This is a thirteen-page report published by OpenAI in August 2026, described on its cover as a "Back-to-School Report | Summer 2026." It combines internal usage telemetry with named profiles of teachers and students, an account of a summer training program, and a short review of external evidence on AI and learning outcomes.
Every usage figure in it is OpenAI's own measurement of its own products, described as "privacy-preserving analysis." The report states no sample definition, no confidence interval and no independent verification for any internal estimate, and gives a date window for only two of them. Its confidence rating here reflects that: the figures are the company's, reported by the company, about the company's products.
Usage figures
Across all age groups, the report identifies "as many as 70 million conversations for self-testing purposes per week," characterised by "repeated attempts at answers, answer-checking, misconception checks, and requests for more practice."
In the United States, prompts related to classwork and homework "peak during the school year at more than 460 million messages per week, with activity climbing on Sunday evenings before the school week begins." During summer months the analysis found "more than 180 million coursework-related messages per week." The hourly distribution chart covers January 5 to July 19, 2026.
Within the self-testing conversations, the report gives behaviour shares: 59% involved iterative learning, 63% included answer-checking, 58% included the learner's own attempt, 51% requested practice, and 36% checked misconceptions. A caption states that "[t]hese estimates describe patterns of use, not learning outcomes," and that a single conversation can contain multiple behaviours. People aged 18–21 "accounted for around 25% of users and 30% of messages in the broader homework analysis."
On the teacher side, an analysis of data from January 1 to July 16, 2026 found "more than 1.9 million messages related to tasks that help educators save time": 900,000 about report cards and progress reports, 800,000 about lesson planning, and more than 100,000 each about substitute plans and teacher-evaluation materials. Separately the analysis identified roughly 1.5 million translation-related messages "so far this year," including approximately 575,000 involving communication with parents and families.
The report's own hedges
The document is explicit about the limits of what its telemetry shows, and the hedges are load-bearing. Of a classroom result in which students scored 23% above the school average on a district writing assessment after a warm-up exercise was used twice a week for two weeks, it states: "That comparison does not establish that AI caused the result." Of the behaviour shares: "This broader analysis does not yet establish how often these interactions reflect productive practice versus quick-answer seeking or cognitive offloading." Of the field generally: "Evidence about AI and learning is still early and mixed. General-purpose chatbot use is not itself a learning intervention; outcomes depend on whether students are required to think, attempt, revise, and receive appropriate teacher guidance."
External evidence cited
Four external studies are summarised, two of them randomized trials with results that cut in different directions.
A randomized Harvard study of "a carefully designed AI physics tutor" produced "more than twice the median immediate learning gains of an active-learning classroom," with higher reported engagement and motivation.
A randomized study of nearly 1,000 high-school mathematics students in Turkey found that "unrestricted AI improved practice performance but reduced later unaided test performance by 17%," while "a tutor configured with teacher-designed hints largely avoided that effect." The report draws the conclusion itself: "whoever does the work does the learning."
A World Bank randomized study in Nigeria found that six weeks of teacher-guided, GPT-4-powered tutoring improved performance on a combined assessment of English, AI knowledge and digital skills.
A nationwide ChatGPT Edu deployment in Estonia is reaching "more than 20,000 students and 4,600 teachers," with the University of Tartu, Stanford and OpenAI studying classroom outcomes. The report states that "Estonia's implementation and research are underway; learning outcomes have not yet been established."
For background the report cites Benjamin Bloom's 1984 "2 Sigma Problem" on one-to-one tutoring and a later review of 89 randomized studies by researchers at Northwestern University and the University of Toronto finding consistent gains from tutoring in reading and mathematics. It cites the 2022 Programme for International Student Assessment for the finding that "47% of socioeconomically disadvantaged students across OECD countries scored below basic proficiency in mathematics, compared with just 14% of advantaged students."
Named practitioners
Six practitioners are profiled, with time-saving estimates attributed to them individually rather than measured. Casey Cuny, California's 2024 Teacher of the Year and an English teacher at Valencia High School, describes an exercise he calls the "Elaboration Conversation" and a classroom rule: "humans draft; AI feedback, humans finish." Ava Morton, an eighth-grade educator in Clayton County, Georgia, adapts texts for students with ADHD, dyslexia, autism and other learning differences and "estimates that ChatGPT saves her at least five hours of planning each week." Kara Crown, a special-education teacher outside Chicago, reports cutting the time to draft individualized education plan goals and forms "from more than an hour per student to less than 30 minutes." Taiyo Inoue, a college mathematics instructor in California, uses Codex for course administration in Canvas and "estimates that these workflows save him four to five hours a week." Doreen Mayrell teaches dual-credit college algebra at Collin College in Texas and builds lesson-specific custom GPTs. Brandon Pieczka, an Iowa State University software-engineering student, used Codex during a Pinterest internship and "documented 20 experiments" by his second week.
The AI Skills Jams
The report describes a series of "AI Skills Jams for K–12 Educators" launched in summer 2026 by OpenAI Academy and the Walton Family Foundation, held in Jonesboro, Georgia; Fairfax, Virginia; Orlando; San Bernardino, California; Phoenix; Salt Lake City; and Las Vegas. Participants "received a year of free access to ChatGPT Education." Initial survey data is reported as "93 percent of participants said they left with something they could use again, and 96 percent planned to apply what they learned within 30 days"; no sample size or response rate is given.
Christina Ordonez, technology department chair at Hoffman Estates High School in Illinois's Township High School District 211, is profiled for extending AI training beyond teachers to "librarians, athletic trainers, hall monitors, bus drivers, and hundreds of other staff members."
The report states a commercial purpose for the program alongside the training one: the workflows educators developed "show us where teachers lose time, where students need more individualized support, and where educators encounter friction when using AI," observations that "can help our product and engineering teams improve tools for schools."
Relationships
- supports: Education — AI Deployment, AI in Education.
- contradicts: its own framing, in part — the Turkey randomized result it reports is evidence against unrestricted student use of general-purpose AI.
- depends-on: OpenAI.
- related: The Turbulent AI Era Is Here. The Choices We Make Now Are Critical. (Bill Gates, August 2026), Cognitive Friction.