AI tutors can improve education when they help students develop knowledge and skills that remain usable without the system. Their educational value depends on the learning task, the quality and accuracy of feedback, the learner's participation, the comparison instruction, and the way achievement is measured. Research on conventional intelligent tutoring systems establishes a substantial history of favorable average outcomes. Research on newer generative AI tutors also contains promising randomized trials, alongside direct evidence that impressive assisted performance can coexist with weaker independent learning. The correct question is therefore how a particular tutor changes a particular learner's durable competence under a specified instructional design. Ma et al., 2014; Kulik and Fletcher, 2016; Bastani et al., 2025.
What Is an AI Tutor in Education?
An AI tutor is an educational system that uses artificial intelligence to provide individualized instructional interaction, such as diagnosing a learner's response, selecting an appropriate task, presenting a hint, explaining a concept, asking a follow-up question, or giving formative feedback. The defining educational function is guided learning. The label alone says little about whether the system performs that function accurately, equitably, or effectively. A successful tutor coordinates subject knowledge, information about the learner, instructional decisions, and evidence that the learner can subsequently work independently.
The term encompasses different technologies. A traditional intelligent tutoring system (ITS) may represent the domain through explicit rules, track mastery of particular skills, and prescribe a narrowly defined sequence of problems. A modern generative AI tutor often uses a large language model (LLM) to produce open-ended explanations and conversational responses. A hybrid system may connect a language model to approved curricula, worked solutions, retrieval databases, mathematical solvers, learning records, and teacher-controlled boundaries. These designs produce different error profiles and different opportunities for valid instructional evaluation. VanLehn, 2011; Latif, Liu, and Zhai, 2026.
An ordinary chatbot becomes a plausible tutor only through the actual instructional process it supports. A polite explanation, a fluent answer, or a convincing conversational tone does not establish that knowledge has been learned. Likewise, a student who completes more exercises with AI assistance may simply have a better tool for completing exercises. Researchers must observe what happens after assistance is withheld, whether students can explain their reasoning, and whether improvement transfers to unfamiliar problems. Bastani et al., 2025.
Intelligent Tutoring Systems, Generative Tutors, and Adaptive Platforms
Traditional ITS architecture frequently contains a domain or expert model, a student model, a pedagogical or tutoring model, and an interface. The domain model encodes target knowledge; the student model approximates what the learner knows or misunderstands; the tutoring model chooses the next instructional action. This structure enables detailed feedback in well-specified domains, but its usefulness depends on accurate content modeling and valid inferences from learner responses. Ma et al., 2014; VanLehn, 2011.
Generative tutors add flexible dialogue. They can rephrase an explanation, invite a student to compare approaches, generate questions, and react to incomplete answers. Their flexibility also means a response can depart from the curriculum, assert an inaccurate solution, misinterpret the student's question, or provide a final answer too early. Systems that constrain generation with verified solutions and teacher-designed prompts address some of these hazards, although the extent of risk reduction must be tested in each context. Bastani et al., 2025; Adejumo et al., 2026.
Adaptive learning platforms optimize task selection or difficulty using a learner model. Such adaptation can be an ingredient of tutoring, but personalization of sequence alone does not establish the benefits of conversational coaching or diagnostic feedback. Similarly, analytics software that predicts risk or measures engagement is a decision-support technology rather than a tutor simply because it uses AI. These distinctions matter for scientific comparisons and for assigning responsibility for errors. Latif, Liu, and Zhai, 2026.
The Scientific History of AI Tutoring
Educational tutoring technologies predate contemporary LLMs. Early computer-assisted instruction showed learners sequences of questions and corrections; later ITS research attempted to represent knowledge and misconceptions explicitly, so the instructional path could respond to student behavior. This tradition generated decades of controlled evaluations, many concentrated in mathematics, science, and other domains with well-defined answers. Its effects cannot be assigned automatically to today's open-domain chatbots: the underlying representations, safeguards, and learner interactions have changed. VanLehn, 2011; Kulik and Fletcher, 2016.
VanLehn's comparative review examined conditions under which tutoring systems and human tutors improve learning compared with other forms of instruction. A central insight is that effective instructional interaction depends on the feedback and support offered during learning, not merely on whether a human or computer is present. Subsequent meta-analyses estimated favorable average achievement effects for ITS, while revealing strong variation across controls and outcome measures. The arrival of LLM-based systems expands the design possibilities and creates a second evidence problem: claims about one tested tutor configuration are not automatically claims about the underlying model, the vendor's entire product line, or future model versions. VanLehn, 2011; Ma et al., 2014; Kulik and Fletcher, 2016.
How AI Tutors Can Support Learning
Diagnostic Feedback and the Next Useful Step
An effective tutor identifies where a learner's reasoning diverges from the target concept and responds with information that makes the next attempt more productive. Feedback can locate an error, identify a missing concept, offer a partial hint, compare procedures, or ask the student to explain a decision. The timing, specificity, and actionability of feedback influence whether it improves learning. When feedback simply supplies the correct answer, the immediate task may be completed while the student has little reason to reconstruct the reasoning. Research on educational feedback supports designing around the learner's next action, and recent AI feedback syntheses suggest substantial potential accompanied by uneven reporting of system design and validity. Hattie and Timperley, 2007; Ba et al., 2025; Maier et al., 2026.
A mathematically plausible AI hint may still diagnose the wrong misconception. A useful evaluation therefore distinguishes correctness of a final answer from correctness of the intermediate explanation, validity of a diagnosis, pedagogical appropriateness for the student's level, and success of the subsequent student response. Human review or alignment with verified domain solutions is especially important where an erroneous worked step can contaminate later reasoning. Bastani et al., 2025.
Scaffolding, Worked Examples, and Productive Effort
Instructional scaffolding supports performance that the learner could not yet achieve alone, then gradually reduces assistance as competence develops. An AI tutor can implement this principle by moving from full worked examples to partial steps, prompts for explanations, and independent retrieval. Such sequencing is a design hypothesis that must be evaluated against the learner's progress. Prolonged availability of detailed solutions can also make assistance a permanent substitute for effort. Bastani et al., 2025; Kestin et al., 2025.
The instructional objective is not to maximize struggle. Learners with insufficient prior knowledge may need explicit demonstration, manageable steps, and accessible explanations before independent practice becomes informative. More advanced learners may benefit from brief prompts, increasingly complex tasks, and opportunities to detect their own mistakes. The appropriate quantity of support depends on the relationship among task difficulty, prior knowledge, cognitive demands, and assessment objectives. VanLehn, 2011.
Retrieval, Explanation, and Transfer
Repeatedly seeing an AI-generated explanation can create familiarity without independent recall. A tutor can instead ask learners to answer from memory, generate an explanation in their own words, and apply a principle to a new situation. Classical research on retrieval practice supports the value of retrieving previously studied material for later retention. But that general result does not prove that every AI-generated quiz, prompt, or dialogue improves retention: the quality of the questions, timing, corrective feedback, and learner response remains decisive. Roediger and Karpicke, 2006.
Transfer deserves its own test. A student may solve a tutor's familiar example yet fail to recognize the same principle in a new representation or problem. Assessments should therefore include unpracticed tasks that require selection of a method, explanation of why it works, and detection of a misleading suggestion. A tutor can facilitate transfer through varied examples and explanatory prompts, but it cannot establish transfer by reporting that a learner engaged in multiple conversations. Ma et al., 2014.
Motivation, Engagement, and Perceived Support
Conversational tutors can provide quick replies, adjust apparent pace, and make practice more approachable. These features may influence willingness to try another problem or ask for clarification. An experimental course study of an instructional AI tutor found higher reported engagement and motivation alongside higher immediate learning gains than an active-learning comparison. Perceptions are educationally meaningful, but satisfaction cannot serve as a substitute for independent achievement or a measure of conceptual understanding. Kestin et al., 2025.
Learners may also attribute greater authority or understanding to a responsive system than its actual performance warrants. A confident tone can increase acceptance of an explanation, including a mistaken one. The appropriate design response is calibrated uncertainty, visible supporting evidence, and opportunities for students to challenge and verify the system, rather than presenting conversational fluency as expertise. UNESCO, 2023.
What Does Research Show About Learning Outcomes?
Evidence must be organized by technology generation, comparison condition, student population, and outcome. Combining every AI-supported educational application into a single effectiveness claim conceals important differences: software can tutor, mark, summarize, translate, sequence, answer, or write, and those activities require different measures of success.
Meta-Analyses of Traditional Intelligent Tutoring Systems
Ma and colleagues synthesized 107 effect sizes involving 14,321 participants. Compared with teacher-led large-group instruction, ITS were associated with an average effect of Hedges' g = 0.42; compared with non-intelligent computer-based instruction, g = 0.57; and compared with textbooks or workbooks, g = 0.35. The review found no statistically significant difference in the comparisons with individualized human tutoring (g = −0.11) or small-group instruction (g = 0.05). These are different comparators, and their effect estimates cannot be treated as one interchangeable measure of superiority. Ma et al., 2014.
A separate meta-analytic review by Kulik and Fletcher examined 50 controlled evaluations and reported a median test-score advantage of 0.66 standard deviations over conventional instruction. Importantly, estimated gains depended substantially on whether studies used locally constructed tests or standardized assessments. A curriculum-aligned short test may be sensitive to the specific topics a system teaches while revealing less about broad achievement or independent transfer. The favorable estimate is useful evidence for specific ITS implementations, not a guaranteed uplift for any AI tutor deployed in another subject or setting. Kulik and Fletcher, 2016.
The established conclusion is that many well-designed traditional ITS can improve measured achievement over specified alternatives. Its limits concern the concentration of studies in structured domains, characteristics of comparison instruction, possible alignment advantages, variation among systems, and relatively weak evidence about long-term effects under routine, large-scale implementation. Ma et al., 2014; Kulik and Fletcher, 2016.
K–12 Systematic Review: Positive Findings With Important Qualifications
A 2025 systematic review in npj Science of Learning considered 28 K–12 studies totaling 4,597 learners. The included studies were quasi-experimental and varied in exposure duration. Overall outcomes were generally favorable, but advantages were smaller when ITS were compared with non-intelligent tutoring systems. The authors emphasized the need for longer interventions, more diverse samples, stronger designs, and better investigation of implementation. These results support conditional educational value, with confidence constrained by study design and heterogeneity. Létourneau et al., 2025.
The age of a learner also changes the interpretation of autonomy. Younger students may struggle to recognize fabricated explanations, regulate requests for help, or protect personal information. Teacher supervision, age-appropriate tools, accessible feedback, and clear boundaries are components of an intervention rather than peripheral implementation details. UNESCO, 2023.
A Randomized University Physics Trial
Kestin and colleagues reported a randomized crossover study in a Harvard undergraduate physics course involving 194 enrolled students. The custom AI tutor was designed around teaching principles also used in the active-learning classroom comparison. In the reported lesson data, the AI-tutored group showed higher median posttest scores, and students reported greater engagement and motivation. The study offers direct causal evidence about that pedagogically designed system, course material, and instructional arrangement. Kestin et al., 2025.
Its external validity is necessarily narrower than the promise of tutoring for every student. A two-lesson crossover experiment with immediate concept tests differs from year-long classroom adoption, elementary literacy instruction, remote low-bandwidth schooling, and retention measured months later. The relevant comparison also combined the tutor's pedagogy, delivery mode, and learning setting, so the trial does not isolate a universal effect of language-model capability. Interpretation should retain those design features rather than generalize the headline beyond the study. Kestin et al., 2025.
A High School Mathematics Trial Revealing a Learning Hazard
Bastani and colleagues randomized approximately one thousand high school mathematics students by classroom to textbook-supported practice or access to one of two GPT-4-based tools. GPT Base resembled an ordinary conversation interface. GPT Tutor included teacher-provided problem solutions and prompting intended to supply hints rather than final answers. On assisted practice tasks, performance rose markedly relative to control: the study reports increases of 48% for GPT Base and 127% for GPT Tutor. These figures describe performance in the assisted exercises, not gains in independent mathematical knowledge. Bastani et al., 2025.
The decisive outcome came when AI access was removed. The GPT Base group performed about 17% worse than controls on a subsequent unassisted exam. The GPT Tutor group's unassisted score was statistically indistinguishable from the control group's, despite its substantial assisted-practice advantage. In this setting, guardrails mitigated a harmful learning effect, yet the experiment did not establish an improvement in the independent examination outcome for the guarded tutor. The causal inference applies to the trial's interventions, student population, curriculum, and assessment rather than to every possible generative tutor. Bastani et al., 2025.
This distinction resolves a major confusion in educational technology evaluation. Assistance can function as an answer-production system that raises the score of the activity being performed. Learning requires a change in the student's capabilities. Systems that make answer requests effortless may increase the former while weakening the practice necessary for the latter. A responsible tutor should therefore be assessed against both assisted completion and independent follow-up. Bastani et al., 2025.
A Teacher-Supported Trial in Nigeria
A World Bank randomized evaluation in Nigeria studied a six-week program supporting secondary-school English learning with a GPT-4-powered assistant and instructional facilitation. The report found an approximately 0.23-standard-deviation gain on English, the primary learning outcome, and a 0.31-standard-deviation gain on a broader assessment including AI knowledge and digital skills. The intervention was a package of technology, teacher support, curriculum alignment, and scheduled use; its effect cannot be cleanly attributed to software alone. The findings are especially relevant because they extend causal evidence beyond highly resourced Western university settings while leaving longer-term retention and large-scale replication as open questions. De Simone and colleagues, 2025.
Newer ChatGPT Meta-Analyses: Promising but Heterogeneous
A 2026 meta-analysis of 35 experimental and quasi-experimental studies involving 4,193 participants reported an average Hedges' g of 0.67 for ChatGPT-associated educational outcomes. The heterogeneity estimate was very high (I² approximately 91%), and the authors identified subject, duration, and instructional mode as moderators. Many included interventions were educational uses of ChatGPT rather than rigorously specified tutoring systems, so the pooled effect must not be rebranded as the average causal impact of AI tutors. Very high heterogeneity signals that a single mean offers limited prediction for a new classroom. Wu and colleagues, 2026.
Another 2026 higher-education meta-analysis assembled 66 eligible studies and 129 effect sizes and found a large average association for ChatGPT-supported undergraduate learning outcomes (g = 1.14, with a reported standard error of 0.185). That result encompasses diverse applications and measures. Its size should trigger careful examination of interventions, outcome alignment, research quality, and possible small-study effects; it is not an estimate that every educational deployment will yield a one-standard-deviation benefit. Mo et al., 2026.
The appropriate synthesis is conditional: carefully designed AI tutoring can improve short-term measured learning in particular settings; traditional ITS have a substantial positive evidence base; generative AI has both supportive trials and a clear controlled warning about independent learning; and evidence about sustained, transferable gains at population scale remains substantially less mature. Ma et al., 2014; Kestin et al., 2025; Bastani et al., 2025.
The Measurement Problem: What Counts as Learning?
Assisted Performance Versus Independent Competence
The strongest evaluation design measures outcomes under multiple conditions. First, it records how learners perform during supported practice. Second, it measures performance without the tool. Third, it introduces delayed testing to determine retention. Fourth, it checks transfer to unfamiliar or altered tasks. A fifth measure may assess error detection: can the learner reject a plausible but faulty AI answer? Each captures a different educational construct. The Bastani trial demonstrates why the second cannot be inferred from the first. Bastani et al., 2025.
Observed improvement may result from instruction, additional practice time, motivational novelty, automated answer retrieval, a more aligned test, or interaction with a human facilitator. Experimental design must identify which explanation is supported. If control learners receive fewer instructional minutes or no comparable feedback, a tutor's apparent advantage may partly reflect additional resources. An active comparator should ideally hold content, time, access to assistance, and opportunity to practice as constant as feasible. Kulik and Fletcher, 2016; Kestin et al., 2025.
Retention, Transfer, and Assessment Alignment
A near-term test built from the same item templates may exaggerate the educational significance of mastery. Delayed unassisted tests and novel item formats reveal whether the student acquired a usable method. Standardized assessment offers one form of external anchor but may be less sensitive to a short, narrowly scoped intervention. Strong evaluations use both appropriately aligned measures and independent measures that assess the intended competence. The discrepancy between local and standardized tests observed in the ITS meta-analysis is a central warning against interpreting all effect sizes as equivalent. Kulik and Fletcher, 2016.
Effect Size, Statistical Precision, and Practical Value
Hedges' g expresses the estimated difference between groups in pooled standard-deviation units, adjusted for small-sample bias. The same numerical effect may imply different consequences depending on baseline knowledge, curriculum difficulty, measurement reliability, cost, and access. A statistically significant average does not guarantee that disadvantaged groups benefit equally. Confidence intervals, classroom-cluster adjustments, prespecified outcomes, attrition, and the total cost of implementation matter for a credible school-level decision. The ITS and modern generative-AI syntheses both show why comparing effect sizes without comparator and measurement context can mislead. Ma et al., 2014; Wu and colleagues, 2026.
Accuracy: Where AI Tutoring Can Go Wrong
Factual Errors, Fabricated Steps, and Unsupported Sources
A generative tutor may supply an incorrect fact, an invalid proof, a misleading interpretation, or a fabricated reference. The problem is especially serious when errors are embedded in intermediate reasoning that a novice cannot audit. In programming education, a 2026 synthesis of 64 empirical studies discussed how hallucinated and misleading AI outputs can impose extra cognitive work during debugging and encourage uncritical reliance. Such findings are domain-specific, but they illustrate a general requirement: systems must be evaluated on the accuracy and pedagogical consequences of their content, not the fluency of the response. Adejumo et al., 2026.
Grounding a tutor in an approved syllabus, verified answers, trusted source material, or formal mathematical tools can make some claims easier to audit. It does not eliminate every failure mode. A system can retrieve correct information and then explain it badly, misjudge the learner's level, or present a valid solution at the wrong moment. Error audits should sample actual learner interactions across topics and levels, including adversarial or unusual questions, rather than relying solely on benchmark scores for general-purpose language models. Bastani et al., 2025.
Pedagogical Accuracy and Misdiagnosis
Accuracy includes judgments about what the student needs next. A tutor might incorrectly infer that a learner lacks a prerequisite, overestimate mastery from a lucky correct response, or use a generic motivational message where a precise misconception needs addressing. Valid tutor assessment therefore asks whether its feedback targets the error, whether it supports another attempt, and whether learners can detect and correct misconceptions after the interaction. The review of 129 AI-feedback studies identified generally encouraging findings but also insufficient transparency about models and instructional mechanisms, limiting reproducibility. Ba et al., 2025.
Stability, Version Changes, and Reproducibility
Commercial models and prompting configurations change. The same learner query can yield different answers across model versions, decoding settings, or retrieval context. Schools need version records, prompt and source documentation where feasible, repeatable test cases, and change-control procedures. A result obtained for a fixed 2024–2025 tutor should be reported with its configuration and cannot automatically certify a modified deployment. This follows directly from the need to evaluate a specified intervention rather than treating the label AI as a stable treatment. Latif, Liu, and Zhai, 2026; UNESCO, 2023.
Cognitive Offloading and Learner Agency
Cognitive offloading occurs when a person delegates part of a mental task to an external resource. In an educational setting, the consequences depend on which process is delegated. Offloading routine access to definitions may free attention for reasoning; delegating the reasoning itself can remove the very activity the lesson was designed to develop. The critical distinction concerns the objective: a tool that improves a finished assignment can still diminish the opportunities to practice the target skill. Bastani et al., 2025.
A productive tutor asks for an initial attempt, offers graded hints, requires the learner to articulate a choice, and eventually withdraws support. It can include a short independent checkpoint after each explanation and a delayed review without AI access. These are plausible design safeguards supported by general learning principles and the contrast between the two conditions in the high-school field experiment; their optimal sequence, duration, and effect across subjects remain empirical questions. Bastani et al., 2025; Roediger and Karpicke, 2006.
Agency also requires a student to know that a response may be wrong and to have meaningful alternatives: a textbook, teacher, peer discussion, independently checkable evidence, and a route to report inappropriate guidance. When learners are graded on AI-mediated output but never on independent competence, school incentives may reward dependence. Institutional evaluation must align assessments with the learning behaviors it wants to promote. UNESCO, 2023.
Differences Across Learners and Educational Settings
Age and Prior Knowledge
Younger students generally require closer instructional oversight, explicit privacy protections, and stronger verification of appropriateness. Prior knowledge affects whether students can evaluate a generated explanation; a novice may need bounded examples and teacher-selected tasks, while an expert learner can more effectively critique alternative solutions. Age comparisons in available reviews should be interpreted cautiously because systems, subjects, dosage, and assessment conditions also differ. Létourneau et al., 2025; UNESCO, 2023.
Subject Matter
Mathematics and programming often permit relatively clear checks on final answers, although explanations and intermediate steps can still be defective. Writing, historical interpretation, open-ended social inquiry, and language learning require additional attention to evidence, argument quality, contextual appropriateness, and rubric validity. Clinical, legal, or safety-critical training adds a need for specialized oversight and carefully validated educational content. A tutor that works on short algebra items cannot be presumed equally accurate in literary interpretation or culturally specific history. Ma et al., 2014; Adejumo et al., 2026.
Accessibility, Language, and Equity
AI tutoring may widen access to immediate explanations, translation, flexible pacing, and alternative representations. These are potential affordances whose actual benefits depend on usability, the quality of language support, device and connectivity access, accessibility compatibility, and teacher capacity. Evaluations should disaggregate outcomes where appropriate by prior achievement, disability-related accessibility needs, language proficiency, and resource availability, while respecting privacy and avoiding unwarranted profiling. The Nigeria trial illustrates a supported deployment in a different context, but also a bundle of facilitating conditions that cannot be assumed everywhere. De Simone and colleagues, 2025; UNESCO, 2023.
An intervention can improve average achievement while disadvantaging a subgroup if it gives poorer feedback in a less represented language, lacks screen-reader compatibility, or requires personal devices students do not possess. Equity evaluation should use actual subgroup performance and documented access barriers, not only general assurances that a tutor is personalized. UNESCO, 2023.
Teachers, Peers, and the Social Environment of Learning
Teacher expertise supplies curricular judgment, recognition of learner circumstances, classroom relationships, and accountability for instruction. AI support can be incorporated into teacher-led practice, homework support, or supplementary tutoring; it can also alter how learners seek help from adults and peers. A system's value must be considered within that social arrangement. The evidence for tutoring benefits cannot by itself establish that removing teachers or collaborative learning would preserve the same results. Kestin et al., 2025; UNESCO, 2023.
Privacy, Student Data, and Governance
AI tutoring often involves sensitive educational information: student identifiers, grades, writing samples, misconceptions, behavioral traces, and free-text disclosures. Even when such data are useful for personalization, their collection creates risks of excessive retention, unintended disclosure, secondary use, profiling, and unequal treatment. Schools should establish clear purposes, minimize collection, restrict access, specify retention and deletion procedures, and review provider terms and technical practices. UNESCO's guidance explicitly treats privacy, age appropriateness, and institutional validation as requirements of responsible adoption. UNESCO, 2023.
A tutor should not quietly become an automated high-stakes assessor. The European Union AI Act identifies several uses of AI in education, including certain systems for evaluating learning outcomes or influencing educational access and level, among categories of high-risk applications. Legal obligations depend on the actual use case and applicable jurisdiction; a general chat-based study helper and an admissions or grading decision system are not interchangeable. This distinction reflects both governance and scientific validity: evidence that a system supplies helpful practice feedback does not validate it for consequential selection decisions. EU AI Act, Recital 56.
Transparent institutional policies should specify approved tools, instructor responsibility, acceptable assistance during assessments, the data sent to providers, human correction and appeals processes, and procedures for investigating inaccurate or biased feedback. Privacy and equity safeguards should be implemented before collecting learner traces at scale, with monitoring throughout operation. UNESCO, 2023.
Designing and Evaluating an Educational AI Tutor
Begin With a Specific Learning Objective
Define the knowledge, procedure, or reasoning skill to be acquired, the learner population, and the setting. Record what students already know and identify the appropriate curriculum standards. The tutor's dialogue should be judged against that objective. If a unit aims to teach a student to solve linear equations independently, a system's ability to output the correct equation solution is only a content-function check; the educational outcome is the student's subsequent ability to solve and explain analogous equations. Bastani et al., 2025.
Establish Instructional Safeguards
A defensible design uses verified curricular content, teacher-reviewed worked solutions where practical, constrained and specific feedback, escalation for unsupported questions, and a preference for hints over immediate disclosure of answers when the learning task requires reasoning. The tutor should invite learner responses and gradually withdraw assistance as mastery develops. Its safeguards must be tested for actual behavior rather than inferred from a system prompt. Bastani et al., 2025; Maier et al., 2026.
Predefine the Evaluation and Comparator
For a credible causal test, preregister primary outcomes where feasible, randomize learners or classrooms appropriately, and compare with a meaningful alternative: teacher-supported practice, high-quality non-AI tutoring, another digital program, or an equivalently resourced instructional method. Record exposure time, additional teacher assistance, student adherence, attrition, and cross-group contamination. When randomizing classrooms, analyses must account for clustered assignment and the smaller effective sample size. Report effect estimates with uncertainty rather than only a significant p-value. Bastani et al., 2025; Kestin et al., 2025.
Measure Both Gains and Costs
Evaluate independent immediate performance, delayed retention, transfer, error detection, engagement, accessibility, time required, teacher workload, response quality, costs, and privacy incidents. Separate prespecified primary endpoints from exploratory outcomes. Collect independent assessments, not simply a vendor's session-completion or satisfaction metrics. Evaluate learning outcomes across student groups and monitor whether benefits persist after routine deployment. Kulik and Fletcher, 2016; Létourneau et al., 2025.
Maintain Human Accountability
Qualified educators should be able to audit the content and correct wrong guidance. Students need a clear explanation of the tool's role, realistic expectations about accuracy, and ways to obtain human support. Governance must cover vendor changes, data handling, accessibility, curriculum fit, and the distinction between formative guidance and high-stakes evaluation. A school can adopt a narrowly supported use case while withholding approval for uses whose evidence and safeguards have not been established. UNESCO, 2023; EU AI Act, Recital 56.
Evidence Status and Unresolved Research Questions
Established evidence: many traditional intelligent tutoring systems improve average measured achievement relative to particular conventional alternatives, and their effects depend on comparator and outcome choice. Established evidence also shows, in at least one rigorous high-school generative-AI field experiment, that access to an unguarded answer-generating tool can worsen independent performance despite improving supported work. These findings have direct experimental or meta-analytic support in the cited settings. Ma et al., 2014; Kulik and Fletcher, 2016; Bastani et al., 2025.
Preliminary and context-dependent evidence: teacher-aligned generative AI tutors can yield immediate learning gains, improved engagement, and useful formative feedback under tested conditions. Trials in university science and supported secondary language programs extend the evidence beyond conventional ITS, while their individual designs still limit claims about universal adoption. Kestin et al., 2025; De Simone and colleagues, 2025; Maier et al., 2026.
Limited evidence: sustained retention over semesters or years; independent transfer across distant domains; reliability for learners using less represented languages; robust causal evidence about accessibility and subgroup equity; system-wide cost effectiveness; and the effects of model upgrades. Evidence is also limited on which specific features of an LLM tutor causally improve outcomes when all other instructional resources are held constant. Létourneau et al., 2025; Latif, Liu, and Zhai, 2026.
Contested generalizations: that an AI tutor can replace a teacher across a full curriculum; that a conversational system reliably matches individualized human tutoring; that positive pooled effects from broad ChatGPT use estimate the effect of guided tutoring; or that personalized conversation necessarily yields deeper comprehension. These propositions require direct designs with credible controls, independent assessment, transparent intervention descriptions, and replication across settings. VanLehn, 2011; Wu and colleagues, 2026.
Research should now compare guarded and unguarded systems under equivalent time and content, test delayed examinations, evaluate diagnostic and explanatory accuracy on actual student interactions, study independent learners and teacher-supported deployment separately, and report costs, privacy safeguards, accessibility, and adverse educational effects. Scientific progress will depend on identifying which combinations of pedagogy, subject knowledge, model design, and social support change learning, rather than treating the AI label as a single intervention.
Frequently Asked Questions
Do AI tutors improve grades?
Some conventional and generative tutoring systems improve scores in controlled studies. The result depends on the test, comparison instruction, implementation, and whether the AI remains available during assessment. Higher assisted practice grades alone do not demonstrate improved independent learning; a high-school trial documented the reverse pattern for an unguarded generative assistant. Ma et al., 2014; Bastani et al., 2025.
Are AI tutors as effective as human tutors?
Some older ITS comparisons in meta-analytic evidence did not show a statistically significant difference from individualized human tutoring, but the studied systems, subjects, and controls were specific. Absence of statistical significance does not prove universal equivalence, and newer conversational tutors require their own head-to-head trials. Ma et al., 2014; VanLehn, 2011.
Can an AI tutor give incorrect explanations?
Yes. Generative tutors may produce factual, procedural, interpretive, or citation errors. Verified course material, teacher-reviewed solutions, constrained prompting, and human escalation can reduce some risks, but claims about accuracy require ongoing testing on authentic student questions. Bastani et al., 2025; Adejumo et al., 2026.
How should students use an AI tutor without becoming dependent?
Start with an independent attempt, request a hint or explanation of an error, explain the solution back in your own words, and solve a similar problem without the tool. Periodically test recall and transfer after a delay. These practices preserve the student's active role and directly address the problem of assisted success without durable competence. Bastani et al., 2025; Roediger and Karpicke, 2006.
Should schools use AI tutors for grading or admissions?
Evidence that a tutor provides helpful formative practice does not establish that it makes valid, fair high-stakes decisions. Consequential assessment requires separate validity, fairness, accountability, and legal review. Some educational assessment and access uses are addressed as high-risk categories under the EU AI Act. EU AI Act, Recital 56.
Related Articles
Cluster Learning and Educational Psychology — the disciplinary research chapter on learning mechanisms, educational theories, assessment, and instruction.
Education in the Age of AI: Learning, Motivation, Assessment, and Cognitive Development — the broader historical and psychological context for educational change in the Artificial Era.
Retrieval Practice: Why Testing Yourself Improves Memory — the established memory-learning mechanism relevant to independent tutor-supported recall.
Spaced Repetition: How the Spacing Effect Supports Long-Term Memory — the canonical account of distributed practice and durable retention.
Learning Analytics in Education: Prediction, Bias, and Student Privacy — Explore prediction models, educational data quality, privacy, bias, and the limits of intervention inference.
Generative AI for Homework: Learning, Dependence, and Cognitive Offloading — Separate assignment completion from genuine learning and assess cognitive offloading and verification practices.
Digital Distraction During Studying: Multitasking and Learning Costs — Examine notifications, media multitasking, divided attention, and experimental learning-outcome estimates.
Adaptive Learning Systems: Personalization and Research Evidence — Review algorithmic personalization, sequencing, learner models, outcomes, equity, and implementation.
References
Adejumo, A., Oyelere, S. S., Sanusi, I. T., and Suhonen, J. (2026). A systematic review of the impact of GenAI on learning performance, AI hallucinations, and problem-solving in computer science education. Computers and Education: Artificial Intelligence, 10, 100570. https://doi.org/10.1016/j.caeai.2026.100570
Ba, S., Yang, L., Yan, Z., Looi, C. K., and Gašević, D. (2025). Unraveling the mechanisms and effectiveness of AI-assisted feedback in education: A systematic literature review. Computers and Education Open, 9, 100284. https://doi.org/10.1016/j.caeo.2025.100284
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., and Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences. https://doi.org/10.1073/pnas.2422633122
De Simone, M. E., Tiberti, F. H., Barron Rodriguez, M., Manolio, F., Mosuro, W., and Dikoru, E. J. (2025). From chalkboards to chatbots: Evaluating the impact of generative AI on learning outcomes in Nigeria. World Bank Policy Research Working Paper 11125. World Bank reproducibility and study record
European Union. (2024). Regulation (EU) 2024/1689 (Artificial Intelligence Act), Recital 56. Official AI Act Service Desk
Hattie, J., and Timperley, H. (2007). The power of feedback. Review of Educational Research, 77(1), 81–112. https://doi.org/10.3102/003465430298487
Kestin, G., Miller, K., Klales, A., Milbourne, T., and 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, 17458. https://doi.org/10.1038/s41598-025-97652-6
Kulik, J. A., and 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
Latif, E., Liu, V., and Zhai, X. (2026). A systematic review of intelligent and robot tutoring systems: Evolution, pedagogical design, and AI-driven classification. Smart Learning Environments, 13, 1. https://doi.org/10.1186/s40561-025-00427-9
Létourneau, A., Deslandes Martineau, M., Charland, P., Karran, J. A., Boasen, J., and Léger, P. M. (2025). A systematic review of AI-driven intelligent tutoring systems (ITS) in K–12 education. npj Science of Learning, 10, 29. https://doi.org/10.1038/s41539-025-00320-7
Ma, W., Adesope, O. O., Nesbit, J. C., and 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
Maier, U., Seibold, M., and Klotz, C. (2026). LLM-generated formative feedback in education: A qualitative systematic literature review. Computers and Education Open, 10, 100374. https://doi.org/10.1016/j.caeo.2026.100374
Mo, F., Huang, J., Yang, Y., Özen, Z., Maeda, Y., and Olenchak, F. R. (2026). Undergraduate students' learning outcomes with ChatGPT: A meta-analytic study. Computers and Education: Artificial Intelligence, 10, 100536. https://doi.org/10.1016/j.caeai.2025.100536
Roediger, H. L., III, and 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
UNESCO. (2023). Guidance for generative AI in education and research. UNESCO. Official publication and guidance
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
Wu, X., Zhu, P., Zhang, J., Yin, M., and Wang, Y. (2026). ChatGPT's impact on student learning outcomes: A meta-analysis of 35 experimental studies. Humanities and Social Sciences Communications, 13, 684. https://doi.org/10.1057/s41599-026-07019-z