced a provided answer. The historical intelligent-tutoring reviews reported differences between locally developed and standardized measures; the choice of test is therefore part of the substantive evidence Kulik & Fletcher (2016). Educational claims should report their outcome scale, its validity evidence, and the scope of interpretation.
Process measurements illuminate possible mechanisms but are not themselves proof of durable learning. Tutor transcripts can document hints requested, time spent reviewing a correction, explanations produced, and whether a student identifies an injected error. Such data should be linked to independently assessed outcomes, because interaction logs can also capture compliance or task completion. A mixed-methods protocol may combine randomized outcome comparisons with detailed observation of reasoning processes. Studies of human tutoring analyzed the interaction between tutor and student behavior as part of the explanatory problem Chi et al. (2001). An adequate measurement architecture therefore follows both sides of the exchange while locating the final educational criterion in the human learner.
Experimental Identification, Longitudinal Designs, and Replication
Randomized controlled trials can estimate the effect of an assigned instructional arrangement, provided assignment, compliance, and outcome measurement are correctly handled. In school and university studies, students may be nested within classes or course sections, and teachers may differ across conditions. Allocating only a small number of sections can leave treatment effects entangled with classroom-level variation even when the within-section sample is large. The four-section macroeconomics trial illustrates why its reported positive outcomes should be interpreted in relation to cluster allocation Babaei-Balderlou & Shakya (2026). Multi-site studies should randomize enough independent units, record instructor effects, define the intended treatment, and analyze clustering rather than treating every student outcome as independently assigned.
An especially informative design separates model sophistication from instructional guidance. One comparison may hold the language model constant and randomize hint-only, stepwise explanation, full-answer, and strategically fading conditions. Another can hold instructional content constant while randomizing adaptive versus fixed selection. A third can test a conventional tutor against a generative tutor under matched exercise and feedback schedules. The high-school mathematics experiment provides evidence that guardrail design affects independent assessment, making these contrasts more than a technical detail Bastani et al. (2025). Pre-registration of primary outcomes, access rules, manipulation checks, and correction procedures is essential when many prompt or interface variants could otherwise be selected after results are observed.
Longitudinal evidence is needed to determine whether assistance establishes an enduring developmental advantage. Follow-up assessments should occur after system access is removed or reduced and should measure independent competence as tasks become more complex. Attrition and voluntary use create selection problems: highly motivated students may be both more frequent users and stronger performers. Logging exposure and randomizing encouragement or access windows can help separate use intensity from initial motivation, although each method carries assumptions. Replication must document model version, prompts, curriculum, feedback policy, and instructor involvement because the nominal product name may remain stable while its behavior changes. The cumulative intelligent-tutoring evidence provides an important reference point but not a substitute for reproducing contemporary interventions Ma et al. (2014) VanLehn (2011).
Reliability, Bias, Governance, and Evidence Limitations
Artificial feedback introduces a reliability problem that applies directly to instruction. A tutoring system may misidentify the student's misconception, provide an incorrect solution, or produce different explanations for equivalent questions. Reliability audits should sample outputs across subject matter, learner proficiency, prompt variation, and repeated attempts, with qualified subject-matter adjudication. Error rate alone is insufficient because a rare confident false correction may be more consequential than a frequent easily recognized ambiguity. When the pedagogical design encourages learners to trust the system, the study should assess whether students can detect misleading feedback and whether correction procedures repair the resulting misunderstanding. Validity of the feedback content is a prerequisite for interpreting educational efficacy.
Adaptive models can reproduce inequities through the data and signals used to infer learner ability. A system may treat language variation or disability-related response patterns as evidence of low proficiency, then deliver unnecessarily simplified instruction. Another model may adapt to engagement proxies that are shaped by internet instability or differences in device access. Fairness evaluation should inspect measurement equivalence, error distributions, and outcomes across relevant learner groups, with explicit protections for privacy. Data minimization is particularly important when educational interaction histories include children's work, confidence reports, or sensitive descriptions of personal difficulty. Educational governance must document retention, access, student choice, teacher oversight, and appeals for incorrect automated judgments.
Current evidence is limited by heterogeneous comparators, frequent use of short interventions, differences in assessment design, and rapid changes in generative AI systems. The 2026 feedback-specific meta-analysis reported an average positive achievement effect alongside substantial between-study heterogeneity Huang et al. (2026). That dispersion means the pooled value cannot predict a specific classroom result without further information about design and context. Conversely, the adverse unassisted mathematics result and university working paper are concrete counterevidence against a universal beneficial-effect claim Bastani et al. (2025) Liu et al. (2026). A credible synthesis maintains the outcome and design of every claim, reports uncertainty when a mechanism has not been isolated, and separates experimentally demonstrated gains from plausible educational applications.
Research Agenda and Testable Predictions
A first research priority is to identify which adaptive feedback policies maximize delayed independent competence under fixed instructional time. One testable hypothesis is that a policy requiring a learner-generated explanation before revealing a detailed answer will yield greater delayed transfer than unrestricted solution delivery on tasks whose target is explanatory reasoning. The hypothesis follows from research on retrieval, learning versus performance, and the high-school mathematics contrast, but its effect size and conditions remain open Roediger & Karpicke (2006) Soderstrom & Bjork (2015) Bastani et al. (2025). A randomized factorial design could vary prior-knowledge level, hint specificity, and fading schedule, then assess independent response quality after several delays. Pre-registered mediation analyses might examine whether the number of learner-generated reasoning steps explains part of any observed difference.
A second priority concerns the conditions under which conversational adaptation is genuinely superior to less complex instruction. A conventional rule-based system may already supply well-targeted corrective information; a more fluent language model may add flexibility, engagement, or error. Researchers should compare equal-content, equal-duration conditions and assess how accurately each system detects misconceptions. The positive college physics trial establishes the feasibility of a pedagogically structured generative tutor under its tested conditions Kestin et al. (2025). A competing hypothesis is that gains depend primarily on the teaching policy and opportunity to respond, with the underlying model contributing a smaller or task-dependent incremental effect. Isolating those factors can strengthen explanations of both successful and unsuccessful implementations.
A third priority is the long-term organization of human agency during routine use. Researchers should test whether extended access changes students' ability to recognize uncertainty, select study strategies, detect bad feedback, and resume work without the tool. Metacognitive help-seeking experiments offer a precedent for training the regulation of assistance Roll et al. (2011). New longitudinal studies should extend that logic to open-ended artificial feedback, include teachers and peer contexts, and test whether benefits generalize across subjects, languages, and access conditions. They should also evaluate cost, teacher workload, privacy, and the distribution of learning gains. The scientifically decisive outcome is a robust, replicable increase in human competence with clearly described conditions of support.
Conclusion
Human learning with adaptive artificial feedback is a distinct research object at the intersection of learning psychology, educational measurement, computational tutoring, developmental science, and human–artificial interaction. Its scientific identity rests on changes in a human learner's independent knowledge, skill, strategy, and self-regulation resulting from an instructional encounter with adaptive computational assistance. Historical research on scaffolding and formative feedback explains the importance of contingent guidance, while intelligent tutoring evaluations establish that well-designed computer-mediated instruction can produce educational benefits under specified comparison conditions Wood et al. (1976) Shute (2008) Ma et al. (2014). Generative systems extend the available interaction patterns and make feedback design, source reliability, and learner regulation more consequential.
The empirical record establishes meaningful heterogeneity. Structured generative tutoring produced favorable learning outcomes in a randomized college physics study, and a feedback-focused meta-analysis reported positive average effects across its included studies Kestin et al. (2025) Huang et al. (2026). In contrast, unrestricted assistance improved supported mathematics practice while reducing subsequent unassisted examination performance in a field experiment; a 2026 university working paper reported adverse course-grade and participation outcomes with course-integrated tutor access Bastani et al. (2025) Liu et al. (2026). These findings become coherent when the instructional intervention is defined by the human actions it encourages, the cognitive steps it supplies, the institutional resources it displaces, and the independent competencies it leaves behind. General educational benefit cannot be inferred from the presence of artificial intelligence alone.
The resulting scientific standard is operational: define the target human competence, isolate the feedback-selection mechanism, measure the learner's contribution during instruction, compare appropriate alternatives, and assess unassisted retention and transfer after assistance is reduced. Record model and curriculum versions, preserve qualified human oversight, and study developmental, motivational, and equity-related variation. Artificial feedback becomes an educationally successful form of cognitive partnership when it measurably increases the human learner's capacity to understand, act, and judge independently across time and context. That conclusion is a research criterion whose adequacy must continue to be tested by controlled longitudinal evidence, rather than a presumption attached to a particular artificial system.
References
- Azevedo, R., & Cromley, J. G. (2004). Does training on self-regulated learning facilitate students' learning with hypermedia? Journal of Educational Psychology, 96(3), 523–535. https://doi.org/10.1037/0022-0663.96.3.523
- Babaei-Balderlou, S., & Shakya, S. (2026). Can structured AI tutoring and study groups improve learning outcomes? Evidence from a randomized controlled trial. The Journal of Economic Education. Advance online publication. https://doi.org/10.1080/00220485.2026.2695361
- Bangert-Drowns, R. L., Kulik, C.-L. C., Kulik, J. A., & Morgan, M. T. (1991). The instructional effect of feedback in test-like events. Review of Educational Research, 61(2), 213–238. https://doi.org/10.3102/00346543061002213
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122
- Bjork, R. A., Dunlosky, J., & Kornell, N. (2013). Self-regulated learning: Beliefs, techniques, and illusions. Annual Review of Psychology, 64, 417–444. https://doi.org/10.1146/annurev-psych-113011-143823
- Butler, R. (1988). Enhancing and undermining intrinsic motivation: The effects of task-involving and ego-involving evaluation on interest and performance. British Journal of Educational Psychology, 58(1), 1–14. https://doi.org/10.1111/j.2044-8279.1988.tb00874.x
- Butler, A. C. (2010). Repeated testing produces superior transfer of learning relative to repeated studying. Journal of Experimental Psychology: Learning, Memory, and Cognition, 36(5), 1118–1133. https://doi.org/10.1037/a0019902
- Chen, S., & Cheung, A. C. K. (2025). Effect of generative artificial intelligence on university students learning outcomes: A systematic review and meta-analysis. Educational Research Review, 48, Article 100737. https://doi.org/10.1016/j.edurev.2025.100737
- Chi, M. T. H., Siler, S. A., Jeong, H., Yamauchi, T., & Hausmann, R. G. (2001). Learning from human tutoring. Cognitive Science, 25(4), 471–533. https://doi.org/10.1207/s15516709cog2504_1
- Deci, E. L., Koestner, R., & Ryan, R. M. (1999). A meta-analytic review of experiments examining the effects of extrinsic rewards on intrinsic motivation. Psychological Bulletin, 125(6), 627–668. https://doi.org/10.1037/0033-2909.125.6.627
- Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., & Willingham, D. T. (2013). Improving students' learning with effective learning techniques: Promising directions from cognitive and educational psychology. Psychological Science in the Public Interest, 14(1), 4–58. https://doi.org/10.1177/1529100612453266
- Graesser, A. C., Chipman, P., Haynes, B. C., & Olney, A. (2005). AutoTutor: An intelligent tutoring system with mixed-initiative dialogue. IEEE Transactions on Education, 48(4), 612–618. https://doi.org/10.1109/TE.2005.856149
- Han, X., Peng, H., & Liu, M. (2025). The impact of GenAI on learning outcomes: A systematic review and meta-analysis of experimental studies. Educational Research Review, 48, Article 100714. https://doi.org/10.1016/j.edurev.2025.100714
- Hattie, J., & Timperley, H. (2007). The power of feedback. Review of Educational Research, 77(1), 81–112. https://doi.org/10.3102/003465430298487
- Huang, Y., Chen, S., Zhang, W., & Chen, M. (2026). Can generative AI feedback effectively enhance learning outcomes? A meta-analysis of 36 experimental and quasi-experimental studies. Education Sciences, 16(6), Article 816. https://doi.org/10.3390/educsci16060816
- Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). AI tutoring outperforms in-class active learning: An RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15, Article 17458. https://doi.org/10.1038/s41598-025-97652-6
- Kulik, J. A., & Fletcher, J. D. (2016). Effectiveness of intelligent tutoring systems: A meta-analytic review. Review of Educational Research, 86(1), 42–78. https://doi.org/10.3102/0034654315581420
- Liu, J., Sweet, T., Chen, M. H., Engelberg, J., Masters, M., Clark, M., Persaud, A., Lancaster, A., Hollingsworth, J., & Rice, J. K. (2026). The effects of course-integrated AI tutoring on student performance and engagement: A randomized university trial (EdWorkingPaper No. 26-1598). Annenberg Institute at Brown University. https://doi.org/10.26300/y3f8-vh05
- Ma, W., Adesope, O. O., Nesbit, J. C., & Liu, Q. (2014). Intelligent tutoring systems and learning outcomes: A meta-analysis. Journal of Educational Psychology, 106(4), 901–918. https://doi.org/10.1037/a0037123
- Narciss, S. (2013). Designing and evaluating tutoring feedback strategies for digital learning environments on the basis of the interactive tutoring feedback model. Digital Education Review, 23, 7–26. https://revistes.ub.edu/index.php/der/article/view/11284
- Roediger, H. L., III, & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249–255. https://doi.org/10.1111/j.1467-9280.2006.01693.x
- Roll, I., Aleven, V., McLaren, B. M., & Koedinger, K. R. (2011). Improving students' help-seeking skills using metacognitive feedback in an intelligent tutoring system. Learning and Instruction, 21(2), 267–280. https://doi.org/10.1016/j.learninstruc.2010.07.004
- Shute, V. J. (2008). Focus on formative feedback. Review of Educational Research, 78(1), 153–189. https://doi.org/10.3102/0034654307313795
- Soderstrom, N. C., & Bjork, R. A. (2015). Learning versus performance: An integrative review. Perspectives on Psychological Science, 10(2), 176–199. https://doi.org/10.1177/1745691615569000
- Sweller, J. (1988). Cognitive load during problem solving: Effects on learning. Cognitive Science, 12(2), 257–285. https://doi.org/10.1207/s15516709cog1202_4
- VanLehn, K. (2011). The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems. Educational Psychologist, 46(4), 197–221. https://doi.org/10.1080/00461520.2011.611369
- Wood, D., Bruner, J. S., & Ross, G. (1976). The role of tutoring in problem solving. Journal of Child Psychology and Psychiatry, 17(2), 89–100. https://doi.org/10.1111/j.1469-7610.1976.tb00381.x
- Zawacki-Richter, O., Marín, V. I., Bond, M., & Gouverneur, F. (2019). Systematic review of research on artificial intelligence applications in higher education—where are the educators? International Journal of Educational Technology in Higher Education, 16, Article 39. https://doi.org/10.1186/s41239-019-0171-0