Industries / Government, education and community / Education
Tutoring and practice
One-to-one or small-group teaching, and the drills, software and study habits students use to practice on their own.
7
Milestones achieved, from 1885 to 2026.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
In depth
The same ground is charted in more depth, as an industry of its own.
The road
AchievedWatch forOpen
Achieved
1885
Spacing beats cramming
Hermann Ebbinghaus memorized lists of nonsense syllables and timed how long it took him to relearn them, testing only himself. He found that memory falls fast in the first hours and then slowly, and that repetitions spread over time work better than repetitions massed in one sitting.
Credit
Hermann Ebbinghaus
1926
A machine that teaches
In the 1920s Sidney Pressey built machines that showed one multiple-choice question at a time and moved on only after the right answer. In 1926 he wrote that such a machine could teach as well as score, because the student learns at once whether an answer is right.
LessonThe machines did not catch on, and in 1932 Pressey wrote that he was "regretfully dropping further work"; B. F. Skinner took up the idea again in a 1958 paper in Science.
Credit
Sidney L. Pressey
1984
The two-sigma claim
Benjamin Bloom reported studies in which students taught by a good tutor scored about two standard deviations above students in regular classes of about 30. A standard deviation measures how spread out scores are; two above average meant the typical tutored student beat 98% of the regular class.
LessonIn 2011 Kurt VanLehn reviewed many experiments and put human tutoring's effect at 0.79 standard deviations, not 2, with step-by-step tutoring software close behind at 0.76.
Credit
Benjamin S. Bloom, University of Chicago
1987
Spaced review in software
In December 1987 Piotr Woźniak in Poznań, Poland, wrote the first SuperMemo program. Its SM-2 rule sets each flashcard's next review at 1 day, then 6 days, then longer gaps scaled to how easy the card has been. Anki, a free flashcard program, built its scheduler on SM-2, and in October 2023 added FSRS, a scheduler that learns each user's memory patterns.
Credit
Piotr Woźniak (SuperMemo); the Anki project
2014
Tutor software, tested by lottery
Cognitive Tutor Algebra I is an algebra course built around tutoring software from research by John Anderson, Albert Corbett, Kenneth Koedinger and Ray Pelletier at Carnegie Mellon. RAND paired similar schools in seven states and picked one of each pair by chance to use it. There was no effect in the first year; in the second, high-school students gained about 8 percentile points.
LessonA one-year trial would have shown no benefit.
Credit
Anderson, Corbett, Koedinger and Pelletier (the research); John Pane and colleagues at RAND (the trial)
2019
Practice matched to level
Mindspark, software from the Indian firm Educational Initiatives, matches each student's exercises to that student's measured level. In a trial with 619 Delhi public middle-school students, those who won a lottery for free places at after-school Mindspark centers scored 0.37 standard deviations higher in math and 0.23 higher in Hindi after 4.5 months.
LessonEach session paired 45 minutes on the software with 45 minutes of help from a teaching assistant, so the result applies to that mix, not to software alone.
Credit
Educational Initiatives; Karthik Muralidharan, Abhijeet Singh and Alejandro Ganimian
2026
An AI tutor, rarely used
In a two-year randomized trial in 18 Tennessee middle schools, students assigned to Khan Academy with its AI tutor, Khanmigo, set to coach rather than give answers, gained about 0.06 to 0.08 standard deviations in math over a school year. The authors say the gains resemble those from Khan Academy practice without AI. The median student messaged the tutor on a third of practice days, and in 17% of the sessions in which they made a mistake.
LessonAccess was not the limit; getting students to use the tutor was.
Credit
Khan Academy (Khanmigo); evaluators Philip Oreopoulos (University of Toronto) and Nina Low, funded by J-PAL, the Smith Richardson Foundation and Canadian research councils
Watch for
AI drafts, tutors approve
In an exploratory randomized trial with 165 students in five UK secondary schools, Google's LearnLM model drafted messages to students on Eedi's math platform, and 17 human tutors approved or edited each one. Students supported this way solved new problems on later topics 66.2% of the time, against 60.7% with human tutors alone. The paper says tutors approved 76.4% of drafts with zero or minimal edits; Eedi's own release says 82.3%.
Achieved whenThe larger trials Eedi has announced, in the US with Imagine Learning in 2026 and in the UK on long-term gains, report learning results, preferably measured by outside evaluators.
Claimed by
Google DeepMind and Eedi, December 2025
A tutor that speaks first
Khan Academy says changes it tested in its tutor raised "cognitive engagement", its automated rating of whether each tutoring exchange is at least active, by 5.09%, while math accuracy held steady. The Tennessee trial authors report that in 2026 Khan Academy also redesigned its student experience to turn the tutor on automatically during practice.
Achieved whenAn outside randomized trial finds that the redesigned tutor raises test scores beyond what Khan Academy practice gives without it.
Claimed by
Khan Academy, May 2026
Open
01
Audit AI tutors' math
A public, repeatable test of AI tutors as students meet them. A script plays a student working through released state test questions, sometimes giving right answers and sometimes common wrong ones, and logs every time the tutor accepts a wrong answer, rejects a right one or states false math. Two teachers check each flagged exchange. Run it through each maker's programming interface or with the maker's permission.
The briefCredit · unclaimed02
A tutor that watches paper
A tutor that watches a student solve algebra on paper through a camera above the desk and speaks up at the first wrong line, before the student reaches a final answer.
The briefCredit · unclaimed
Open 01 · the brief
Audit AI tutors' math
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public, repeatable test of AI tutors as students meet them. A script plays a student working through released state test questions, sometimes giving right answers and sometimes common wrong ones, and logs every time the tutor accepts a wrong answer, rejects a right one or states false math. Two teachers check each flagged exchange. Run it through each maker's programming interface or with the maker's permission.
Why it is still open
The tutor accuracy figures found come from the makers' own tests; Khan Academy's May 2026 post says only that "math accuracy held steady" in its internal tests. The independent Tennessee trial measured learning, not the tutor's mistakes. This research found no side-by-side audit of the tutors students use.
Built with what exists
Released test questions that several US states publish each year; a web agent to drive tutors that have no programming interface (see the Note "Deleting 80% to start over"); SymPy, an open-source math library, to check answers; a shared spreadsheet for the teachers' checks.
First step
Put 50 questions to one tutor by hand, save every exchange, and have a teacher mark the errors to set the categories.
Done looks like
Three tutors, 500 questions each, two student personas; error rates by type with counts and full transcripts published; the test rerun after each tutor's next major update.
Open 02 · the brief
A tutor that watches paper
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A tutor that watches a student solve algebra on paper through a camera above the desk and speaks up at the first wrong line, before the student reaches a final answer.
Why it is still open
The closest published tool found, MathMistake Checker (AAAI 2025), finds mistakes step by step in written answers, and its paper reports no accuracy figures. This research found no tutor that follows work on paper as it is written and reports a measured accuracy.
Built with what exists
A document camera or a phone on a stand; a vision-language model such as Gemini or Claude to read each new line; SymPy to check whether each line follows from the one before; text-to-speech.
First step
Record 20 solutions to linear equations with planted errors and test how well the model picks out each new line from the video.
Done looks like
On 200 recorded solutions, written by adult volunteers who copy real student errors, the tutor flags the first wrong line within 10 seconds in at least 90% of cases, with no more than one false alarm per 10 correct solutions, and the videos and results are published.