BRENNEN

Industries / Government, education and community / Education / EdTech

Practice and tutoring software

Software that gives a learner exercises, hints and feedback without a teacher at hand, from drill programs to AI tutors.

5

Milestones achieved, from 1963 to 2023.

3

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1963

    Drill and practice by computer

    Patrick Suppes began computer-assisted instruction at Stanford in 1963, and his group's first demonstration programs ran that year on a PDP-1 computer. In 1967 he founded Computer Curriculum Corporation to sell such courses to schools. By late 1978, by his own count, more than 150,000 students in 24 states used its reading and math drill-and-practice courses almost every day, most of them described as disadvantaged or handicapped.

    suppescorpus.stanford.edusuppescorpus.stanford.edubangordailynews.com

    Credit

    Patrick Suppes, Institute for Mathematical Studies in the Social Sciences, Stanford University; Computer Curriculum Corporation

  2. 2012

    Language drills in an app

    Duolingo, founded by the engineers Luis von Ahn and Severin Hacker, opened its free language-learning app to the public on 19 June 2012. The company describes its approach as mobile-first and gamified, and it counts how many days in a row each learner practices. In the last three months of 2025 it averaged 52.7 million daily users and 133.1 million monthly users, 12.2 million of them paying, by the company's figures in its annual report to the US Securities and Exchange Commission. These counts measure use, not learning.

    sec.gov

    Credit

    Luis von Ahn and Severin Hacker; Duolingo

  3. 2016

    Homework hints raise scores

    ASSISTments, free online homework software from Worcester Polytechnic Institute, gives students hints and immediate feedback and shows teachers which problems their class got wrong. In a randomized trial in 43 Maine schools with 2,850 seventh graders, students whose teachers used it, with training, scored 0.18 standard deviations higher on an end-of-year math test than students doing homework as usual. Students with low prior math scores gained most.

    files.eric.ed.gov

    Credit

    Worcester Polytechnic Institute (ASSISTments); Jeremy Roschelle, Mingyu Feng and Robert Murphy of SRI International and Craig Mason of the University of Maine (the trial)

  4. 2018

    Students walk out on Summit

    Summit Learning was a self-paced online curriculum, designed by Facebook engineers and funded by Mark Zuckerberg and Priscilla Chan, according to the New York Post. In November 2018 nearly 100 students at Brooklyn's Secondary School for Journalism walked out of class against it; one freshman said he spent close to five hours a day on it. New York City's education department said the school would drop it in grades 11 and 12, and that another school had already dropped it. In McPherson, Kansas, more than 30 middle-school students protested outside on 29 January 2019, and the district began re-evaluating the program.

    LessonStudents objected that the software asked them to teach themselves on screens, and their protests cut its use.

    nypost.comspectator.mcpherson.edu

    Credit

    Summit Learning; the students of both schools

  5. 2023

    Chatbots undercut homework help

    Chegg sold step-by-step homework solutions written by paid experts. On 1 May 2023 it told investors that a spike in student interest in ChatGPT since March was hurting its growth in new customers. Its quarterly revenue fell from $187.6 million in early 2023 to $77.7 million in the third quarter of 2025, and in October 2025 it cut 388 jobs, about 45% of its staff, citing "the new realities of AI and reduced traffic from Google".

    LessonA paid library of answers lost its customers once a free chatbot could write answers on demand.

    sec.govsec.govsec.gov

    Credit

    Chegg (figures from its filings); OpenAI (ChatGPT)

Watch for

  1. AI tutor beats active class

    In a randomized trial with 194 students in Harvard's introductory physics course for life-science students, in fall 2023, each student had one lesson from a custom AI tutor at home and one active-learning class, in consecutive weeks. The authors report that students learned more from the tutor, in a median of 49 minutes against a 60-minute class, with an effect of 0.73 to 1.3 standard deviations by their estimate.

    Achieved whenAnother team repeats the result with its own students, over more than two lessons.

    pmc.ncbi.nlm.nih.gov

    Claimed by

    Greg Kestin, Kelly Miller and colleagues at Harvard University, who built the tutor, June 2025

  2. Six weeks of AI tutoring

    In a randomized trial in Edo State, Nigeria, first-year senior secondary students in a six-week after-school English program, held twice a week, used Microsoft Copilot, built on GPT-4, with teachers guiding each session. They scored 0.31 standard deviations higher on a test of English, AI knowledge and digital skills, and 0.23 higher in English alone. The authors say they cannot separate the chatbot's effect from the rest of the program.

    Achieved whenAn outside evaluator finds similar gains in regular classes or at larger scale.

    econpapers.repec.orgblogs.worldbank.org

    Claimed by

    Martín De Simone and colleagues, World Bank, which ran and evaluated the program, May 2025

  3. AI tutor on feature phones

    Yiya Solutions, an education group in Uganda, says its AirVoice tutor lets learners dial in from a basic feature phone, with no internet, and talk through science and technology lessons by voice. In a pilot with 100 students in two districts, it reports that 76% of students' exchanges with the tutor were scored "very accurate", meaning the tutor understood the learner and replied as expected; the case study does not say who scored them or how.

    Achieved whenAn outside evaluator publishes how accurately the tutor understands learners and what they learn from it.

    team4tech.org

    Claimed by

    Yiya Solutions, in a case study published by Team4Tech, November 2025

Open

  1. 01

    Test placement with scripted students

    A harness that sends 40 scripted test students through the adaptive placement tests of math practice products and checks where each product places them. Each scripted student is a small program with a known skill profile written as answer rules, for example "adds fractions with like denominators correctly, but adds the denominators when they differ". The harness reads each question, works out the answer that profile would give and enters it. Run it only through teacher or demo test accounts, with each maker's permission; no real student takes part.

    The brief
    Credit · unclaimed
  2. 02

    Spoken answers over phone lines

    A public test set and scoreboard for tutors that work over ordinary phone calls. Adult volunteers read scripted math answers, such as "three quarters", "forty-seven" or "I don't know", over real mobile calls in English and in two languages of countries where phone tutoring has been tried, such as Kiswahili and Luganda, at set levels of background noise and with scripted wrong answers and long pauses. Each speech-to-text system, and each tutor that allows testing, is scored on whether it hears and marks every answer as the script says it should.

    The brief
    Credit · unclaimed

Open 01 · the brief

Test placement with scripted students

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A harness that sends 40 scripted test students through the adaptive placement tests of math practice products and checks where each product places them. Each scripted student is a small program with a known skill profile written as answer rules, for example "adds fractions with like denominators correctly, but adds the denominators when they differ". The harness reads each question, works out the answer that profile would give and enters it. Run it only through teacher or demo test accounts, with each maker's permission; no real student takes part.

Why it is still open

A 2025 review describes simulated students being used to test teaching approaches and train teachers, and a 2026 study found that one AI model playing a whole class overrates weak students and underrates strong ones, a reason to use rule-based profiles with known answers instead. This research found no published test that runs simulated students through commercial placement tests and compares the placement with a known profile.

arxiv.orgaclanthology.org

Built with what exists

Python; SymPy, an open-source math library, to compute answers; an AI model to tag each question with the skill it tests; a browser agent such as Browser Use to read questions and enter answers where a product has no programming interface (see the Note "Deleting 80% to start over"); a spreadsheet of profiles and their expected placements.

  1. First step

    Write five profiles for grade 4 to 6 fractions and take them by hand through one free placement test, recording the level it reports for each.

  2. Done looks like

    40 profiles run three times each through at least three products; for each product, the share of runs placed within one grade level of the profile's known level, the number of questions it took, and how often repeat runs agree, published with the profiles and the harness code.

Open 02 · the brief

Spoken answers over phone lines

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test set and scoreboard for tutors that work over ordinary phone calls. Adult volunteers read scripted math answers, such as "three quarters", "forty-seven" or "I don't know", over real mobile calls in English and in two languages of countries where phone tutoring has been tried, such as Kiswahili and Luganda, at set levels of background noise and with scripted wrong answers and long pauses. Each speech-to-text system, and each tutor that allows testing, is scored on whether it hears and marks every answer as the script says it should.

Why it is still open

Phone tutoring by people raised math learning by 0.33 standard deviations on average across five replication trials, at about $12 per student, by J-PAL's account, so an AI version that hears answers reliably could reach many more children. Yiya Solutions reports that 76% of its AI phone tutor's exchanges were scored "very accurate" but does not publish its test, and an automated phone tutor on GitHub, IVR Tutor, plays recorded lessons with keypad quizzes and reports no test. This research found no public set of spoken answers recorded over phone lines in these languages.

povertyactionlab.orgteam4tech.orggithub.com

Built with what exists

A phone-calling service such as Twilio or Africa's Talking to place and record calls; speech-to-text that covers the chosen languages, such as Meta's MMS model or ElevenLabs Scribe v3 (see the Note "Fewer word errors"); Pipecat, an open-source framework for voice agents, to run each system the same way; a script sheet giving the right mark for every answer.

  1. First step

    Record 20 adult volunteers saying 25 scripted number answers in Kiswahili over a phone line and measure how often two speech-to-text systems get the numbers right.

  2. Done looks like

    2,000 recorded answers from at least 40 consenting adult volunteers in three languages, published under an open license; for at least three speech-to-text systems, the share of answers heard and marked as the script says, by language and noise level.