Industries / Government, education and community / Education / EdTech
Assessment
Testing what learners know with software: scoring essays and answers by machine, giving tests on screen, and checking that work is the learner's own.
6
Milestones achieved, from 1964 to 2025.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1964
A computer grades essays
Ellis Page began Project Essay Grade at the University of Connecticut in December 1964, with pilot funding from the College Entrance Examination Board from February 1965. His program counted features that track writing quality without reading for meaning, such as essay length, average word length and the number of common words. On 272 essays from grades 8 to 12, typed onto punched cards, these counts matched the pooled judgments of at least four teachers with a multiple correlation of 0.71.
LessonPage wrote that his programs largely ignored the order of words, so the scores rested on features that go with good writing rather than on the writing itself.
Credit
Ellis B. Page, University of Connecticut
1999
Machine as second marker
Educational Testing Service (ETS) has used its e-rater program to score essays since 1999. On the GMAT, an entrance exam for business schools, e-rater serves as a second rater alongside one human rater, so a person still reads every essay.
Credit
Educational Testing Service
2018
Australia shelves machine marking
ACARA, Australia's curriculum authority, planned to mark the writing section of NAPLAN, the national school tests, with automated essay scoring as the tests moved online, at first alongside a human marker. After objections from teachers' groups, the Education Council of education ministers decided in December 2017, and announced in January 2018, that automated essay scoring would not be used for NAPLAN writing.
LessonThe council said any future change would be "informed by further research into automated essay scoring", so evidence had to come before a national test was marked by machine.
Credit
COAG Education Council (chair Susan Close); NSW Teachers Federation (opposition)
2021
Proctoring misses planted cheaters
At the University of Twente in the Netherlands, 30 computer science students took a mock exam under Proctorio, software that watches test-takers through their webcam and flags suspicious behavior by algorithm; six had been asked to cheat. Proctorio flagged none of the six, and staff reviewing the recordings caught one. A 2022 study of four proctoring programs used for US law school and bar exams found that all their anti-cheating measures "can be trivially bypassed".
LessonThe Twente authors compared proctoring to a placebo: it may deter cheating because students believe it works, not because it catches cheaters.
Credit
Bergmans, Bouali, Luttikhuis and Rensink, University of Twente; Ben Burgess, Edward W. Felten and colleagues, Princeton University
2023
Detectors flag non-native writers
OpenAI released a classifier for AI-written text on 31 January 2023, saying that in its own test it caught 26% of AI-written text and wrongly flagged 9% of human text, and withdrew it on 20 July 2023 "due to its low rate of accuracy". A Stanford team tested seven detectors that year: they wrongly labeled an average of 61.3% of 91 essays written by non-native English speakers for the TOEFL English test as AI-generated, while judging 88 US eighth-grade essays almost perfectly.
LessonDetectors that score how predictable the words are mistake the plainer vocabulary of non-native writers for machine text.
Credit
OpenAI; Weixin Liang, James Zou and colleagues, Stanford University
2025
Outside test finds accurate detector
Economists Brian Jabarian and Alex Imas of the University of Chicago, who declare no conflicts of interest, ran three commercial detectors (Pangram, OriginalityAI and GPTZero) and one open-source detector on 1,992 human passages written before 2020 and on matched AI passages. Pangram had near-zero false positives and false negatives and was the only tool to keep false positives at or below 0.5% without losing accuracy. The human passages were reviews, blogs, news, novel excerpts and résumés; none were student essays.
Credit
Brian Jabarian and Alex Imas, University of Chicago Booth School of Business; Pangram
Watch for
AI drafts national test items
ETS says its Wilbur system, which uses an AI model to draft test questions that people then review and revise, has been evaluated for NAEP, the US national assessment, in reading, math and science. In three studies, experts accepted a substantial share of the drafts as usable starting points but judged none ready for an operational test, and reviewers agreed on accept-or-reject calls 58% to 80% of the time.
Achieved whenItems drafted this way pass NAEP's pilot testing with students and appear in an operational NAEP test, with their statistics published.
Claimed by
Chris Meador and colleagues, Educational Testing Service, October 2026
Watermarks to flag AI text
Google says more than 10 billion pieces of content have been marked with SynthID, a watermark added to output from its AI models, and in May 2025 it opened early access to SynthID Detector, a portal that checks uploads, including text, for the mark. The announcement gives no accuracy figures, and the portal looks only for Google's own mark.
Achieved whenAn outside test reports how often the text watermark is still found after the kinds of edits students make, and teachers can use the portal.
Claimed by
Google DeepMind, May 2025
Open
01
Test the writing-replay tools
A public test of tools that record how a document was written, such as Turnitin Clarity's draft playback and writing report, Grammarly Authorship, and Google Docs version history. Adult volunteers write 100 short essays: half drafted in the usual way, half made by retyping or lightly editing an AI draft. Teachers who do not know which is which judge each replay, and each tool's labels are tallied.
The briefCredit · unclaimed02
Plant cheaters for AI proctors
A repeatable controlled test of remote proctoring services against today's ways of cheating. Adult volunteers take a mock exam under each service; a known share cheat by set methods, such as a phone running a chatbot out of camera view, camera glasses or a second person in the room, and the rest act honestly, some of them nervously. Count what each service flags. Run it with each vendor's permission, on test accounts.
The briefCredit · unclaimed
Open 01 · the brief
Test the writing-replay tools
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of tools that record how a document was written, such as Turnitin Clarity's draft playback and writing report, Grammarly Authorship, and Google Docs version history. Adult volunteers write 100 short essays: half drafted in the usual way, half made by retyping or lightly editing an AI draft. Teachers who do not know which is which judge each replay, and each tool's labels are tallied.
Why it is still open
Turnitin announced Clarity in March 2025, and UCLA offered it to instructors in Winter and Spring 2026; UCLA says it shows pasted text, writing time and a draft playback, and does not detect AI writing. The writer and teacher Anna Mills noted in 2024 that AI output retyped by hand is labeled human by Grammarly Authorship. Lab studies in 2024 and 2026 trained detectors on typing patterns to spot transcribed AI text, and in June 2026 GPTZero reported 98.5% accuracy for its own detector of text typed in by "autotyper" programs, tested by GPTZero on seven such programs. This research found no outside test of the process tools that schools use.
plagiarismtoday.comgptzero.medts.ucla.eduannamills.substack.comapi.crossref.org
Built with what exists
Trial or institutional accounts for each tool; an AI model to write the drafts that volunteers retype; a scoring sheet for the teachers; Python to compute catch and false-flag rates.
First step
Have five volunteers each write one essay and retype one AI draft in Grammarly Authorship, then compare the reports.
Done looks like
100 sessions by consenting adult volunteers; for each tool, and for the blind teacher review, the share of retyped essays caught and the share of honest essays wrongly flagged, with the essays and replays published.
Open 02 · the brief
Plant cheaters for AI proctors
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A repeatable controlled test of remote proctoring services against today's ways of cheating. Adult volunteers take a mock exam under each service; a known share cheat by set methods, such as a phone running a chatbot out of camera view, camera glasses or a second person in the room, and the rest act honestly, some of them nervously. Count what each service flags. Run it with each vendor's permission, on test accounts.
Why it is still open
The one controlled test found with planted cheaters, at the University of Twente, covered one service and six cheaters and was run before chatbots were widely used. The 2022 study of four proctoring programs took their software apart rather than testing planted cheaters in live sessions. In August 2026 Psytech SA, which promotes the vendor Podium's tools, reported from Podium's own experiment with 415 online participants on an ability test, that people who got answers from AI on a second device were invisible to its behavior monitoring. This research found no outside controlled test of several services since 2023.
Built with what exists
A mock exam on a learning platform's quiz tool; each vendor's trial or education account; phones running chatbots; consumer camera glasses; recordings of every session as ground truth; a test plan published before the runs.
First step
Ask two proctoring vendors for test access and run ten pilot sessions on one, three of them with planted cheating.
Done looks like
At least 60 volunteer sessions per service, 20 of them with planted cheating; for each service, the share of cheaters flagged and of honest takers flagged, beside the rates of human reviewers watching the same recordings, all published.