Industries / Government, education and community / Education
Testing and grading
Writing exams, marking answers and turning results into grades, from classroom quizzes to national exams and international comparisons.
7
Milestones achieved, from 581–618 to 2023.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
581–618
Officials chosen by exam
Under the Sui dynasty, China began filling government posts through competitive written exams instead of family rank. The system lasted about 1,300 years, until the Qing court ended it in 1905.
Credit
the Sui dynasty court
1915
One right answer per item
Frederick J. Kelly built the Kansas Silent Reading Test so that every item was wholly right or wholly wrong and any teacher could score it quickly. Many items asked the child to circle one word from a short printed list. Kelly reported giving it to 9,252 children in 19 Kansas cities in May 1915, and the historian Cathy Davidson calls him the author of the first multiple-choice test.
Credit
Frederick J. Kelly, Kansas State Normal School at Emporia
1937
Machines score pencil marks
IBM began selling the 805 Test Scoring Machine, which found pencil marks on an answer sheet because pencil graphite carries electric current. IBM says it made scoring a 150-question test 5 to 10 times faster, and it stayed in production until January 1963.
Credit
Reynold B. Johnson, a Michigan high-school teacher hired by IBM in 1934; IBM
1993
Tests that adapt
ETS moved the GRE General Test, a US graduate-school admission exam, to a computer-adaptive format, in which the computer picks each question based on how the taker did on earlier ones. In late 1994 the test-prep company Kaplan showed ETS parts of about 200 questions that people taking the test had memorized, and ETS suspended the adaptive test and then limited it for several months.
LessonAn adaptive test reuses a limited pool of questions, so takers can memorize them and pass them on.
Credit
Educational Testing Service
2000
One test, 32 countries
The OECD ran the first PISA survey, in which 15-year-olds in 32 countries took the same written reading, math and science tasks under independent supervision. It has repeated every three years since and moves to a four-year cycle after 2025.
Credit
OECD
2020
England drops its grading formula
With exams canceled, Ofqual, England's exam regulator, used a statistical model to adjust teachers' estimated A-level grades toward each school's past results. Ofqual's own table shows 39.1% of 718,276 A-level grades came out lower than the teacher's grade, about 280,000 grades by this file's arithmetic. On 17 August 2020 Ofqual switched to teacher grades and apologized.
LessonA model built to keep national results in line with past years moved hundreds of thousands of individual grades, and it was withdrawn within days of results day.
Credit
Ofqual
2023
Texas machine-scores written answers
The Texas Education Agency began scoring open-ended answers on STAAR, the state's school tests, with an "automated scoring engine", on a limited basis in December 2023 and fully in spring 2024. Humans rescore a quarter of responses. The agency said it would need fewer than 2,000 temporary human scorers, down from about 6,000, and expected to save $15 million to $20 million a year.
Credit
Texas Education Agency
Watch for
AI marks like examiners
Researchers at the company Medly AI built a set of 32,534 student answers to its own GCSE mock exams, each marked by two examiners, across 328 questions in five subjects. They report that the best off-the-shelf AI models agreed with the examiners more closely than the examiners agreed with each other, measured by quadratic weighted kappa, an agreement score where 1 is perfect: 0.75 against 0.65 in English, 0.86 against 0.84 in math and 0.74 against 0.69 in science.
Achieved whenAn exam board, or Ofqual, repeats the test on live exam scripts and publishes the result. In January 2026 Ofqual said AI is "nowhere near ready to take over high stakes marking".
Claimed by
Malachy Fox, Kavi Samra and Paul Jung, Medly AI, June 2026
Oral exams run by AI
At NYU Stern, a voice AI ran personal oral exams for 36 students in fall 2025 and 37 in spring 2026, and a panel of three AI models graded each transcript. The authors say grading cost under $1 per exam and a voice subscription covered the speaking time.
Achieved whenAnother institution runs such exams and publishes how often the AI grades match human examiners' grades.
Claimed by
Panos Ipeirotis and Konstantinos Rizakos, NYU Stern, March 2026
Open
01
Proofread the exam paper
A checker that works through an exam paper and its mark scheme before the exam is sat. It solves each question independently several times and flags missing information, impossible or ambiguous questions, wrong mark-scheme answers, and differences between the standard paper and its braille or large-print versions.
The briefCredit · unclaimed02
Score early reading in Hausa
An app that listens to a child reading an early-grade reading assessment in Hausa (letter sounds, simple words and a short passage), marks each item right or wrong, and counts words read correctly per minute. Early-grade reading assessments are short one-to-one tests in which a trained assessor listens and marks each item by hand. Collect recordings only through a school or research partner, with parents' consent and ethics approval.
The briefCredit · unclaimed
Open 01 · the brief
Proofread the exam paper
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A checker that works through an exam paper and its mark scheme before the exam is sat. It solves each question independently several times and flags missing information, impossible or ambiguous questions, wrong mark-scheme answers, and differences between the standard paper and its braille or large-print versions.
Why it is still open
In summer 2025, England's exam boards reported 72 errors in the content of GCSE, AS and A-level question papers, 17 of them in Ofqual's most serious category, which covers errors that can make a meaningful answer impossible. Forty-three were found only after the exam, and nine occurred only in braille versions. Ofqual's January 2026 statement sees promise for AI in the quality assurance of marking, and this research found no exam board using AI to check papers before they are sat.
Built with what exists
An AI model that works through math and science, such as Claude Fable 5.2 or Gemini, with code execution to recheck calculations (the Note "Seven more flagged" covers a Claude-based reviewer of long documents); liblouis, the open-source braille translator, to turn braille papers back into print for comparison; the past papers and mark schemes that exam boards publish.
First step
Take ten past papers from one board in one subject, plant ten errors, and measure what a single model run finds.
Done looks like
On 50 published past papers into which a teacher has planted 100 errors of the kinds Ofqual counts, the checker finds at least 80 with no more than one false flag per paper; any real errors it finds in published papers are listed and sent to the boards.
Open 02 · the brief
Score early reading in Hausa
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An app that listens to a child reading an early-grade reading assessment in Hausa (letter sounds, simple words and a short passage), marks each item right or wrong, and counts words read correctly per minute. Early-grade reading assessments are short one-to-one tests in which a trained assessor listens and marks each item by hand. Collect recordings only through a school or research partner, with parents' consent and ethics approval.
Why it is still open
This research found such work for Kiswahili, where a 2026 NeurIPS poster reports an 18.3% word error rate on child speech, and for isiXhosa, where a 2025 dataset is available on request, plus a Gates Foundation-funded pilot in South Africa that has not published results. It found none for Hausa.
Built with what exists
Meta's MMS speech model, which lists Hausa among 1,162 languages, under a non-commercial license; NVIDIA's NeMo FastConformer models, the design used in the Kiswahili work; a commercial speech-to-text service as a baseline (the Note "Fewer word errors" covers ElevenLabs Scribe v3, rated on 102 languages).
First step
Record 20 adult Hausa speakers reading a sample passage with planted errors, and measure MMS's word error rate on them.
Done looks like
On 300 recordings, each scored by two trained assessors, the app's item marks agree with the assessors as often as the assessors agree with each other, and its words-correct-per-minute is within 5 of their average for at least 90% of readers.