AI Tutor vs Certified Teacher: What the Research Says for Parents

The studies, in one table

SD means standard deviation, the unit researchers use to say how far one group moved compared with another.

What is an AI tutor, and what is a certified teacher?

"AI tutor" covers more than one kind of product. The US Department of Education's 2023 report on AI split classroom technology into tools "specifically designed for educational use" and "general technologies that are widely used in educational settings". A general chatbot, such as ChatGPT or Claude, sits in the second group. Tools built for tutoring sit in the first: IES, a US federal research arm, calls them "Socratic-style AI chatbots" that "ask users probing questions rather than simply telling students the right answers." Khan Academy says Khanmigo works "without giving you direct answers". That is the vendor describing its own product, not evidence that it works. Older intelligent tutoring systems, in VanLehn's 2011 review, mostly give feedback at each step of a problem rather than checking only the final answer.

A certified teacher is a credential held by a person, and in Canada it comes from one province. In Ontario, the Ontario College of Teachers "licenses, governs and regulates" the profession, and every publicly funded school teacher must be an Ontario Certified Teacher (OCT). The licence follows an initial teacher education program, takes the form of a Certificate of Qualification and Registration, and stays in good standing through an annual fee. It also names the divisions a teacher is qualified for, from Primary (Grades K-3) to Senior (Grades 11-12).

So "Canadian certified" means certified by a particular province, and a careful answer names which one. The College says teachers certified in any Canadian province or territory can also get certified in Ontario, and since 1 January 2026 they can apply "As of Right" for deemed certification valid for up to six months. In the US, licences are issued state by state.

That credential is the one part of the comparison a parent can check on a public register, which is why it matters more here than the hourly price.

What did the studies actually find?

Search whether AI tutoring works and you will meet two sets of headlines that seem to contradict each other. Both are quoting real studies.

The large gains. At Harvard, Kestin and colleagues ran a randomized controlled trial in an undergraduate physics course; 194 students were eligible for the study. Students "learn significantly more in less time when using the AI tutor", with median learning gains over double those of the in-class group. The tutor was custom-built with "expert-crafted, question-specific prompts written by instructors experienced with the content." The authors add their own limit: they "do not presume that structured AI tutoring will always outperform in-class active learning in all contexts".

In Nigeria, a World Bank team gave senior secondary students Microsoft Copilot, powered by GPT-4, for English practice over six weeks. The abstract reports a gain of 0.23 standard deviations in English, the main outcome, and 0.31 on an assessment that also covered knowledge of AI and digital skills. The authors equate the gains to 1.5 to 2 years of business-as-usual schooling. Teachers were in the room, though, and the team's own blog says: "We cannot isolate the effect of the chatbot alone."

The small gains. The largest and longest school trial in the research we read ran for two years in 18 Tennessee middle schools. Students used Khan Academy with Khanmigo in daily remedial math sessions. Math achievement rose by "about 0.06 to 0.08 standard deviations over a school year", and the authors write that "These gains resemble those from Khan Academy practice without AI assistance." The median student messaged Khanmigo on only a third of the days they practised. The authors' conclusion: "The binding constraint appears to be engagement".

A paper published in 2026, on a trial in US Grades 1 to 5, found the same problem from the other side. Nearly half of the control group never used the AI literacy platform. Adding an in-person tutor raised engagement by 71-80%, yet "the intervention did not improve reading achievement".

Why they disagree. College physics students with an instructor-written tutor are not middle schoolers on a remedial program, and six weeks with a teacher present is not two years of daily use. The test matters too: Kulik and Fletcher found that results for intelligent tutoring systems "depended to a great extent on whether improvement was measured on locally developed or standardized tests". And scale matters. Kraft, Schueler and Falken's meta-analysis of 282 randomized trials found that studies matched to large-scale US programs aimed at standardized test performance had pooled effects "only a third to a half as large" as the full sample, driven by "stark declines in pooled effect sizes as tutoring program scale increases".

The human side has a longer record, and who teaches matters in it. Nickow, Oreopoulos and Quan found effects "stronger, on average, for teacher and paraprofessional tutoring programs than for nonprofessional and parent tutoring." The Education Endowment Foundation in the UK notes that programs using teaching assistants or volunteers "may be less effective than those using experienced and specifically trained teachers."

Strong causal evidence on AI in schools is still thin. The Stanford SCALE Initiative found "only 20 high-quality causal studies" among more than 800 papers, and as of late 2025 reported no high-quality causal study of student AI use in US K-12 classrooms. The Tennessee and Grades 1 to 5 trials came out after that cut. IES wrote in August 2026 that strong causal studies "are still extremely limited".

Where do AI tutors go wrong?

The most common failure is handing over the answer. Bastani and colleagues gave nearly a thousand high school math students in Turkey either a ChatGPT-style interface (GPT Base) or a version with learning safeguards (GPT Tutor). With access, grades rose 48% and 127% respectively. When access was removed, GPT Base students did worse than students who never had it: a "17% reduction in grades for GPT Base". Without guardrails, students "subsequently perform worse on their own". The safeguards in GPT Tutor largely mitigated that drop.

The same team tested the plain tool on 57 practice problems, ten times each. It "gives a correct answer only 51% of the time on average; it makes logical errors 42% of the time and arithmetic errors 8% of the time". The tool was GPT-4, and current models may do better. The safeguarded version used teacher-written solutions, so it could not give wrong feedback. The Harvard team called hallucination a "significant challenge" for education and did not rely on GPT-4 alone to write solutions. OpenAI's own terms say users "should not rely on Output from our Services as a sole source of truth or factual information".

IES adds that students who used general-purpose chatbots while learning "were more likely to have lower exam performance", even when they felt the tools helped. SCALE found tools "designed with pedagogical guardrails" showed "more promising outcomes than general-purpose chatbots that provide answers directly".

What the research says about AI with a person in the loop

The evidence here is limited. IES lists "Teacher-mediated and AI-augmented tutoring" as one of three general patterns, and says it "may promote student learning". SCALE's examples of guardrailed tools are tutoring systems that give hints or guide reasoning. And the largest gain in the studies above came from a stand-alone AI tutor that instructors built for the Harvard physics course.

Our own view is that AI is best used as a tool that supports a teacher's judgement. That is a view, and the studies below do not settle it.

Tutor CoPilot, from Stanford, was "the first randomized controlled trial of a Human-AI system in live tutoring", with 900 tutors and 1,800 K-12 students from under-served communities. The AI offered suggestions to tutors during live sessions, and each tutor chose what to use. Students whose tutors had it were 4 percentage points more likely to master topics, and students of lower-rated tutors gained 9 points. Tutors with the tool were "less likely to give away the answer to the student", and they flagged suggestions "that are not grade-level appropriate".

IES also describes human-supervised AI tutoring, in which students use an AI model fine-tuned for teaching while expert tutors supervise them directly. Those students "performed as well as or better than those working with human tutors alone". The US Department of Education's 2023 report went further: "the Department firmly rejects the idea that AI could replace teachers". That report is the Department's 2023 position, written under a previous administration.

The Grades 1 to 5 trial above is the counterweight: a person in the loop is no guarantee of a gain.

Age limits and privacy before your child signs up

Read the age rules before anyone signs up.

• ChatGPT. The OpenAI Terms of Use, effective 1 January 2026, say "You must be at least 13 years old or the minimum age required in your country to consent to use the Services", and anyone under 18 needs a parent or legal guardian's permission.
• US law. The COPPA rule defines a child as "an individual under the age of 13", and an operator must obtain verifiable parental consent before collecting a child's personal information.
• UNESCO. Its 2023 guidance on generative AI in education recommends an age limit of 13 for independent use of these tools.
• Khanmigo for families. The paying account holder must be 18 or older and live in the United States, with a US billing address. A parent can then turn on access for up to 10 children. A family without a US billing address cannot open that account; whether schools offer another route was not checked.
• Canada. The Children's Privacy Code from the Office of the Privacy Commissioner of Canada says organisations should "Make high privacy settings the norm", "Limit how much data they collect about you" and "Avoid making profiles about you". Treat it as the regulator's guidance.

In Ontario, the Ministry of Education's index of Policy/Program Memoranda, last updated 20 October 2025, contains none on artificial intelligence. Our guide to AI rules in Canadian schools covers what that means in a classroom.

How to check that a teacher is certified

In Ontario, search the teacher's name on the Ontario College of Teachers' public Find a Teacher register. The College's own pages describe what the certificate carries: it stays in good standing through an annual membership fee, and its Basic Qualifications name the divisions the teacher is qualified for. Check both, and check that the division fits your child, for example Junior (Grades 4-6) for a Grade 5 student.

Elsewhere in Canada, the licence comes from the regulator in the province that issued it, so ask the teacher which province that is and search there. Our walkthrough on how to check teacher certification in Canada has the steps, and what "provincially certified" means explains the wording.

In the US, each state runs its own lookup. New York's is the Certification Lookup on the State Education Department's site, with a separate route for out-of-state applicants. For the wider difference between the two roles, see certified teacher vs tutor.

Questions to ask, at the kitchen table

None of these needs anything more than the device your child already uses.

• Ask your child to show you a recent AI conversation. Did the tool give the answer, or ask a question back?
• Give the same kind of problem on paper the next day, with the app closed. That is the moment the Turkey trial measured, and where the unguarded tool fell short.
• Check one AI answer yourself against the textbook or the teacher's worked example. One wrong step tells you how much checking the tool needs.
• Read the age line in the terms with your child, before the account exists.
• Ask a teacher two questions. Which province or state issued your licence, and which grades does it cover? Then look it up.
• Ask how the teacher uses AI, if at all, and what they check before your child sees it.