Twenty Years of Machine Translation: What Happened to Translators, and What It Says About AI and Text Work

Glasp Research

Glasp Research

Oct 31, 2026

15 min read

Twenty Years of Machine Translation: What Happened to Translators, and What It Says About AI and Text Work

Summary

Machine translation has been free and getting better since 2006. If any profession can tell us what AI does to text work, it is translators. We compiled US public data from 2005 to 2025 and read it next to the causal studies and the latest industry surveys. What the record shows:

  • Official headcount has not collapsed. US wage-and-salary employment of interpreters and translators was 51,350 in May 2016, 58,870 in May 2019 and 52,060 in May 2025 (BLS OEWS). Survey-based employment including the self-employed was about 97,500 in 2024, statistically indistinguishable from 2022 (ACS, our tabulation). For the Google Translate era, one causal study estimates that machine translation slowed employment growth rather than cutting jobs. For the LLM era, the official series are still inconclusive.

  • The clearest signals so far come from places official statistics do not look: online freelance platforms, where demand for writing and translation work fell 20 to 50 percent relative to trend after ChatGPT in one study, and industry surveys, where half of European language companies reported falling 2024 revenue. In the US survey data, the share of translators under 30 fell from 18.5 percent (2018 to 2022) to 15.1 percent (2023 and 2024), just outside sampling noise in an exploratory test.

  • New work appeared alongside the old. Post-editing of machine output became 27 percent of European language-company revenue in 2024, and independent translators describe it as discounted. Whether it pays less per unit than the work it replaced has not been measured.

  • On one large platform, the shelter that creative translation had from neural MT did not hold against ChatGPT.

For other text work, the record suggests watching prices, freelance volume and the age of new entrants together with headcount.

Why translators

Three dated shocks hit the same occupation. Google Translate launched with statistical MT on 28 April 2006 (Google Research blog). Google announced neural MT in production on 27 September 2016 (Google Research blog), and DeepL launched in 2017. ChatGPT was released on 30 November 2022 (OpenAI). Each made a usable draft cheaper. Translators also sit at the top of current AI exposure rankings: a 2025 Microsoft study of Copilot conversations places interpreters and translators first among detailed occupations by its "AI applicability" score (Tomlinson et al., 2025, appendix Table L.4). This is the profession with the longest exposure history and among the highest present exposure.

Before reading any outcome data we wrote down hypotheses and what would make the history useless for predicting AI: a fall of more than 10 percent in employment or relative pay between 2022 and the latest year, or a large fall in written-translation demand with no offsetting growth elsewhere; no split between high-stakes and commodity work; or no sign that earlier waves created new tasks. That pre-registration is scored below.

Wave 1: statistical MT (2006 to 2015)

The best causal evidence for this period is Frey and Llanos-Paredes (2025). Using about 695 US local labour markets and the spread of the Google Translate app after 2010, they estimate that a 1 percentage point increase in Google Translate use reduced translator employment growth by 0.71 percentage points. Their counterfactual is about 28,000 translator positions not created over 2010 to 2023. In their data, total US employment grew about 13 percent over 2010 to 2023, translator and interpreter employment about 3 percent. Wages slowed after adoption and recovered from about 2016, with no significant effect over the whole period. Job postings asking for Spanish, Chinese and German also grew more slowly where MT use was higher.

That is displacement in the form of jobs that never appeared, not layoffs. Our own ACS series cannot test it: before 2018 the Census codes translators together with "miscellaneous media and communication workers", and that bundle grew from 49.6 per 100,000 workers in 2005 to 71.2 in 2016, with median earnings between about 0.78 and 0.90 of the all-worker median (our tabulation). All the bundle tells us is that the broad group did not shrink.

Machine translation also created demand. On eBay, introducing a new MT system raised exports on the platform by 10.9 percent (Brynjolfsson, Hui and Liu, 2019; the 2018 working paper reported 17.5 percent). Whether the extra cross-language commerce later created paid work for human translators is plausible but not shown by that study.

Wave 2: neural MT (2016 to 2022)

Neural MT was a large jump in quality, and the official count rose anyway. OEWS wage-and-salary employment went from 51,350 (May 2016) to 53,150, 57,140 and 58,870 (May 2017, 2018, 2019), then dropped to 52,170 by May 2021 (BLS OEWS). That drop straddles a change in the OEWS estimation method (May 2021) and the pandemic, which closed or moved online many of the schools, clinics and courts where interpreters work. We have not tested that explanation, and we do not attribute the drop to MT either.

The platform evidence is sharper. Yilmaz, Naumovska and Aggarwal (2026) study a large online labour market and find that Google's neural MT reduced demand for "analytical" translation tasks with no detectable effect on "intuitive" tasks such as transcreation, the creative adaptation of marketing copy. On that platform, routine text was repriced and creative work was sheltered.

In the ACS, the share of translators working in professional, scientific and technical services (NAICS 54, which includes translation agencies) rose from 34.5 percent in 2005 to 2007 to 44.1 percent in 2014 to 2016 and 48.5 percent in 2018 to 2019 (our tabulation). The pre-2018 figures are for the broader bundle, and NAICS 54 also holds self-employed translators who report their own business, so this cannot tell agency work from freelancing. Within the clean series from 2018, the further rise to 2023 to 2024 (+2.1 points, 95 percent interval -2.0 to +6.2) is within sampling noise.

Wave 3: LLMs (2023 onward)

What official statistics show: almost nothing yet

Interpreters and translators (SOC 27-3091), 2022 against the latest year:

  • OEWS wage-and-salary employment: 52,160 in May 2022, 52,060 in May 2025, -0.2 percent (BLS OEWS).

  • OEWS median wage relative to all occupations: 1.16 in May 2022, 1.18 in May 2025, +1.8 percent (BLS OEWS).

  • ACS employment including the self-employed: 94,567 in 2022, 97,547 in 2024, +3.2 percent, 95 percent interval -8.7 to +15.0 percent (our tabulation).

  • ACS employment per 100,000 workers: 60.1 (2021 to 2022 average), 57.9 (2023 to 2024 average), -3.6 percent, 95 percent interval -12.1 to +4.8 percent (our tabulation).

  • ACS median earnings relative to all workers: 0.76 in 2022, 0.76 in 2024, +0.5 percent (our tabulation).

ACS intervals use the Census replicate weights and treat survey years as independent.

Our pre-registered threshold for "the analogy breaks" was a fall of more than 10 percent in employment or relative pay. The official series do not cross it. But with two post-ChatGPT survey years, the ACS intervals still include a fall of about 9 to 12 percent, so the official data cannot yet tell a 10 percent fall from none.

The ACS count is about twice the OEWS count. Part of the gap is that the ACS includes the self-employed; the rest comes from design (households asked about a person's main job versus employers asked about filled positions). The ACS relative earnings figure is below 1 because it covers total annual earnings including part-time and part-year work; for those working 35 hours or more a week it is 0.80 to 0.91 over 2018 to 2024.

What other evidence shows: freelance demand and entry are moving

  • Platforms. Teutloff et al. (2025), studying an online freelance platform from January 2021 to September 2023, find that after ChatGPT demand for skills it can substitute, such as writing and translation, fell 20 to 50 percent relative to trend, while demand in complementary clusters grew. On Upwork, Hui, Reshef and Zhou (2024) find writing-related freelancers lost about 2 percent of monthly jobs and 5.2 percent of monthly earnings (figures from the authors' school summary; the published abstract gives no numbers). Demirci, Hannane and Zhu (2025) find a 21 percent drop in job posts for automation-prone writing and coding work within eight months, relative to manual-intensive jobs. These are platform populations, not the whole occupation.

  • The creative shelter weakens. In the same study that found neural MT spared transcreation, ChatGPT produced "substantially larger and more uniform reductions" across both routine and creative translation (Yilmaz, Naumovska and Aggarwal, 2026).

  • Industry. The European Language Industry Survey 2025 (1,322 responses, 654 independents) reports that MT is now used in more than half of professional translation work; that half of language companies saw 2024 revenue fall, by 19 percent on average; and that 23 percent of independents are considering ending their freelance activity. Post-editing rose 7 points to 27 percent of language-company revenue, but independents describe it as "discounted and less-valued". Respondents are self-selected.

  • UK surveys. In the Society of Authors survey of January 2024, as quoted by the European Writers' Council, 36 percent of translator respondents had already lost work to generative AI and 43 percent said the income from their work had decreased in value (787 responses across all member types; the number of translators among them was not reported). ITI's pulse surveys, as shown in search-index text of pages that now return 404, put the share with less work from existing clients at 46 percent (autumn 2023), 49 percent (spring 2024) and 51 percent (autumn 2024).

  • Entry, in the US data. In our ACS series the under-30 share of translators averaged 18.5 percent over 2018 to 2022 and 15.1 percent over 2023 and 2024 (about 930 to 990 sampled workers per year). The difference, -3.4 points, has a 95 percent interval of -6.0 to -0.7. This comparison was chosen after the annual figures had been seen and is one of 16 we report, with no adjustment for multiple comparisons, so treat it as a lead, not a finding. The self-employed share also fell (27.5 to 24.3 percent, 2018-19 to 2023-24), within sampling noise. The direction matches Brynjolfsson, Chandar and Chen (2026), who find employment of 22 to 25 year olds in AI-exposed occupations 19 percent below the path of less-exposed peers, driven by fewer hires.

Scoring the pre-registration

  • H1 (wage-and-salary employment kept growing 2006 to 2019; MT did not cause aggregate displacement): split. Growth is partly supported (OEWS 51,350 to 58,870 over 2016 to 2019; Frey and Llanos-Paredes report about +3 percent over 2010 to 2023). No-displacement is rejected: about 28,000 positions not created is displacement relative to the counterfactual. Whether interpreting carried more of the growth is untestable, because both share one code.

  • H2 (after 2022, displacement shows first in freelance and platform data rather than in OEWS): supported as far as the data go, with two caveats: the platform studies group translation with other writing, and the gap may reflect the lag and noise of official series.

  • H3 (high-stakes work holds better than commodity work; post-editing becomes a large, lower-paid share): partly supported. Creative translation held under neural MT but not under ChatGPT on one platform; legal, medical and live interpreting work was not measured; post-editing became a large revenue share, but "lower-paid" rests on respondents' descriptions.

  • F1 (official employment or relative wage down more than 10 percent, or written-translation demand down more than 20 percent within two years with no offsetting growth): not met. OEWS does not cross -10 percent. The platform clause is not established as written: the 20 to 50 percent fall covers writing and translation together, and post-editing revenue grew, which is the offset the clause rules out. The window is short, and the ACS interval still includes a fall near the threshold.

  • F2 (no tier split): not met for neural MT, where one platform study shows a split; for LLMs the split narrowed on the same platform.

  • F3 (no new tasks in earlier waves): not met. Post-editing is a measured revenue line in industry surveys.

What this suggests for other text work

The translator record gives five hypotheses, each with an indicator, for writers, editors, technical writers, support agents, analysts and anyone else whose output is text. They are things to watch, not forecasts the record establishes.

  1. Headcount may lag. In the Google Translate era, better MT slowed growth in official counts rather than cutting them, and since ChatGPT the platform measures have moved while official counts have not. Indicator: freelance demand and prices next to official employment.

  2. Entry may be the first margin to move. Indicator: the age mix and new hires, not the total. Among our comparators, editors show the largest fall in employment share: 88.8 per 100,000 workers in 2021 to 2022 versus 77.8 in 2023 to 2024 (-12.4 percent, 95 percent interval -17.9 to -6.8; our tabulation), although editors were already in long decline (121.3 in 2005 to 2007) and this comparison was also chosen after looking.

  3. New review work may appear. Post-editing became a large revenue line in translation, and the people doing it describe it as discounted. Indicator: whether "AI output review" becomes a separately priced task and what it pays per unit compared with writing.

  4. High-end work may get less shelter from general tools than from narrow ones. Neural MT spared creative translation on one platform and ChatGPT did not. Indicator: demand for the top tier of a text occupation, not only its average.

  5. Demand expansion is the counterweight. Cheaper translation grew cross-border trade on eBay; cheaper text may create new text needs. Whether that offsets the repricing for human workers is open, and it should show first in volumes, not headcount.

Where the analogy breaks

  • Translation has a floor that much text work lacks: in-person interpreting in hospitals, schools and courts, much of it tied to language-access requirements. We did not verify the legal details, and the ACS cannot separate interpreters from translators, so the size of that floor is unmeasured.

  • Interpreting is live and local; written translation is not. Liability, certification and procurement rules differ between a court, a hospital and an e-commerce listing.

  • Translation is judged against a source text. Much other text work has no source to check against, so the need for human review may behave differently.

  • LLMs can write the original text in the target language, which can remove the translation step altogether.

  • MT improved one task; LLMs improve many at once, so the move to adjacent tasks that translators made (review, terminology, localization management) is itself exposed.

  • The platform evidence covers freelancers on particular platforms, who are not representative of the occupation, and two survey years after ChatGPT are not enough for official series to settle anything.

Data and limits

The ACS series is our tabulation of public Census microdata (1-year PUMS, 2005 to 2024; 2020 is missing because no standard 1-year file was released). The tables, scripts and pre-registration are kept by Glasp Research and are available on request. Limits:

  • Coding. Before 2018 translators are coded together with miscellaneous media workers, and the bundle jumps 19 percent between 2016 and 2017, likely a classification change. Comparisons that cross 2018 use different groups.

  • Sampling. Each year has about 930 to 990 sampled translators. Replicate-weight standard errors cover employment and the under-30, self-employed and industry shares for 2018 to 2024, not medians. The interval tests were chosen after the annual tables had been read and are not corrected for multiple comparisons.

  • Earnings. ACS earnings are total annual earnings from all jobs in each year's dollars, not translation pay, hourly rates or per-word prices. Per-word prices, task volume, hires and post-editing work are not observed in either official series, and NAICS 54 is a weak proxy for agency work.

  • OEWS. Values for May 2016 to May 2023 were read from BLS pages through a fetch tool because BLS refuses scripted access; May 2025 came from the BLS public API; earlier years and May 2024 could not be retrieved. BLS does not encourage using OEWS as a time series, and it switched to a model-based estimation method (MB3) with the May 2021 estimates (BLS OEWS FAQ), so year-to-year OEWS figures here are indicative.

  • Surveys. The Society of Authors and ITI figures come from a third-party quotation and from search-index text, because the primary pages refused automated access or no longer exist. DeepL's launch is given by year only. Industry surveys are self-selected.

  • Coverage. UK, EU and Japanese series were not built; the non-US evidence is survey-based only, and the platform studies are global freelance markets, not US employment.

AI use

This article was researched and written by Claude (an AI model by Anthropic) in an automated session: it searched the public literature, downloaded public Census microdata, wrote the tabulation scripts and drafted the text. OpenAI's Codex reviewed it in three rounds, and Claude applied the agreed fixes. Numbers marked "our tabulation" were computed by those scripts; numbers from the literature were checked against the sources listed below where they were reachable, and the secondary ones are named as such in the text. No Glasp user data was used.

References

Comments

Add a comment
    Glasp Research

    Written by Glasp Research