Industries / Government, education and community / Education / Tutoring
One-to-one teaching
A tutor works with one student or a small group and decides what to teach next from what the student does in the moment.
6
Milestones achieved, from 1982 to 2024.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1982
Tutors learn too
Peter Cohen, James Kulik and Chen-Lin Kulik pooled 65 independent evaluations of school tutoring programs. Tutored students outperformed control students on examinations, and the children who served as tutors also came to understand the subject better. Neither group's self-esteem changed much.
Credit
Peter A. Cohen, James A. Kulik and Chen-Lin C. Kulik
2014
Tutors without teaching degrees
Minnesota Reading Corps places AmeriCorps members, who need no background in literacy or teaching, in schools to tutor struggling readers from kindergarten to third grade for 20 minutes a day, using set lessons and a one-minute reading check each week. In a randomized trial run from 2011 to 2014 by NORC at the University of Chicago for the federal Corporation for National and Community Service, tutored kindergarten, first and third graders scored significantly higher than untutored classmates after one semester. A 2017–18 replication with 622 students found effects of 0.85 and 0.81 standard deviations in kindergarten and first grade and 0.28 in second and third grade.
LessonEffects were large for the youngest children and small or not significant for older ones.
Credit
Minnesota Reading Corps and ServeMinnesota; NORC at the University of Chicago (evaluator)
2020
Ninety-six trials pooled
Andre Nickow, Philip Oreopoulos and Vincent Quan pooled 96 randomized trials of tutoring from preschool to grade 12 and found an average effect of 0.37 standard deviations on learning. Effects were larger when the tutors were teachers or paraprofessionals (trained tutors who are not certified teachers) than when they were volunteers or parents, larger in the early grades, and larger for tutoring during the school day than after school.
LessonAbout 80% of the trials took place during school hours, so most of the evidence is not about the after-school tutoring that families buy.
Credit
Andre Nickow (Northwestern University), Philip Oreopoulos (University of Toronto) and Vincent Quan (J-PAL North America)
2021
Daily tutoring for teenagers
Saga Education put tutors, mostly recent college graduates with strong math skills and no teacher training, into Chicago public high schools, each working with two students for one class period every day. Tutors trained for about 100 hours and were paid $16,000 plus benefits for the nine-month year. In two randomized trials with 2,633 and 2,710 ninth and tenth graders, math test scores rose 0.16 and 0.37 standard deviations, at $3,500 to $4,300 per student per year by the authors' count.
LessonTutoring raised teenagers' scores, not only young children's, when it was daily and built into the timetable.
Credit
Saga Education and Match Education; Jonathan Guryan, Jens Ludwig and colleagues (Northwestern University and the University of Chicago Crime Lab and Education Lab)
2024
Software takes alternate days
A randomized trial with about 4,000 students in two large school districts, in 2018–19 and 2019–20, tested groups of four in which two students worked with a tutor while the other two used math practice software, swapping every other day. Math scores rose 0.23 standard deviations, almost as much as Saga's two-to-one model, at about 30% lower cost per student. The University of Chicago Education Lab ran the trial with Chicago and New York City public schools and Saga Education, and says the model halves the number of tutors needed.
Credit
Monica Bhatt, Jonathan Guryan, Salman Khan, Michael LaForest-Tucker and Bhavya Mishra; Saga Education
2024
Early reading gains reverse
Reading Recovery, developed by the New Zealand literacy researcher Marie Clay in the 1970s, gives the weakest first-grade readers 12 to 20 weeks of daily 30-minute one-to-one lessons from specially trained teachers. A randomized trial under the US Investing in Innovation (i3) program found that after five months participants had made more than 130% of the average first grader's reading growth. A follow-up of 9,879 third graders and 6,345 fourth graders found that participants then scored 0.19 to 0.43 standard deviations below similar students who had just missed the cutoff for the program.
LessonEarly tutoring gains can fade or reverse, so a program needs to be followed for years; less than a quarter of the students could be traced to third grade, which weakens the follow-up.
Credit
Marie Clay; Henry May and colleagues, Center for Research in Education and Social Policy, University of Delaware
Watch for
AI hints for human tutors
Tutor CoPilot shows a tutor suggested replies, drawn from several teaching strategies, when the tutor asks for help during a chat-based math session. In a randomized trial in spring 2024 with FEV Tutor, an online tutoring company, and a school district in the US South, students whose tutors had the tool were 4 percentage points more likely to pass a short end-of-session quiz (66% against 62%), and 9 points more likely with lower-rated tutors, at about $20 per tutor per year. The study found no significant effect on end-of-year math scores. FEV Tutor put the tool into all its sessions but closed in January 2025.
Achieved whenAn outside team repeats the gain with another tutoring provider, or a trial finds gains on end-of-year tests.
Claimed by
Rose Wang, Ana Ribeiro, Carly Robinson, Susanna Loeb and Dora Demszky, Stanford University, latest version October 2026
Open
01
Hints for spoken sessions
A tool that listens to a live voice or video tutoring session, transcribes it as it goes, and, when the tutor presses a key, shows two or three short suggestions for the next move, as Tutor CoPilot does in text chat. Test it only in role-played sessions in which consenting adult volunteers play students from scripts with planted mistakes.
The briefCredit · unclaimed02
Notes for the next tutor
An assistant that reads a student's past session transcripts and writes a one-page handover note for a tutor who has never met the student: what was covered, what the student still gets wrong, and what was promised for next time. Test it on made-up students only: experienced tutors write five-session histories of fictional students, each with a hidden list of ten facts the next tutor must know.
The briefCredit · unclaimed
Open 01 · the brief
Hints for spoken sessions
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A tool that listens to a live voice or video tutoring session, transcribes it as it goes, and, when the tutor presses a key, shows two or three short suggestions for the next move, as Tutor CoPilot does in text chat. Test it only in role-played sessions in which consenting adult volunteers play students from scripts with planted mistakes.
Why it is still open
Tutor CoPilot was built for a chat-based platform, and its authors list speech (student and tutor voice) and vision (the shared whiteboard) as future work. This research found no published voice version or test of one.
Built with what exists
Streaming speech-to-text such as ElevenLabs Scribe v3 (see the Note "Fewer word errors") or OpenAI's Whisper; an AI model such as Claude or Gemini; the Tutor CoPilot authors' public code (github.com/rosewang2008/tutor-copilot) as a starting point; a small web page that sits beside the video call.
First step
Write ten scripts of a student making common fractions mistakes, record two volunteers acting them over a video call, and time transcription plus suggestion for each tutor request.
Done looks like
In 100 scripted 20-minute role-played sessions, suggestions appear within 3 seconds of the key press in at least 95% of requests, and two experienced tutors, scoring blind, rate at least 70% of suggestions usable without edits, with a published log of every request, suggestion and rating.
Open 02 · the brief
Notes for the next tutor
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An assistant that reads a student's past session transcripts and writes a one-page handover note for a tutor who has never met the student: what was covered, what the student still gets wrong, and what was promised for next time. Test it on made-up students only: experienced tutors write five-session histories of fictional students, each with a hidden list of ten facts the next tutor must know.
Why it is still open
Students often meet a different tutor each session. In a study of 15,959 learners on one online language-tutoring platform, about 40% changed to a new tutor every session, and more changes went with slower gains in speaking scores. The Tutor CoPilot trial's provider matched tutors to students almost at random each session. Stanford's tutoring toolkit recommends consistent pairings. This research found no published test of AI-written handover notes.
Built with what exists
An AI model such as Claude or Gemini, or an open-weights model such as Mistral Large 4 (see the Note "A trillion open parameters") so a program can keep notes on its own servers; a scoring sheet; two scorers.
First step
Write three fictional five-session histories with their hidden fact lists and check whether one model's notes recover the facts.
Done looks like
Across 30 fictional histories, the notes contain at least 9 of the 10 hidden facts and no false statements in at least 27 of 30, scored by two tutors who did not write the histories; and volunteers playing the new tutor answer more of 10 questions about the student correctly after 10 minutes with the note than after 10 minutes with the raw transcripts.