Skip to content

AI Tutoring

Pattern · All patterns
Evidence · 329 claims (152 for, 127 mixed, 50 against) · 222 studies (80 causal, 47 review, 27 design, 25 theoretical, 19 quant-synthesis, 13 associational, 11 qualitative), q1–q4 · 46 of 222 report an effect size · 314 claims rest on one study

Description and scope

Tutoring delivered by software rather than a person: step-based intelligent tutoring systems that track a learner's work on each step and give hints and feedback, and conversational tutors built on large language models that answer, question and explain in dialogue. The decisions it covers are how much the tutor gives away (answers, hints, Socratic questions), how it keeps the learner doing the thinking, how it checks understanding rather than accepting a correct-looking answer, how it guards against over-reliance and wrong answers, and when it refers the learner to a human tutor, mentor or instructor.

It draws on three older literatures that conversational AI tutors inherit: pedagogical agents (animated or embodied on-screen characters that explain, model, give feedback and motivate), teachable agents (learning by teaching a computer agent that learns from the learner, as in Betty's Brain, and the protégé effect), and computer-supported collaborative learning (CSCL), where software scripts, scaffolds or takes part in a learning conversation as a peer or partner.

The study configurations below are evidence; the response-dependent policy is an untested design proposal. Use it to gather better evidence about a learner group and a tutor configuration, not to certify that a tutor "works" or that a learner has mastered something because the tutor says so.

What the evidence does and does not test. The wiki's strongest evidence is on step-based intelligent tutoring systems against conventional instruction: three meta-analyses (Ma et al. 2014, k = 107; Kulik & Fletcher 2016, 50 evaluations; VanLehn 2011) agree on a moderate advantage over classroom instruction, textbooks and other computer-based instruction and no detected difference from one-to-one human tutoring, with much smaller effects on standardized than on locally developed tests. For LLM tutors the wiki records a handful of randomized studies, mostly single-session and immediate: an unguarded GPT-4 interface raised practice scores and lowered later unaided exam scores, while a guardrailed tutor removed the harm without a detected gain (Bastani et al. 2025); a research-designed physics tutor beat an in-class active-learning lesson on an immediate post-test (Kestin et al. 2025); over five weeks, a structured tutor raised a post-test over a holdout group free to use consumer chatbots, in a small preregistered field experiment (Yang et al. 2026; 86 randomised, 46 analysed); and with about 2,400 adults practising GRE items, AI tutoring and expert human tutoring each beat no tutoring by similar amounts in roughly one session each (Northcutt et al. 2026). No study recorded here follows an LLM tutor across a whole course to a delayed or transfer outcome. Pedagogical-agent evidence rests largely on reviews of older multimedia experiments; teachable-agent evidence on one meta-analysis of learning by teaching and a few small studies; CSCL evidence here is two meta-analyses of collaboration scripts by one research group (Vogel et al. 2016; Radkowitsch et al. 2020): a small effect on domain learning and a larger one on collaboration skills, against unscripted collaboration, with no AI partner in the studies. So the configuration this page defends is narrow: a tutor that works step by step, withholds final answers by default, escalates help contingently, is judged on unaided and delayed performance, and has a named human route.

Mastery learning decides when a learner moves on; self-regulated learning concerns the learner's own planning and monitoring; practice with simulated people covers AI role-play partners and simulated patients or clients; mentor support in self-paced online learning covers the human side of the handoff. This pattern is about the tutor's moves inside a task.

Inputs: establish the design brief

Field Record or elicit
Learner and starting observation Task-specific starting knowledge (novice, intermediate, advanced, or ?); an unaided attempt at a representative task, with reasoning; prior use of AI tools and how the learner says they use them (for answers, hints, checking). Preserve unknowns.
Context and activities Setting and mode (classroom, online-instructor-led, online-self-paced, workplace-clinical); whether a human tutor, mentor or instructor is reachable and how fast; time per session and number of sessions; the exact tutor configuration: model or system, system prompt or tutoring rules, answer-withholding rules, hint ladder, problem-specific content given to the tutor, agent persona or voice, logging.
Learner-valued goal Ask what the learner wants from the tutor (finish the assignment, pass the assessment, be able to do this at work) and why. Record disagreement with the designer objective: a learner who wants answers and a designer who wants independent skill want different tutors, and the evidence below says which one each will get.
Designer objective The capability (procedure, concept, complex-skill, judgement), whether the domain is well structured (steps can be checked) or ill structured, and which aids are permitted when the capability is assessed.
Outcome Separate measures for assisted performance during practice, unaided performance at the end, a delayed unaided measure and, where it matters, transfer to a changed task. Say whether the test is developed locally (close to the tutored problems) or standardized. Add a reliance or verification measure if over-reliance is a risk.

A request to "add an AI tutor" does not specify these inputs. Ask for them before choosing a tutor configuration or predicting a gain. Check first that the tutor's answers on the course's own problems are correct: in the wiki, LLM-generated hints were wrong, too general or gave the answer away about a third of the time in one coding study, and expert reviewers discarded a quarter to nearly half of generated hints [~M] in another. Do not take a learner's stated intention as a forecast of how they will use the tutor: among 98 adolescents, learning-oriented intentions did not predict enacted help-seeking [~M], while those who asked for "just the final answers" did request more answers. Read the log as well as asking.

Sequence and conditional policy

  1. Elicit an unaided attempt first. Before the tutor opens, ask for the learner's own attempt and reasoning on a representative task. Keep it short and record it as a learning exposure. This is the baseline against which tutored practice is later read.
  2. Configure the tutor for the objective, not for helpfulness. For a structured task, set the tutor to work step by step, to withhold final answers by default, and to give problem-specific hints (teacher- or designer-written content in the prompt, or a validated hint bank). Ground it in the course's own materials (slides, notes, worked solutions) and have it say which one it is drawing on; a tutor answering from general knowledge invents content (see the grounding row below). Start contingent: the least help that moves the learner, beginning with a cue or leading question, more only after a failed attempt, less after success. Do not pass identity labels about the learner (group membership, a disability or motivation label) into the tutor's prompt; personalise from the learner's work instead. Test the tutor on the course's own problems before learners see it, and do not judge it by how helpful its turns are rated: turns that leaked the answer were rated more helpful and followed by less independent work [~M]. Include in that test a learner who insists on the answer: in a teacher-configured K-12 tutor, direct requests for answers doubled the rate of direct answers (26.1% against 12.5%) [~M], although the tutor kept scaffolding in most turns. Show the learner the tutor's answer or a short summary of its reasoning, not a full reasoning trace: in a preregistered experiment with 559 adults, full traces lowered AI-assisted reasoning performance against answer-only (d = −0.54) and summary traces (d = −0.48) [-M] while raising trust in the AI [~M].
  3. Tutored practice, dose by session. Early sessions (the first one or two in a unit): 20–40 minutes of tutored step-based practice with hints available and an explanation request at each step. Middle sessions: raise the threshold for help (a hint only after an attempt; no worked step without a learner explanation) and add prompts to check the tutor's claims. Late sessions: tutor off for at least one unaided practice set before the assessment. Fade per skill on the learner's own performance rather than by calendar where the tutor can track skills: in a step-based geometry tutor, fading worked steps adaptively, skill by skill, beat fixed fading and plain tutored problem solving [+W] (on the delayed test in the classroom replication). Move a learner on when unaided work on a changed problem meets the agreed criterion, not when tutored practice does. The session doses are a proposal; that one claim tests fading inside a step-based tutor, and none tests a dose or a fading schedule for a conversational AI tutor.
  4. Read the log, not only the score. Record what was asked (answer, hint, explanation, check), how many hints, whether the full answer was given, and how often the learner verified the tutor. Where the tutor tracks skills, read accuracy per skill, not overall: overall percent correct can stay flat while each skill's accuracy rises, because later tasks are harder [~M], and a skill with no rising curve points to a fault in the tutor's model of the skill, as log analysis found in an algebra tutor [+M] and as a data-derived model that improved mastery when used to redesign a geometry unit [+M] illustrates. Watch for learners who never signal confusion: in one tutor deployment a "passive consumer" pattern was visible only in longitudinal analytics [~M], not in any single session. Use the table below.
  5. Hand off to a human on named triggers. Repeated impasse at the same step, a learner asking for the answer repeatedly, a tutor error the learner accepted, distress, or a question the tutor cannot answer. Name who takes over and how fast. As a proposed default, change approach after a second failed turn at the same impasse and hand off after about three: in 6,630 impasse turns with guided LLM tutors, next-turn recovery fell from 39.7% at onset to 26.2% at the third turn and 18.8% from the sixth, about 13% lower odds with each turn [~M], and help-seeking impasses fell most steeply once the first help had failed (an association in logs, not a test of a trigger). Two studies recorded here support human involvement alongside software; none tests a specific trigger.
  6. Reobserve unaided and later. Score an unaided, changed task at the end of the unit and again after a delay, separately from tutored practice. If tutored practice is high and unaided performance is not, treat it as a configuration problem, not as learning.
Observation Discriminating probe Proposed action, with its uncertainty
Tutored practice scores high, unaided scores flat or lower Compare the log: share of turns asking for the answer or receiving it; give one matched unaided item mid-unit. If answer-taking dominates, tighten answer-withholding and require an attempt before a hint. This is the pattern Bastani et al. observed with an unguarded interface; the guarded configuration removed the harm but did not show a gain, so expect no more than parity until an unaided measure shows otherwise.
Learner asks for the answer repeatedly, or skips Ask what they are trying to get done and by when; check whether the task is beyond reach with the current hints; check where in the session it happens. If the task is out of reach, step back (a worked example or a simpler step) rather than release the answer. If the goal is completion under time pressure, negotiate the goal; refer to a human mentor if it persists. Expect drift: within a session, requests shifted from conceptual and procedural help toward answers and verification [~M], so hold the answer-withholding rule late in a session, not only early. Expect it most from learners who are struggling: in 482 biology students' trace data, the lowest-scoring quartile opened worked solutions more often while hint use did not differ by quartile [~M] (small effect, associational). The branch itself is untested.
Learner accepts a wrong tutor answer Ask them to check the step against a source or a second method. Add a short reflection or verification prompt; warning labels alone did not reduce reliance in one study recorded here. Do not rely on learners checking unprompted: in one classroom log verification was 16% of requests on average and the median learner made none [~M], although most had said they wanted the AI to check their understanding. Log the error and fix the tutor's content.
Learner rarely asks for help but fails unaided Compare hint use with errors; ask whether they avoided help. Expect it: most students in one course said they sometimes avoided asking for help when they needed it [~M]. A warning that the tutor can be wrong raised hint requests in one classroom study; a metacognitive help-seeking tutor improved some kinds of help-seeking [~M] but produced no domain learning gain [~M] in another. Prompt help at impasse rather than waiting to be asked: proactive feedback raised feedback uptake [+M] in a within-subjects pair-programming study, and in a randomised distance-education experiment with 120 learners a tutor that reached out raised use of lexical [+M] and metacognitive help tools [+M] whether the tutor was a person or the computer, while use of the conceptual help tool did not change [~M]. Still an untested branch for learning outcomes.
Learner stops asking the tutor questions Look back at what the tutor answered when they did ask: relevant, correct? Route open questions to a human or a curated source and fix the tutor's grounding. In AutoTutor, answers to student questions were often irrelevant or wrong, and student questions extinguished quickly [-M].
Correct with hints at the first step, stuck later Probe the later step alone. Escalate the hint ladder for that step only; change approach (address the error) rather than repeat the same question: in one tutor-log study a scripted question repeated after a failed one halved the odds of recovery [-M]. If the impasse persists past about three turns, hand off (step 5).
Fluent explanation in dialogue, weak on a changed problem A changed problem without the tutor. Add varied unaided practice; the dialogue may have carried the reasoning. Untested branch.

Choosing configurations from evidence

Do not rank these results against one another: comparators (no tutoring, textbooks, classroom instruction, active learning, an unguarded chatbot), outcomes (practice, immediate unaided, standardized, transfer) and learners differ. A precise abstention names the missing comparison, outcome or observation and the next way to resolve it.

Fitting the design to a situation

Ask for four facts when the brief omits them: is the task structured enough for the tutor to check each step?, what is assessed, with what aids, and when (immediately, at the end of the course, later at work)?, is a human tutor, mentor or instructor reachable, and how quickly?, and what will the tutor actually say, tested on the course's own problems?

If the brief says… Then change… Basis
Learners are novices in the domain Step-based tutoring with contingent hints and worked steps available; a coach or tutor agent rather than open chat Agents as coaches and tutors are most useful with novices [~M]; Guidance that helps novices becomes redundant with expertise [~S]
Learners are already competent and practising for fluency or transfer Reduce step-level hinting; use the tutor for challenge problems, explanation checking or as a learner to teach Guidance that helps novices becomes redundant with expertise [~S]; Learning by teaching [+S]
Adult or professional learners (higher education, workforce) Keep the answer-withholding default; most LLM-tutor trials recorded here were undergraduates or high-school students, and the harm-after-withdrawal result was in high school. Leave decisions with the learner: adult learners asked for AI that points to what to improve rather than giving the final answer GPT-4 access harmed the unaided exam [-M] (high school); AI tutor beat active learning [+M] (undergraduates); ITS fostered declarative and procedural knowledge in vocational education [+M] (systematic review, no pooled effect); adult learners want human-in-the-loop AI [+M] (focus groups)
Children or young adolescents Prefer a structured tutor or teachable agent with teacher oversight to open chat; test the tutor's safety over long multi-turn conversations and against requests to cheat, not only single prompts, and do not rely on an off-the-shelf safety filter Teachable agent improved fourth-graders' science reasoning [+M]; warning about AI mistakes raised seventh-graders' hint requests [+M]; multi-turn interactions raised attack success on conventional risks three to five times [-M] and safety classifiers passed exam-answer and cheating requests as safe [-M] (model evaluations, no learners)
The goal is a procedure in a well-structured domain (mathematics, physics problems, programming) A step-based tutor with problem-specific hints written or checked by a teacher Guardrailed tutor with teacher-designed hints [+M]; ITS improve achievement [+S]; ITS help especially where knowledge is structured [+M] (review, no effect size)
The goal is an ill-structured judgement (writing, design, ethics, clinical reasoning) Use the tutor for questioning and critique of the learner's draft, not for producing it; unlock the AI only after the learner's own attempt; an AI-improved version of the learner's own practice draft, after the attempt, can teach; assess independent work Requiring engagement before the AI unlocked cut verbatim AI text in final essays (27% against 47%, d = −0.47) [+M]; AI-assisted cover-letter practice improved unaided writing a day later [+M] (after a lesson on the principles); Bounded support gave the best independent writing [~M]; originality gains vanished unaided [~M]; an LLM-assisted essay group did worse than a no-tool group across four sessions [-M] (small preprint, 18 completed the fourth session); scaffolded GenAI augmented rather than replaced judgement in a design-thinking course [+M] (student reflections)
The assessment is unaided, or delayed, or at work Schedule tutor-off practice before it; measure unaided performance separately from tutored practice; do not forecast from studies whose test allowed the AI Practice up, unaided exam down [-M]; Effects small on standardized tests [~M]; g = 0.76 with AI allowed in the test, −0.06 without [~M] (programming)
Online self-paced or asynchronous, no instructor watching Answer-withholding by default, a hint ladder, logged reliance signals, and a named human handoff with a response time; add reflection prompts before accepting the tutor's answer Reflection prompt cut acceptance of wrong advice [+M]; Human-AI tutoring raised MAP growth over AI alone [+M]
Learners will use a general chatbot anyway Offer a structured tutor as well, and teach a structured way to use any chatbot (attempt first, ask for hints, check claims) rather than relying on a warning; make the learner's own reliance visible to them A structured tutor beat a holdout group that freely used consumer chatbots, over five weeks [+M] (small field experiment); Structured prompting improved explanations; unprompted access did not [+M]; warning labels insufficient [-M]; students named awareness of their reliance as the most useful outcome [+M] (six interviews, programming course)
Short time: one session or a single lesson A research-designed tutor can replace a lesson for an immediate outcome; do not extrapolate to a course AI tutor beat active learning in less time [+M] (one lesson, immediate test)
A course of many weeks Fade tutor help across sessions on each learner's performance, add tutor-off checkpoints, and plan the curriculum the tutor follows rather than leaving it to the tutor; no claim here follows an LLM tutor over a course with real learners Adaptive fading beat fixed fading in a step-based tutor [+W]; tutoring agents' simulated students plateaued within 5–10 of 30 days, mostly for lack of a curriculum [-M] (simulated learners only)
Human tutors or mentors are available but stretched Give them AI suggestions in real time. Routing learners below a threshold to proactive human contact is plausible but not shown to beat on-request help; if you do it, compare the two Tutor CoPilot raised mastery 4 points [+S]; Proactive tutoring had larger relative growth [~M]; no significant proactive–reactive difference at the cutoff [~M]
No human is available at all Narrow the tutor's scope to checkable tasks and accept that it may underperform a human at helping Conversational agents were well behind humans at helping the student [~M]
High stakes (certification, patient or client safety) Do not let the tutor's judgement of mastery stand alone; validate its content and its errors, and assess with a human-scored or proctored task GPT-4 feedback flawed in about 35% of hints [~M]; convincing incorrect explanations [-M]
A voice tutor is planned instead of text Choose voice to buy engagement and questioning, text to buy the same measured mastery for less; measure both if it matters Voice and text gave equivalent weekly mastery; voice doubled turns and questions at 2.8 times the cost [~M] (within-student, one five-week field experiment)
Learners writing in a second language want feedback on drafts AI feedback can extend a human tutor's feedback; keep the human for follow-up questions and spoken practice, and measure progress rather than assuming parity GPT-4 and human tutor feedback produced no detected difference over six weeks [~M] (equivalence not tested); students preferring the human cited follow-up questions and speaking practice [~M]
An animated or avatar agent is planned Spend effort on what the agent says and its voice before its face; avoid adding visual load on complex content Talking head added little [~M]; voice, not visual presence [~M]; visual-channel overload [-M]
Group work with an AI participant Give the AI a bounded role (tutor or critic) and have the group verify its output together; do not use one AI agent to check another; time its prompts to pauses rather than interrupting Tutor plus peers gave the best unaided accuracy [+M]; conversational-agent prompts improved individual and pair outcomes [+W] but were felt as disruptive [~W] (one experiment, reported second-hand); judge agents propagated peer errors [-M]; groups verifying chatbot solutions together supported problem solving [+M] (exploratory, grade 12 physics)
Health professions (clinical interviewing, diagnosis) Use multi-agent simulated patients with scaffolding for interview practice, with faculty review of errors; test long dialogues, where role-play failures cluster Multi-agent simulated-patient training raised OSCE-aligned scores (71.8% vs 55.6%) [+M]; role-play failures rise after turn 6 [-M]
Teacher education Use the tutor to support novice tutors or teachers in the moment (suggested moves), not to replace them Tutor CoPilot [+S] (tutors of K-12 students; not tested with teacher candidates)
Business or management courses (case analysis, decisions) Use the tutor to question the learner's analysis and to check claims against case data; keep the decision and justification the learner's Checking AI claims against sources predicted decision-making gains [+M] (secondary school, socio-scientific issues; carried to business untested)
Technology and computing courses Configure hint-not-solution behaviour and test it; small models frequently give direct code solutions Hint-not-solution was the hardest dimension, 6% to 81% across models [-M]; learning by teaching an LLM improved assignments, not exams [~M]; developers who used AI assistance scored 17% lower on a later comprehension test [-M] (cited study, steepest for full delegation); in a project course, students who coded without AI kept high code comprehension while the heaviest AI users showed little [-M] (descriptive, instructor-rated)
Learners with low self-regulation Add more external structure (planning prompts, check-ins, hint budgets) rather than open access Gen-AI effect on higher-order thinking was larger for high-SRL students (ES 0.86 vs 0.28) [~M]
The tutor would answer from its general knowledge, with no course materials given to it Ground it in the course's slides, notes and worked solutions, have it cite which it used, and tune retrieval per discipline ungrounded ChatGPT self-study taught 6 of 23 students a fabricated answer [-M]; a robot tutor given the lecture slides cited them and made no false statements, while the control made ten [+M] (six transcripts); students and lecturers want the tutor to have the official materials [+M]; the best retrieval settings differed between STEM and humanities [+M] (automated evaluation)
The tutor is to be given learner profile labels (group membership, disability, "unmotivated") Leave the labels out of the prompt; personalise from the learner's work, and audit the tutor's feedback across groups for less criticism and more praise LLM feedback gave students marked by group or label less constructive criticism and more praise for the same essays [-M] — a bias in the model's treatment of the label, not a property of the learners
Learners receiving special-education services or learning in a second language Keep a step-based tutor inside teacher-led instruction and track these learners' unaided results separately; the evidence is a district comparison, not a trial higher state-test pass rates for Cognitive Tutor students receiving special education (35.7% against 10.9%) and with limited English (27% against 18.9%) [+M] (no effect size, not randomised); potential for marginalised students, contingent on safeguards [~M] (review)
Young readers practising oral reading A speech-recognition tutor that listens and supplies a word or hint at a miscue can give practice time a human cannot; keep a teacher reading the reports speech-recognition tutoring supports word identification, attention and motivation [+M]; time efficiencies over human tutoring [+M] (both from a vendor evidence summary, no effect sizes)
Learners are disengaged, or game the hints Personalise problem contexts to learners' interests; when the log shows gaming, give the skipped material as supplementary exercises; check engagement over weeks, not only in the first session an agent that responded to detected gaming with exercises on the bypassed material halved gaming and improved those students' learning [+W] (one classroom study, as a review reports it); interest-personalised problems reduced gaming (d = 0.35) [+M] and raised accuracy on the hardest skills [+M] (ninth-grade algebra, one randomised study); staff of a technology-driven tutoring programme reported engagement as a challenge [-M]

The evidence by question

Further evidence, not yet read against this model

Claims linked to this pattern that the model above does not use. All were read against the model on 2026-10-19; each note says why it stays here (most restate a study the model already cites, test a tutor's outputs rather than learners' learning, or argue rather than measure).

Intelligent tutoring systems - Meta-analytic evidence indicates ITS yield positive outcomes [+M] — a review restating Ma et al. (2014), already used. - ITS improve upon typical classroom instruction [+M] — a review's summary sentence; the meta-analyses above carry the comparison. - ITS improve outcomes particularly with immediate actionable feedback [+M] — a position paper's reading of VanLehn and Kulik & Fletcher; the feedback bullet uses the primary evidence. - ITS are a mature implementation of personalized learning [+M] — a review's characterisation, no comparison. - Personalized learning environments outperform standardized settings [+W] — second-hand review assertion with no design described. - AutoTutor achieved gains comparable to human tutors [+M] — restates the AutoTutor result already used. - Human tutoring d ≈ 0.79 (VanLehn, as reported) [+M] — restates VanLehn (2011), already used. - Individualized BKT reduced questions needed for mastery [+M] — about knowledge tracing for mastery decisions; belongs with mastery learning; second-hand, no design reported. - Classroom-scale data is insufficient for tutor-model misconception acquisition [-M] — about training a model of misconceptions, no learners. - Tutor models trained on ten misconceptions keep accuracy [+M] — same study, model accuracy only. - The field moved toward step-based problem solving and collaboration support [+M] — a description of research trends, not of effects.

LLM and generative-AI tutors - Reviews citing Bastani et al. report harm after withdrawal of GPT-4 access [-M] — second-hand accounts of Bastani et al. (2025), whose own claims the model uses. - Pedagogically designed AI tutors outperformed active learning (as reported) [+M] — restates Kestin et al. (2025), already used. - An engineered AI tutor doubled learning versus active learning (as reported) [+M] — restates Kestin et al. (2025). - A prompt-engineered chat tutor improved performance in a physics course (as reported) [+M] — restates Kestin et al. (2025). - GAI value is contingent on pedagogical integration [~M] — a narrative summary that agrees with the model; no new data. - Immediate performance is not evidence of independent learning [-W] — a review's reading that agrees with the model's horizon rule (the marker records polarity toward AI tutoring, not toward this page's policy); no new data. - AI substitution may impede long-term skill development [-W] — an argument; the model rests the same point on trials. - LLM teaching effectiveness varies by domain [~M] — LLM teacher agents teaching simulated students; suggests testing per domain, no learners. - Adaptive mode showed a warm-up per-question increase [+M] — a within-session detail of the 50-person chatbot experiment already used. - AI hallucination as a pedagogical resource [+W] — an argument for teaching verification; the model's verification prompts rest on measured results. - A review reports over-reliance on AI dialogue systems may weaken decision-making [-W] — two small qualitative studies reported in a review; agrees with the model's default. - Reviewed literature reports over-reliance, errors and uncritical trust in AI for mathematics [-M] — a review's risk list; agrees with the model's default. - Excessive GenAI reliance risks undermining foundational knowledge [-M] — a document analysis recommending supervised assessment of foundational computing knowledge; consistent with the unaided-assessment row. - Overreliance on automated writing evaluation may lower critical thinking [-W] — about automated writing evaluation, a review's risk statement. - LLM cognitive debt weakens critical thinking [-W] — an argument citing the essay-writing study the model uses directly. - Reactive approaches to cognitive debt are insufficient [-W] — an argument; no data. - Contradictory meta-reviews on AI and critical thinking [~M] — notes design limits in that literature; no effect. - Brain connectivity scales down with external support in essay writing [-M] — a neural measure from the essay study whose performance result the model uses; not a learning outcome. - AI substitution: inclusion and time trade-offs [~W] — a conjecture, not tested. - Placebo effect of perceived AI capabilities [~M] — cited HCI work, not tested in a learning setting here. - Verification visibility interventions reduce overreliance in a model [+M] — an agent-based simulation, no learners; a hypothesis for the group-work row. - Student reflections on a copilot comment workflow [~M] — perceptions of a code-generation exercise, not tutoring. - Over-reliance on AI scaffolding risks learned helplessness [-W] — an argument; the fading step already answers it. - Adaptive learning dominance limits student agency [-W] — an argument; no data. - GenAI personalization and bias tensions [~W] — framing only, no evidence. - Students exercise ethical agency in GenAI use [+M] — about students' ethics of AI use, not tutoring moves. - Four exploratory student AI-user profiles [~M] — exploratory survey clusters of attitudes; no outcome. - An agentic AI instructor sustained engagement in a field deployment [+W] — a qualitative report of engagement, no comparison. - Embedded AI let a novice troubleshoot in real time [+W] — an anecdote. - AI in university physics serves six pedagogical functions [+M] — a review's description of uses, no effects. - Phase VI AI systems face six challenges [-W] — agrees with the gap the page states: controlled studies of transfer over time remain scarce.

Building and evaluating LLM tutors - Emotion features help only with theory-grounded strategies [~M] — a prompt ablation scored on model outputs, no learners. - A task-specific prompt beat PedCoT prompting [+M] — model accuracy on 61 test cases, no learners. - Prompt design can generate cognitively activating feedback [+M] — a demonstration; no learning outcome measured. - GPT-4 feedback was useful for identifying scoring errors [+M] — about AI grading of exams, outside this page's scope. - Automated verification of AI-drafted content gave mixed results [~M] — about AI-generated course content, outside this page's scope. - RL reward misalignment failure modes [-W] — an analysis of risks (hints creating dependency, easy content chosen for engagement); agrees with judging on unaided outcomes; no data. - Within-session RL improved response quality [+M] — response-quality ratings over 20 conversations per condition, not learning. - Teachers used Purpose for goals and Rules for pedagogical behaviour [+M] — how 27 teachers configured a chatbot; useful when recording the tutor configuration, no outcome. - Higher pedagogical scores associated with exercise conversion [+M] — a small same-day association in a live deployment; not strong enough to set a configuration. - Equation handling dominates AI tutor discourse [+M] — a description of what students do in one physics tutor. - Scripted prompts nearly eliminate persona drift [+M] — LLM persona consistency, no learners. - Systems adapt perceptibly to a learner persona [+M] — raters judging generated videos, no learning outcome. - Adaptive sessions converge on simulated mastery [+M] — simulated personas only. - Socratic test pilot: conversation forced explanation [+M] — a conversational exam (assessment, not tutoring); student perceptions only. - Socratic test pilot: students accepted the scaffolding [+M] — same pilot, perceptions only. - Socratic test: one pilot instructor saw differentiation [+W] — one instructor's report. - Knowledge-based design robot perceived as more responsible [+M] — perception with 12 per group; the same study's transcript result is used in the grounding row. - Knowledge-based design: no difference in acceptance or motivation [~M] — same small study; no detected difference, equivalence not tested. - Twelve design requirements for tutoring social robots [+M] — interview-derived requirements; the materials requirement is used in the grounding row. - A learning-assistant pilot rated chat and quiz modes positively [+M] — eight students' ratings. - MAIC script generation outscored baselines [+M] — generated lecture scripts, outside this page's scope. - Computer-generated self-explanation prompts recommended [+W] — a recommendation for future research, not a test. - Subjective ratings diverged by persona in an adaptive system [~M] — ratings of a simulated conversational partner, no learning outcome.

Feedback, hints and help-seeking - Self-questioning improves comprehension [+S] — about teaching learners to generate their own questions about text, not a tutor's move; a tutor could prompt it, untested here. - Staged disclosure and diagnostic error patterns [~W] — illustrative cases of how a tutor's hidden diagnosis diverged from its feedback; no prevalence estimate.

Pedagogical agents - Minimal social cues trigger social reactions [~M] — explains why agents draw social responses; no learning outcome. - Avatars said to increase tutor engagement (vendor white paper) [+W] — vendor assertion; the agent meta-analyses carry this question.

Collaboration and humans in the loop - Personalized learning as a human–AI collaboration [+M] — a review's conclusion that agrees with the human-handoff step; no new test. - Practitioners reject AI replacing instructor expertise [+M] — practitioner beliefs (survey of 72). - Students prefer hybrid AI and instructor feedback [+M] — preference of 30 students for essay feedback; a learner-valued input, not an outcome. - AI-driven PBL enhances educator-student collaboration [+W] — a review theme, no effect. - Over-reliance on AI may reduce human interaction [-M] — a policy argument; agrees with the human-handoff step.

Role-play and simulated people (not about this page: these concern AI role-play agents and simulated patients, covered by practice with simulated people) - RL training reduced missed red flags in a clinical agent [+M] (not about this page) - Role-play agents degrade consistently under attack across models [-M] (not about this page) - Authority challenge and emotional manipulation are the most effective attacks [-M] (not about this page) - Automated judging aligns with experts with a conservative bias [~M] (not about this page) - Persona leakage strongest for communication style and personality [-M] (not about this page) - The review reports that highly competent PALs increased learners' application of acquired knowledge while low-competent PALs increased self-efficacy [+W] — attached 2026-10-17 from Yanghee Kim et al. (2006); restates the Kim, Baylor et al. (2006) experiment whose claims the agents bullet already uses.

Linked after the first re-read; read 2026-10-19 (second pass)

Learning outcomes and reliance - When AI assistance is withdrawn, performance tends to decline [~W] — read 2026-10-19: a position paper citing one study, no effect size; agrees with the horizon rule the trials above already carry. - AI-assisted performance does not transfer to unassisted performance [~W] — read 2026-10-19: a theoretical article's report of a cited experiment, no effect size; same point as the unaided-assessment row. - AI assistance can improve short-term performance while reducing independent capability [-W] — read 2026-10-19: a conceptual synthesis, no new data. - AI support can improve performance while weakening metacognitive accuracy [~W] — read 2026-10-19: a workshop paper's report of two cited studies; the trace and confidence results in the over-reliance bullet measure this directly. - Prior genAI use shows some evidence of poorer unaided reasoning [~W] — read 2026-10-19: the preregistered tests were null and the one significant estimate is exploratory; does not change the default. - GenAI access reduces perceived difficulty of the assisted task [~W] — read 2026-10-19: perceived difficulty only, same lab study. - Early reliance relates to both benefit and negative impact [-W] — read 2026-10-19: self-reported benefit and harm in a 118-student survey; no learning measure. - Early reliance and evaluation literacy interact [~W] — read 2026-10-19: same survey, self-report outcomes. - Exploratory profile interaction with AI-assisted gains [~W] — read 2026-10-19: interactions with prior knowledge and university ranking were not significant; nothing to act on. - Boundary conditions on verification and reliance [~W] — read 2026-10-19: a review's list of expected moderators, which it calls analytic expectations rather than established effects. - Longitudinal trust studies omit skill outcomes [~W] — read 2026-10-19: an argument that agrees with the gap stated under Elements and limits. - Prompt–response interfaces limit learning value without guidance [+W] — read 2026-10-19: a perspective citing earlier work; the structured-prompting trial in the model carries the point. - Standalone LLM tutors lack control over the learning trajectory [-W] — read 2026-10-19: a review chapter's argument; agrees with the course-length row's call for a planned curriculum. - LLM use for extra explanations did not significantly affect grades [~W] — read 2026-10-19: a non-significant correlation in one programming course; not evidence of no effect. - Essays under a standard chatbot sat closest to AI-only essays [~W] — read 2026-10-19: exploratory embedding analysis from the Engage-to-Unlock study, whose text-coverage result is in the ill-structured row. - Withholding answers improves output-bound accuracy, help-seeking does not [+M] — read 2026-10-19: about letting quiz-takers answer "don't know", not about a tutor's moves; no bearing on the configuration.

Tutoring systems and deployments - Mathspace summer users showed significant gains in one district [~W] — read 2026-10-19: a pre-post brief of 27 students with no comparison group and low engagement. - ChatGPT class mean 81.81 against 70.45 [+W] — read 2026-10-19: descriptive means from two classes of 31, low rigour; does not change the expectation. - Final-test score distributions differed by class [+W] — read 2026-10-19: frequency tables from the same study. - Teaching a ChatGPT agent: fewer code submissions [+W] — read 2026-10-19: descriptive log counts from the teachable-agent study already cited; no test. - Betty's Brain strategies can be detected from activity sequences [~W] — read 2026-10-19: a method result reported in a commentary; supports reading the log (step 4) but measures no outcome. - Proactive prompts and information items explain little of help-tool use [~W] — read 2026-10-19: a regression inside the distance-education experiment whose main results are in the help-seeking row. - Math Agent concision raised behavioural engagement and cut latency [~W] — read 2026-10-19: engagement and speed, not learning. - An adaptive router reduced turns in long conversations [~W] — read 2026-10-19: efficiency in a live A/B test; no learning outcome. - Scaffolding level is associated with the student's next move [~W] — read 2026-10-19: a turn-level association that agrees with the leakage results already used. - Use of worked examples was discouraged in the Cognitive Tutor Algebra course [~W] — read 2026-10-19: a historical note about instructors' beliefs, not evidence; the fading result in step 3 tests worked steps directly. - RESPE pilot: privacy concerns and repetitive scaffolds [~W] — read 2026-10-19: ten students' comments on a regulation-scaffolding prototype; a reminder to tell learners who sees their data. - Teachers see AI tutors as help at a scale no teacher can give [~W] — read 2026-10-19: teachers' perceptions in interviews; a stakeholder input, not an outcome. - Scenario-based tutor-training lessons gave a pooled 7.4% gain [+W] — read 2026-10-19: about training human tutors with AI-scored lessons; closer to teacher education than to AI tutoring of learners. - Interpretable AI mechanisms give clearer learning benefits [+W] — read 2026-10-19: a review of AI game-based learning; consistent with preferring a configuration you can state, no direct test here. - Counterfactual probabilities quantify the effect of hints and skills [~W] — read 2026-10-19: a modelling method for assessment with support; no tutoring outcome.

Building and evaluating LLM tutors (no learners) - Deep-thinking ratio does not predict tutoring quality [~W] — read 2026-10-19: a model-side metric; benchmark only. - Emotion-aware prompting improves pedagogical quality ratings [+W] — read 2026-10-19: LLM-judged quality, no learners. - Responsiveness and helpfulness trade off in a long-horizon benchmark [~W] — read 2026-10-19: simulated learners; agrees with the withholding default. - Parameter training beat prompt engineering on Socratic scaffolding [+W] — read 2026-10-19: a 30-question benchmark; no learners. - Education-oriented tuning raised use of longitudinal student evidence [+W] — read 2026-10-19: benchmark scores of a model family. - Teaching dialogues conclude within 30 turns [~W] — read 2026-10-19: a benchmark design choice between simulated agents. - Learner context changes an AI tutor's explanations [+W] — read 2026-10-19: a two-persona demonstration; personalisation from the learner's work stays the default. - AutoTutor's best semantic matcher combines LSA, word overlap and regular expressions [~W] — read 2026-10-19: engineering of an older system. - Policy in reviewed systems is expressed as assistance boundaries [+W] — read 2026-10-19: a scoping review's description of 90 programming-support systems; useful vocabulary for recording the configuration, no outcome.

Feedback, preferences and perceptions - Feedback source (human tutor or GPT-4) had no significant effect on post-test writing [~W] — read 2026-10-19: the between-groups test from the same study as the second-language writing row; adds nothing beyond it. - Blind ratings favoured GenAI feedback on genuineness [~M] — read 2026-10-19: perceptions of 41 students; no learning outcome. - AI-assisted instructor feedback kept its rated usefulness [~W] — read 2026-10-19: AI helping instructors write feedback, a human-in-the-loop use outside the tutor's moves; usefulness ratings only. - AI feedback reduces the perceived risk of exposing weaknesses [+W] — read 2026-10-19: a review's report of learners' experience; useful when eliciting why a learner prefers the tutor, not an outcome. - Learners selectively filtered AI feedback [~W] — read 2026-10-19: thematic coding of reflections; no outcome. - Trust in customised agents followed accurate, contextual feedback [~W] — read 2026-10-19: perceptions; consistent with the grounding row. - Structured prompting raised perceived clarity and lowered novelty [~W] — read 2026-10-19: user-experience ratings from the structured-prompting study already used. - Students found ChatGPT useful but lacking depth against TA answers [~W] — read 2026-10-19: one student's comment. - Students see LLM chatbots as useful but requiring critical use [~W] — read 2026-10-19: self-report. - AI chatbots give feedback, step-by-step solutions and self-quizzing [+W] — read 2026-10-19: a review's description of uses, no effects. - ChatGPT-generated synthesis lists education benefits [+W] — read 2026-10-19: text generated by ChatGPT and printed in a review; not evidence. - Personalised guidance for all students at once is a flipped-learning gap [+W] — read 2026-10-19: an article's motivating statement, no data. - Explanations widen the human–AI fairness gap [~W] (not about this page) — read 2026-10-19: acceptance of AI grading decisions; AI grading is outside this page's scope. - In Study 1, university students using ChatTutor scored significantly higher than teaching-as-usual students on a delayed conceptual knowledge test four weeks after the lesson, but did not differ significantly from ChatGPT students [+M] — attached 2026-10-20 from Guido Makransky et al. (2025), which proposed "Use theoretically informed GenAI chatbots to scale personalized scaffolding of generative learning activities where instructor formative assessment is not feasible"; tests this page's relationship.

Illustrative design instance and observation record

Illustration only, invented, not a tested course. An online, self-paced introductory statistics module for working adults, with a mentor who answers within one business day. The designer's objective is that each learner can choose and run the right test on a new data set and explain the result; one learner says she wants to pass the module quickly while working full time. Agree that the module assessment is unaided and proctored, and that a delayed check a month later matters to the programme. The first activity asks each learner to choose a test for a short scenario without help. The tutor is configured to work step by step, never to give the final test choice or computed value, to give problem-specific hints written by the course team, and to ask the learner to check its claims against the course notes. Sessions one and two: about 30 minutes of tutored practice; sessions three and four: hints only after an attempt; session five: an unaided practice set with the tutor off. The log flags a learner who asks for the answer three times on one step, or accepts a tutor statement later marked wrong, and sends a message to the mentor.

Record: unaided first attempt → tutor configuration and content actually used → per-step requests (answer, hint, explanation, check), hints given, answer released or not → tutor errors and whether the learner accepted them → handoffs to a human and their response time → unaided score on a changed task at the end → delayed unaided score. Preserve the learner's stated purpose and any change in it. An agent may summarize this record, but must not fabricate missing inputs or assign numerical mastery probabilities without a calibrated model.

Elements and limits

Hints, feedback, immediate feedback, scaffolding, fading scaffolding, Socratic questioning, self-explanation, worked examples, pedagogical agent signaling, peer teaching, collaborative problem solving and metacognitive reflection.

This pattern is scoped to the tutor's moves within tasks and the handoff to people. It does not cover AI grading, AI-generated course content, proctoring or AI role-play as a practice partner (see practice with simulated people). The wiki records no randomized comparison of an LLM tutor across a whole course with a delayed or transfer outcome (the longest is a five-week field experiment with 46 analysed and an end-of-period post-test), no test of a handoff trigger, and no test of fading a conversational tutor's help across sessions (the one fading result is for worked steps in a step-based tutor); no trial of grounding a tutor in course materials against an ungrounded one with a learning outcome; and no test of withholding learner identity labels from a tutor's prompt beyond the feedback-bias audit it rests on. Much of the newer LLM-tutor evidence here is benchmark or simulated-student evaluation of the tutor's outputs, which says what a tutor does, not what learners learn. The diagnostic branches above remain to be tested with actual learners.

Claims

Examples

Key Sources

  • 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. doi:10.1037/a0037123
  • Kulik, J. A., & Fletcher, J. D. (2016). Effectiveness of intelligent tutoring systems: A meta-analytic review. Review of Educational Research, 86(1), 42–78. doi:10.3102/0034654315581420
  • VanLehn, K. (2011). The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems. Educational Psychologist, 46(4), 197–221. doi:10.1080/00461520.2011.611369
  • 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. doi:10.1207/s15516709cog2504_1
  • Graesser, A. C. (2016). Conversations with AutoTutor Help Students Learn. International Journal of Artificial Intelligence in Education. https://doi.org/10.1007/s40593-015-0086-4
  • Bastani Hamsa, Bastani Osbert, Sungu Alp, Ge Haosen, Kabakcı Özge, Mariman Rei. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences of the United States of America. https://doi.org/10.1073/pnas.2422633122
  • 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 15:17458. https://doi.org/10.1038/s41598-025-97652-6
  • Zara Contractor and Germán Reyes. (2026). Experimental Evidence on the Learning Impact of Generative AI. https://arxiv.org/abs/2607.08849 doi:10.48550/arxiv.2607.08849
  • Grace Liu, Brian Christian, Tsvetomira Dumbalska, Michiel A. Bakker, Rachit Dubey. (2026). AI Assistance Reduces Persistence and Hurts Independent Performance. Published as a conference paper at COLM 2026. https://arxiv.org/abs/2604.04721
  • Pardos ZA, Bhandari S (2024) ChatGPT-generated help produces learning gains equivalent to human tutor-authored help on mathematics skills. PLoS ONE 19(5): e0304013. https://doi.org/10.1371/journal.pone.0304013
  • Reddig, J. M., Arora, A., & MacLellan, C. J. (2025). Generating In-Context, Personalized Feedback for Intelligent Tutors with Large Language Models. Int J Artif Intell Educ. https://doi.org/10.1007/s40593-025-00505-6
  • Hou, X., Weng, Y., Yeh, C. H., Lee, D. L., Zheng, L., Li, F., Wang, W., & Liu, Y. (2026). Adaptive Scaffolding Needs Contingency: An AI Tutor That Escalates and Fades on What the Learner Does. arXiv preprint. https://arxiv.org/abs/2609.22993 doi:10.48550/arxiv.2609.22993
  • Shuyi Fan, Boyuan Deng, Mengyu Xu, Jiale Liu, Hongyang Zhang, Qiaoxin Yang, Chongyang Gao. (2026). Rethinking LLM-Judged Helpfulness as a Pedagogy Signal: A Pre-Registered Audit Across Tutor Models. arXiv. https://arxiv.org/abs/2607.28128 doi:10.48550/arxiv.2607.28128
  • Belland, B. R., Walker, A. E., Kim, N. J., & Lefler, M. (2017). Synthesizing Results From Empirical Research on Computer-Based Scaffolding in STEM Education: A Meta-Analysis. Review of Educational Research, 87(2), 309–344. doi:10.3102/0034654316670999
  • Wisniewski, B., Zierer, K., & Hattie, J. (2020). The power of feedback revisited: A meta-analysis of educational feedback research. Frontiers in Psychology, 10, 3087. doi:10.3389/fpsyg.2019.03087
  • Kluger, A. N., & DeNisi, A. (1996). The effects of feedback interventions on performance: A historical review, a meta-analysis, and a preliminary feedback intervention theory. Psychological Bulletin, 119(2), 254–284. doi:10.1037/0033-2909.119.2.254
  • Ren S (2026) College students’ metacognitive awareness of generative-AI reliance: an experimental study of decision confidence and attribution bias. Front. Psychol. 17:1926110. https://doi.org/10.3389/fpsyg.2026.1926110
  • Xiaoyu Hou, Bo Xiao, Hexu Liu and Shane Mueller. (2026). The Role of Instructional Guidance in Generative AI-Assisted Learning: Empirical Evidence from Construction Engineering Education. https://arxiv.org/abs/2606.05509 doi:10.48550/arxiv.2606.05509
  • Tomohiro Nagashima, Mirella Hladký, and Vera Rief. (2026). Warning About AI Fallibility Increases Help-Seeking in an Intelligent Tutoring System. https://arxiv.org/abs/2606.03822 doi:10.48550/arxiv.2606.03822
  • Wong, S. S. H., & Qiu, S. X. (2026). Think First, ChatGPT Later: Guiding Human–AI Collaboration for Learning Gains in Independent Human Creativity. Educational Psychology Review. https://doi.org/10.1007/s10648-026-10118-7
  • Chen X (2026) Layer-sensitive cognitive offloading in generative AI-assisted writing: supported performance and independent no-AI outcomes. Front. Psychol. 17:1906199. https://doi.org/10.3389/fpsyg.2026.1906199
  • Zhao Yan, Yue Yuhe, Sun Zhonghua, Jiang Qiang, Li Gangsheng. (2025). Does Generative Artificial Intelligence Improve Students’ Higher-Order Thinking? A Meta-Analysis Based on 29 Experiments and Quasi-Experiments. Journal of Intelligence. https://doi.org/10.3390/jintelligence13120160
  • Rose E. Wang, Ana T. Ribeiro, Carly D. Robinson, Susanna Loeb, and Dora Demszky. (2025). Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise. arXiv:2410.03017v2. https://arxiv.org/abs/2410.03017 doi:10.48550/arxiv.2410.03017
  • Ashish Gurung, Ge Gao, Jordan Gutterman, Danielle R. Thomas, Shivang Gupta, Lee Branstetter, Emma Brunskill, Vincent Aleven, Kenneth R. Koedinger. (2026). Improving Hybrid Human-AI Tutoring by Differentiating Human Tutor Roles Based on Student Needs. arXiv preprint. https://arxiv.org/abs/2605.11155 doi:10.48550/arxiv.2605.11155
  • Koedinger, K., Aleven, V., Roll, I., & Baker, R. (2008). In vivo experiments on whether supporting metacognition in intelligent tutoring systems yields robust learning. https://learnlab.org/
  • W. Lewis Johnson & James C. Lester. (2015). Face-to-Face Interaction with Pedagogical Agents, Twenty Years Later. Int J Artif Intell Educ. https://doi.org/10.1007/s40593-015-0065-9
  • Nicole C. Krämer & Gary Bente. (2010). Personalizing e-Learning. The Social Effects of Pedagogical Agents. Educ Psychol Rev. https://doi.org/10.1007/s10648-010-9123-x
  • Woo, H.L. (2009). Designing multimedia learning environments using animated pedagogical agents: factors and issues. Journal of Computer Assisted Learning, 25, 203-218. https://doi.org/10.1111/j.1365-2729.2008.00299.x
  • Kim, Y., Baylor, A. L., & PALS Group. (2006). Pedagogical Agents as Learning Companions: The Role of Agent Competency and Type of Interaction. Educational Technology Research and Development, 54(3), 223-243. https://doi.org/10.1007/s11423-006-8805-z
  • Kobayashi, K. (2019). Learning by preparing-to-teach and teaching: A meta-analysis. Japanese Psychological Research, 61(3), 192–203. doi:10.1111/jpr.12221
  • Doris B. Chin, Ilsa M. Dohmen, and Daniel L. Schwartz. (2013). Young Children Can Learn Scientific Reasoning with Teachable Agents. IEEE Transactions on Learning Technologies. https://doi.org/10.1109/TLT.2013.24
  • Angxuan Chen, Yuang Wei, Huixiao Le, Yan Zhang. (2024). Learning-by-Teaching with ChatGPT: The Effect of Teachable ChatGPT Agent on Programming Education. arXiv. https://arxiv.org/abs/2412.15226 doi:10.48550/arxiv.2412.15226
  • Castro-Alonso, J. C., Wong, R. M., Adesope, O. O., & Paas, F. (2021). Effectiveness of Multimedia Pedagogical Agents Predicted by Diverse Theories: a Meta-Analysis. Educational Psychology Review. https://doi.org/10.1007/s10648-020-09587-1
  • Biswas, G., Segedy, J. R., & Bunchongchit, K. (2016). From Design to Implementation to Practice a Learning by Teaching System: Betty's Brain. International Journal of Artificial Intelligence in Education. https://doi.org/10.1007/s40593-015-0057-9
  • Radkowitsch, A., Vogel, F., & Fischer, F. (2020). Good for learning, bad for motivation? A meta-analysis on the effects of computer-supported collaboration scripts. International Journal of Computer-Supported Collaborative Learning, 15, 5–47. https://doi.org/10.1007/s11412-020-09316-4
  • Freydis Vogel & Christof Wecker & Ingo Kollar & Frank Fischer. (2016). Socio-Cognitive Scaffolding with Computer-Supported Collaboration Scripts: a Meta-Analysis. Educ Psychol Rev (2017) 29:477–511. https://doi.org/10.1007/s10648-016-9361-7
  • Ritter, S., Anderson, J. R., Koedinger, K. R., & Corbett, A. (2007). Cognitive Tutor: Applied research in mathematics education. Psychonomic Bulletin & Review, 14(2), 249-255. https://doi.org/10.3758/bf03194060
  • Li, S., Zeng, C., Liu, H., Jia, J., Liang, M., Cha, Y., Lim, C. P., & Wu, X. (2025). A meta-analysis of AI-enabled personalized STEM education in schools. International Journal of STEM Education. https://doi.org/10.1186/s40594-025-00566-y
  • Walkington, C. A. (2013). Using adaptive learning technologies to personalize instruction to student interests: The impact of relevant contexts on performance and learning outcomes. Journal of Educational Psychology, 105(4), 932–945. doi:10.1037/a0031882
  • Kalyuga, S., & Renkl, A. (2010). Expertise reversal effect and its instructional implications: introduction to the special issue. Instructional Science. https://doi.org/10.1007/s11251-009-9102-0
  • Yan, L., Greiff, S., Lodge, J. M., & Gašević, D. (2025). Distinguishing performance gains from learning when using generative AI. Nature Reviews Psychology, 4, 435–436. https://doi.org/10.1038/s44159-025-00467-5
  • Urban, M., Děchtěrenko, F., Lukavský, J., Hrabalová, V., Svacha, F., Brom, C., & Urban, K. (2024). ChatGPT Improves Creative Problem-Solving Performance in University Students: An Experimental Study. Computers & Education. https://doi.org/10.1016/j.compedu.2024.105031
  • Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. Preprint. https://arxiv.org/abs/2506.08872 doi:10.48550/arxiv.2506.08872
  • Ahtisham, B., Vanacore, K., Napoli, A., Arens, J., Ionova, K., Cohn, C., Salehi, S., & Kizilcec, R. (2026). Examining Variation in How Guided AI Tutors Resolve Student Impasses. arXiv. https://arxiv.org/abs/2609.38346 doi:10.48550/arxiv.2609.38346
  • Abdelghani, R., Kaiser, P., & Murayama, K. (2026). Regulating the AI Tutor: Intentions, Help-Seeking, and Self-Regulated Learning in Adolescent GenAI Use. NextGen Learning Interfaces Workshop, AIED 2026. https://arxiv.org/abs/2606.08568 doi:10.48550/arxiv.2606.08568
  • Tan, M., Phalen, L., & Demszky, D. (2026). Marked Pedagogies: Examining Linguistic Biases in Personalized Automated Writing Feedback. LAK 2026. ACM. doi:10.1145/3785022.3785113
  • Vonschallen, S., Kaufmann, K., Oberle, D., Eyssel, F., & Schmiedel, T. (2026). Teachy Mini: Development and Preliminary Evaluation of a Knowledge-Based Generative Social Robot for Higher Education. arXiv. https://arxiv.org/abs/2607.22345 doi:10.48550/arxiv.2607.22345
  • Zhao, J., Yang, K., Gao, H., & Jiang, G. (2026). Generative artificial intelligence and the transformation of medical education: a scoping review. Front. Educ. 11:1907589. https://doi.org/10.3389/feduc.2026.1907589
  • Yang, L., Liu, H., Li, S., Jia, R., Xiao, Y., Chen, G., & Lu, L. (2026). Evaluating Scaffolding-Oriented Multi-Agent Large Language Model System for Clinical Interview Training. https://arxiv.org/abs/2609.10939 doi:10.48550/arxiv.2609.10939
  • Shihao Yang, Marshall Van Alstyne, and Chrysanthos Dellarocas. (2026). When AI Tutors Speak: Evidence from a Randomized Field Experiment. https://arxiv.org/abs/2609.23958 doi:10.48550/arxiv.2609.23958
  • Curtis Northcutt, Inaara Hasmani, Kevin Feng, Trevor Khangi, Andreas Plesner, Jonas Mueller. (2026). StudentBench: AI and human tutoring yield equivalent GRE learning gains. https://arxiv.org/abs/2609.28470 doi:10.48550/arxiv.2609.28470
  • Benjamin Lira, Todd Rogers, Daniel G. Goldstein, Lyle Ungar, and Angela L. Duckworth. (2025). Coach not crutch: Evidence that AI can improve writing skill despite reducing effort. https://arxiv.org/abs/2502.02880 doi:10.48550/arxiv.2502.02880
  • Maier, S., Gunzenhäuser, M., Schweisthal, J., Schneider, M., & Feuerriegel, S. (2026). A meta-analysis of the effect of generative AI on productivity and learning in programming. arXiv preprint. https://arxiv.org/abs/2605.04779 doi:10.48550/arxiv.2605.04779
  • Fernandes, D., Buschek, D., Tankelevitch, L., Kosch, T., & Welsch, R. (2026). Explaining Too Much? Understanding How Large Language Model Reasoning Traces Influence Performance and Metacognition. https://arxiv.org/abs/2605.25856 doi:10.48550/arxiv.2605.25856
  • Bruno De Lièvre, Christian Depover & Pierre Dillenbourg. (2006). The relationship between tutoring mode and learners' use of help tools in distance education. Instructional Science. https://doi.org/10.1007/s11251-005-6076-4
  • Escalante, J., Pack, A., & Barrett, A. (2023). AI-generated feedback on writing: insights into efficacy and ENL student preference. International Journal of Educational Technology in Higher Education, 20:57. https://doi.org/10.1186/s41239-023-00425-2
Page metadata
Field Value
type pattern
id ai-tutoring
title AI Tutoring
description A reusable policy for tutoring by software, from step-based intelligent tutoring systems to conversational AI tutors and pedagogical or teachable agents: decide how much the tutor gives away, keep the learner doing the reasoning, judge it on unaided performance rather than assisted practice, and decide when a human tutor or mentor takes over.
status review
grain_size course
canonical True
generated by claude/unspecified on 2026-10-19
trust unverified — no one has confirmed this page's content

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