AI-Supported Tutoring and Feedback: The Evidence in 2026

AI-supported tutoring has moved from an area where practice ran far ahead of evidence to an area where three major 2025 syntheses now give teachers something real to work with. Huang and colleagues meta-analysed intelligent tutoring systems; Létourneau and colleagues reviewed AI-driven tutoring in K–12; Achuthan and colleagues examined AI's relationship with learner autonomy. The picture that emerges is clearer than it was two years ago, and more specific than the headlines suggest. This evidence review sets out what AI-supported tutoring and feedback actually are, what the 2025 evidence shows, and the implementation choices that separate genuine learning gain from expensive distraction.

What AI-supported tutoring and feedback are

Two broad classes of tool belong to this category.

Intelligent tutoring systems (ITS) are structured software environments that diagnose a student's current knowledge, select appropriate next items, provide step-by-step hints, and adapt difficulty based on performance. They are typically bounded to a specific domain (algebra, reading comprehension, anatomy) with pre-modelled content and predictable behaviour. Examples include Carnegie Learning's Cognitive Tutor for mathematics and various reading ITS programmes.

AI-supported feedback tools — usually built on large language models — provide open-ended feedback on student work: writing, problem-solving, code, speech. They are less bounded than ITS and more conversational. Behaviour is probabilistic rather than deterministic; teacher oversight becomes much more important.

The category also includes hybrid systems where an ITS-style structure is augmented with conversational LLM-based feedback, and where human teacher oversight is built into the platform.

What the research actually shows

The 2025 syntheses give the most recent and structured picture.

Huang and colleagues (2025) examined the effects of intelligent tutoring systems on educational outcomes. Their finding: moderate positive effects on academic achievement across subjects, with stronger effects in mathematics, science, and skill-heavy domains where ITS content is well-modelled. Effects depend on how deeply integrated the ITS is with classroom teaching, and on teacher oversight of the data the system produces.

Létourneau and colleagues (2025) in npj Science of Learning published a systematic review of AI-driven intelligent tutoring systems in K–12 education. Their conclusion: the evidence base is promising but uneven, with strongest findings in well-bounded content domains and particular concern about implementation fidelity, equity of access, and teacher oversight. Much of the positive effect depends on the ITS being a supplement to strong classroom teaching rather than a replacement for it.

Achuthan and colleagues (2025) in Frontiers in Education meta-analysed AI and learner autonomy across self-regulated and self-directed learning outcomes. Their finding: positive effects on autonomy measures, with caveats about assessment designs that prevent students from outsourcing thinking rather than doing it.

The defensible synthesis: AI-supported tutoring has real evidence for learning gain when the tool is well-bounded, teacher oversight is present, and assessment is designed to measure what the student can do — not what the student can get the AI to produce for them. The evidence for less-bounded LLM-based feedback is newer and more heterogeneous; teacher oversight and well-designed assessment are even more important there.

Core design principles

The recent syntheses converge on six principles for productive implementation.

• Use bounded tools for structured practice. ITS for algebra, reading fluency, grammar drills, anatomy labelling, clinical reasoning cases. These are the domains with the most consistent evidence.
• Maintain teacher oversight of the data. ITS platforms produce rich data on where students struggle. Teachers who review and act on that data multiply the effect; teachers who ignore it leave most of the gain on the table.
• Design assessments that require in-person performance. If your summative assessment can be outsourced to AI, the formative AI practice won't transfer. Assessment design is now a pedagogical question.
• Treat AI feedback as supplement, not replacement. AI feedback on a writing draft works well before a teacher review, not instead of it. Students learn when they integrate teacher and AI feedback, not when AI feedback stands alone.
• Preserve thinking load. The learning is in the thinking. If the AI does the thinking and the student reviews the output, transfer evaporates. Tool design and task design must keep the student doing the cognitive work.
• Teach AI literacy. Students need to evaluate AI output, detect hallucinations, notice plausible-but-wrong confidence, and integrate AI use into disciplined learning. This is a new layer of study skills, not automatic.

The classroom routines that carry the evidence

Five routines are visible in strong implementations.

• ITS deployment in structured practice windows. Twenty to forty minutes of ITS practice in class, with the teacher circulating and intervening on the students the system flags as struggling.
• Pre-submit AI feedback on drafts. Students get AI feedback on a first draft; they revise; they submit; the teacher gives the final feedback on the revision. The AI compresses the time to first feedback without replacing the teacher's diagnostic judgement.
• AI-generated retrieval practice. Teachers use AI to generate varied retrieval questions for the content; students practise and the teacher curates which questions are kept.
• Clinical-reasoning or case-based AI simulations. In medical, nursing, legal, and business education, AI-driven simulations produce structured practice in professional reasoning with feedback.
• Teacher-reviewed AI tutoring transcripts. For students using an AI tutor outside class, teachers sample or review transcripts to spot misconceptions the AI missed, corrected wrongly, or failed to challenge.

Classroom examples across phases

Primary. Year 4 mathematics. Students work with an ITS for twenty minutes twice a week on arithmetic facts. The teacher reviews the weekly data and reteaches the small group of students whose error patterns the system has flagged.

Secondary. Year 11 English. Students draft an essay, run it through a teacher-approved AI feedback tool with a structured rubric prompt, revise on the basis of the feedback, and submit. The teacher marks the revision and notes in class the patterns of common missed improvements — the part the AI did not catch.

Tertiary. First-year medical school. A clinical-reasoning AI simulation presents cases with branching outcomes. Students work in pairs through two cases per week. Faculty sample and debrief the transcripts during small-group teaching, focusing on reasoning moves the AI did not score.

Where AI-supported tutoring fails

The failure modes are consistent and — unusually for educational technology — mostly predictable.

• Outsourcing rather than thinking. Students paste the prompt, copy the answer, submit. The learning is zero. Assessment design is the single biggest control.
• Unbounded use with novices. A student with no baseline knowledge using an unbounded LLM often produces confident-looking output they cannot evaluate. The AI's confident wrong answers compound rather than correct the novice's misconceptions.
• No teacher oversight. ITS platforms that no one reviews become expensive worksheets. The data is the point.
• Equity of access. Unequal device, bandwidth, and home-AI-access patterns mean AI tools can widen rather than narrow existing gaps. The Létourneau and colleagues (2025) review flags this as a primary concern in K–12.
• Tools marketed on engagement rather than evidence. Platforms that produce engagement data but have not been tested against learning outcomes. Ask what the tool was evaluated against.

Best fit and poor fit

Best fit: structured, bounded domains (algebra, arithmetic, reading fluency, grammar, vocabulary, anatomy, clinical reasoning, coding practice); pre-submit feedback on drafts; AI-generated retrieval question banks curated by the teacher; case-based simulations in professional education.

Poor fit: first-time teaching of wholly novel complex content; classrooms where assessment cannot be designed to require in-person thinking; contexts where device or bandwidth access is unreliable or unequal; unsupervised use by students without AI literacy training.

Teacher requirements, assessment, and resources

AI-supported tutoring is professional-development-heavy and design-heavy. Teachers need to understand what the tools produce, how to review the data, and how to redesign assessment so that AI supplement becomes AI partnership rather than AI substitution.

Assess with in-person performance tasks, oral defences, unpaced written work produced in class, and practical demonstrations. If an assessment can be passed by pasting a prompt, the assessment itself is now the problem, not the tool.

How TAyumira supports AI-supported tutoring and feedback

TAyumira is itself an AI-native tool for teachers and integrates these evidence-based principles. When you generate a lesson, TAyumira produces:

• An in-class activity whose assessment design is AI-resistant (requires in-person thinking and performance)
• AI-generated retrieval sets and practice items for the content, with teacher-review prompts
• A structured pre-submit feedback prompt for student drafts with an AI literacy framing
• Teacher data-review prompts tied to common misconception patterns for the content
• A balance between AI-supported practice (outside class) and teacher-led formative assessment (in class)

Start for free — the Free tier covers the full workflow.

FAQ

What is the effect size of AI-supported tutoring?

Huang and colleagues (2025) reported moderate positive effects from intelligent tutoring systems on academic achievement, stronger in mathematics, science, and skill-heavy domains. Létourneau and colleagues (2025) in npj Science of Learning reviewed K–12 AI tutoring and described the evidence as promising but uneven. Achuthan and colleagues (2025) reported positive effects of AI on learner-autonomy measures with assessment-design caveats.

Is AI tutoring a replacement for teachers?

No, and the evidence does not support that framing. The strongest effects in the 2025 syntheses come from implementations where AI tools supplement strong classroom teaching, not replace it. The teacher's role shifts toward data review, design of AI-resistant assessment, and targeted intervention — but does not reduce.

What is the difference between intelligent tutoring systems and AI feedback tools?

Intelligent tutoring systems are bounded software environments with pre-modelled content and predictable behaviour; they produce structured practice in specific domains. AI feedback tools (typically LLM-based) give more open-ended feedback on student work; their behaviour is probabilistic and requires greater teacher oversight. Both have evidence supporting well-designed use, but the design principles differ.

How do I assess learning if students can use AI?

By redesigning assessment. In-person performance tasks, oral defences, unpaced writing in class, practical demonstrations, and well-designed formative assessment that requires process, not only product. If the final assessment can be generated by an AI from a prompt, the assessment itself — not the tool — is now the problem.

Is AI tutoring appropriate for primary classrooms?

With caution. Bounded ITS for structured practice (arithmetic fluency, phonics practice) has evidence in primary. Unbounded LLM use without heavy teacher oversight is less well-evidenced and raises concerns about equity of access and appropriate supervision. Létourneau and colleagues' 2025 K–12 review flags these as the primary concerns.

Related evidence reviews

• Adaptive Teaching and Inclusive Design Evidence
• Retrieval Practice and Spaced Practice Evidence
• Formative Assessment Evidence
• Interactive Video Learning Evidence

Sources

• Huang, X., et al. (2025). Effects of intelligent tutoring systems on educational outcomes.
• Létourneau, A., et al. (2025). A systematic review of AI-driven intelligent tutoring systems in K–12 education. npj Science of Learning.
• Achuthan, K., et al. (2025). Artificial intelligence and learner autonomy: A meta-analysis of self-regulated and self-directed learning. Frontiers in Education.
• VanLehn, K. The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems. (Foundational ITS reference.)

Try one AI-supplemented routine this week

Pick one writing or problem-solving task in your current unit. Design an assessment that requires in-person performance. Use an AI tool for pre-submit feedback in a structured way with students, teach them how to evaluate what the tool says, and mark the revised submission. If you want AI-resistant assessments and structured AI-feedback prompts generated for your topic, create a free TAyumira account.