Nov 11, 2026
9 min read
2 views

A look at public enrolment, exam and employment series around three dates: neural Google Translate (November 2016), DeepL (August 2017) and ChatGPT (November 2022). It describes what the series show and names no cause.
Glasp Research. Keisuke Watanabe, Glasp Inc.
Machine translation became good and free in steps: Google's neural system from November 2016, DeepL in August 2017, ChatGPT in November 2022. We wanted to see what public series show for language learners and for translators.
Before collecting the series for this article, we wrote down a rule and committed it. One exception: an earlier run had already shown us the US employment figures for 2016 to 2025, so the labour side was not blind. For each date we compare the average yearly growth rate in the six years before with the rate from that date to the latest year. We call it a break only if the two differ by at least 2 percentage points a year. Where a series changed its definition inside a window, that window is not scored. The rule describes timing. It cannot tell us why anything moved.
The Modern Language Association counts course enrolments in languages other than English at US colleges. The count was 1,673,566 in fall 2009, 1,418,584 in fall 2016 and 1,182,562 in fall 2021. Per 100 college students, using the federal count of total enrolment, that is 8.2, 7.1 and 6.3. The MLA's own ratio, which leaves out Latin and Ancient Greek and for-profit colleges, is 8.7, 7.4 and 6.5. No later census has been published. The MLA's own report says enrolments have been "in a sustained decline" since "the historic peak in enrollments in 2009".
The fall in this total started seven years before neural translation. The yearly rate was -2.3% from 2009 to 2016 and -3.6% from 2016 to 2021, a steepening of 1.2 points, below our threshold. Per student the steepening is 0.4 points. French and German steepened by more than 2 points, but they were also already falling before 2016. Korean grew in both periods.
High school Advanced Placement exams tell a different story. AP Spanish Language exams rose from 155,258 in 2016 to 182,670 in 2025. AP Chinese rose from 12,524 to 18,312. But all AP exams grew faster than Spanish, so Spanish fell as a share of all AP exams, from 3.3% to 3.0%.
In UK GCSE exams, entries in modern languages fell from 362,338 in 2009 to 321,483 in 2016. French and German fell by 3.3% and 5.4% a year from 2010 to 2016. Our rule cannot score GCSE across 2016, because reformed language GCSEs were first awarded in England in 2018. As plain levels: entries were 338,300 in 2026, above the 2016 figure. French fell to 125,678 and German to 33,473, while Spanish doubled from 67,070 in 2009 to 137,236.
At A level, modern language entries were 35,364 in 2009, 31,183 in 2016 and 24,560 in 2026. As a share of all A level entries they fell from 4.2% to 3.7% to 2.7%. The share fell 1.3% a year before 2016 and 3.1% a year after. That gap is just under our threshold at the November 2016 date and just over it at the August 2017 date. But reformed A levels in French, German and Spanish were first awarded in England in 2018, a change of the same kind that stopped us scoring GCSE, and we did not split the A level series for it. So these two results should be read as plain levels, not as scores.
Japan's public numbers mostly go the other way. In the national time-use survey, the share of people aged 10 and over who studied English in the past year was 9.6% in 2011, 11.9% in 2016 and 12.8% in 2021. Applicants for the Eiken English test grew from 2.51 million in fiscal 2011 to 3.39 million in fiscal 2016 and 4.54 million in fiscal 2025. Growth was slower after 2016, but the earlier years include a one-year jump of 22% in fiscal 2015, and the organiser notes that its earlier counts may not be on the same basis, so the two periods may not be like for like. TOEIC Listening and Reading takers were 2.25 million in fiscal 2017, 1.79 million in fiscal 2022 and 1.81 million in fiscal 2025. These are the yearly counts IIBC gives for its public and in-house tests, which we found from fiscal 2017 on. IIBC's long-run chart uses a larger count (2.48 million for fiscal 2017), so the two series cannot be joined.
Self-study also grew. Duolingo reported 37.0 million monthly active users in late 2020 and 88.4 million in late 2023. These are company-reported figures, and there is no earlier series to compare with.
The US Bureau of Labor Statistics counts wage and salary jobs for interpreters and translators. The self-employed are left out; the bureau's projection files put them at between about a sixth and a third of the occupation. Jobs rose from 40,000 in 2009 to 51,350 in 2016 and 58,870 in 2019. They were 52,160 in 2022 and 52,060 in 2025. Per 100,000 US jobs that is 30.6, 36.6, 40.1, 35.3 and 33.5. In 2025 the median wage was $60,170, against $50,980 for all occupations.
Our rule cannot score this series at any date. In 2021 the bureau changed its estimation method, and every window after 2016, or before 2022, crosses that change. Read across the change anyway, jobs grew 2.5% a year from 2010 to 2016 and 0.2% a year from 2016 to 2025. From 2022 to 2025 the change is 100 jobs, well inside the survey's error of about 3%.
The bureau's own ten-year projections, published every year or two, shifted too. Its 2010 to 2020 projection expected 42% growth. The 2016 to 2026 projection expected 18%, and the 2021 to 2031 one 20%. The 2022 to 2032 projection expected 4.3%, and the latest, 2025 to 2035, expects 2%.
Studies find falls in translation work after neural Google Translate and after ChatGPT. We found none for DeepL. On one translation platform, ordinary translation jobs fell 13 to 20% after neural Google Translate, while creative "transcreation" work did not fall. After ChatGPT, one study of freelance postings found demand for Western European language translation down 23% against its expected trend. Another, on a freelance platform it does not name, found translators' earnings down 29.7% against a matched group of other freelancers. A study of the 2010 Google Translate app, which compares US local labour markets and not a platform, estimated that about 28,000 fewer translator and interpreter positions were created in the US from 2010 to 2023 than in a scenario without Google Translate. These studies cover single platforms or compare regions. The national series above cannot confirm or reject them. An earlier Glasp Research piece looked at the longer history of machine translation and translators' work:
https://glasp.co/posts/twenty-years-of-machine-translation-translators-and-ai
We name no cause. Most of the language declines started before neural translation, and the rule shows when rates changed, not why. Over the same years exam rules changed, college enrolment fell, and a pandemic closed classrooms.
The ChatGPT windows are three or four years long, and there is no US college census after 2021. The UK labour series groups translators with authors and writers, so we do not use it. We found no study that links machine translation to language-course enrolment.
The labour side was not blind, as said above. The figures from the platform studies are as our first pass read them from working-paper copies; later independent checks confirmed three of the four papers' figures in the papers themselves (29.7%, 23%, about 28,000); the fourth (13 to 20%) is as our first pass read it.
We wrote the rule before collecting the series and committed it to our own repository first; that is our record, not a public registry. AI agents collected the data in parallel from public files on 10 November 2026, and we computed the rates with a short script. Four sites refused automated access and were skipped: HESA, for UK university enrolments, the SEC's filing system, Duolingo's investor pages, and the publisher's page for one journal article. One UK exam board file refused a direct download and was read through a different fetch tool. The Eiken figures for fiscal 2011 and 2016 come from the organiser's press releases, not from the page linked below, which shows recent years only. For Duolingo we used the company's press releases. For a few of its requests to one UK exam board's site, our run went faster than our own one-second limit.
Modern Language Association, enrolment census fall 2021 (Table 1a):
NCES Digest of Education Statistics, Table 303.10:
https://nces.ed.gov/programs/digest/d24/tables/dt24_303.10.asp
College Board AP exam volume, 2009 to 2019 and 2015 to 2025 (the second link returned "not found" when re-tested on 11 November 2026):
https://reports.collegeboard.org/media/xls/2019-Exam-Volume-Change_1.xls
https://reports.collegeboard.org/media/xlsx/ap-exam-volume-changes-2015-2025.xlsx
JCQ results tables, 2026 GCSE and A level (earlier years from JCQ's yearly results files of the same kind):
Statistics Bureau of Japan, Survey on Time Use and Leisure Activities, time series table on learning, via e-Stat:
https://www.e-stat.go.jp/stat-search/file-download?statInfId=000032229292&fileKind=0
Eiken Foundation of Japan, test-taker statistics:
https://www.eiken.or.jp/eiken/about/situation/
IIBC, TOEIC Program data and analysis, 2020 and 2026 editions:
https://www.iibc-global.org/hubfs/library/default/iibc/press/2020/p155/pdf/2020DAA.pdf
BLS Occupational Employment and Wage Statistics, national file for May 2025 (one such file per year was used):
https://www.bls.gov/oes/special-requests/oesm25nat.zip
BLS Employment Projections archive:
https://www.bls.gov/emp/data/projections-archive.htm
BLS Occupational Outlook Handbook, interpreters and translators (the 2025 to 2035 projection and the May 2025 median wages):
https://www.bls.gov/ooh/media-and-communication/interpreters-and-translators.htm
UK government note on A level reform dates:
Yilmaz, Naumovska and Aggarwal, INSEAD working paper 2023/24/EFE:
https://sites.insead.edu/facultyresearch/research/doc.cfm?did=70468
Teutloff and others, Journal of Economic Behavior and Organization 235 (2025):
https://discovery.ucl.ac.uk/10204198/1/1-s2.0-S0167268124004591-main.pdf
Qiao, Rui and Xiong, arXiv:2312.04180:
https://arxiv.org/pdf/2312.04180
Frey and Llanos-Paredes, Lost in Translation (2025):
This article was researched and written by Claude (an AI model by Anthropic) in a Glasp research run. No Glasp user data was used. Checks: a rule written and committed before any series was read; every figure read from the publisher's file, except Duolingo's, which came from copies of its press releases on two finance sites; a second pass by the same model re-opening each saved source and comparing each number in this draft; on 11 November 2026 a separate AI run that did not write the text recomputed every rate and share from the collected series and found one wrong label, now corrected, but could open only some of the linked sources; on 12 November 2026 a further AI check read the downloaded source files and confirmed the MLA totals and quotations, the 2016 AP counts, the TOEIC fiscal 2017 and 2025 counts and three of the four platform-study figures. The UK exam counts, the 2025 AP counts, the TOEIC fiscal 2022 count, the BLS figures, the Japanese survey and Eiken figures and Duolingo's figures still rest on the first pass. No human has checked these figures.