How Machine Translation and AI Interpreting Work, and What's Left for Translators
Neural and LLM translation, quality estimation, post-editing and speech-to-speech interpreting, explained step by step, with the eight core tasks of interpreters and translators graded by evidence and what US job data does and doesn't show.
Checked against ISO 18587:2017 (edition 1), Microsoft's Interpreter agent documentation (updated October 8, 2026) and Google Cloud Translation's price list as published that day
Role scorecard
Interpreters and Translators SOC 27-3091.00
- Core tasks AI does in production1 of 8
- Most mature evidenceIn production
- Cost$20 per million characters (Google NMT list price)
- Benefit36% higher productivity post-editing NMT (peer-reviewed, 6 translators)
As of · Tasks: O*NET 31.01
| Core task (O*NET) | What AI or robots do | Extent | Evidence grade | Example case | Source |
|---|---|---|---|---|---|
| Follow ethical codes that protect the confidentiality of information. | Software doesn't follow a code of ethics; vendors describe how their systems handle data instead | Not yet | Vendor claim | Microsoft states that Teams Interpreter never stores voice samples or biometric data, 2026 | 2 |
| Translate messages simultaneously or consecutively into specified languages, orally or by using hand signs, maintaining message content, context, and style as much as possible. | Speech-to-speech translation devices interpret face-to-face conversations; no sign-language case found | Does it | In production | AI translation devices at the "Beijing Service" counter, Beijing Daxing International Airport, 2025 | 3 |
| Listen to speakers' statements to determine meanings and to prepare translations, using electronic listening systems as necessary. | Speech recognition turns the speaker's words into text for the translation step, a step inside the device's interpreting rather than the task as a whole | Assists | In production | Same Daxing counter devices, which take face-to-face voice input, 2025 | 3 |
| Compile terminology and information to be used in translations, including technical terms such as those for legal or medical material. | Term-extraction tools and AI help terminologists prepare termbase entries | Assists | In production | European Commission DG Translation terminologists preparing IATE data (6.5 million terms, 2025) | 4 |
| Refer to reference materials, such as dictionaries, lexicons, encyclopedias, and computerized terminology banks, as needed to ensure translation accuracy. | Translation memories retrieve earlier translations; machine translation output is offered as an optional resource | Assists | In production | DG Translation translators with the Euramis database and eTranslation | 4 |
| Check translations of technical terms and terminology to ensure that they are accurate and remain consistent throughout translation revisions. | Machine translation applies a user's glossary so set terms come out the same way; the checking stays with a person | Assists | In production | eTranslation's glossary upload, run by the European Commission | 5 |
| Identify and resolve conflicts related to the meanings of words, concepts, practices, or behaviors. | Large language models tested on picking the right sense of ambiguous words; cultural mediation not shown | Not yet | Prototype | University of Edinburgh study of LLM disambiguation, WMT 2023 | 6 |
| Compile information on content and context of information to be translated and on intended audience. | Translation models can be steered with example translations a person supplies; no case of AI compiling the brief itself | Not yet | Vendor claim | Google Cloud Translation's adaptive translation, which tailors output to supplied examples | 7 |
Extent and grade follow our evidence grades: "Does it" needs a production case where AI performs the task on its own; "Assists" means AI helps a person do it; "Not yet" means only prototypes, vendor claims or nothing. Rows are O*NET's eight core tasks for 27-3091.00; written translation of documents ("Read written materials … and rewrite material into specified languages") is a supplemental task in O*NET, so it isn't a row, though the narrative covers it. The one "Does it" row rests on a single deployment, whose device maker the airport doesn't name; listening is marked "Assists" because speech recognition is one step inside that interpreting. Generally available products from Microsoft and Google are described below but not graded as deployments, since no operator confirmation was found. Sign-language interpreting by AI wasn't found in production.
Costs and benefits
| Neural machine translation, list price | $20.00 per million characters above the first 500,000 a month (Google Cloud Translation, vendor list price)8 |
|---|---|
| LLM translation, list price | $10 per million input characters and $10 per million output characters (Google Translation LLM, vendor list price)8 |
| AI speech interpreting in meetings | 20 hours per user per month included with a Microsoft Copilot license (Microsoft, vendor-stated)2 |
| Human interpreters and translators | Median annual wage $60,170, May 2025 (BLS); per-word or per-hour rates: not disclosed in a reachable primary source9 |
| Post-editing productivity | 36% higher with neural MT and 18% with statistical MT than translating from scratch; keystrokes down 23% and 9% (peer-reviewed experiment, 6 professional literary translators, English to Catalan, authors declare no commercial conflict of interest)10 |
How it works
1. Neural machine translation: encode, then decode
Machine translation (MT) is software that turns text in one language into another. Today's systems are neural networks trained on millions of sentence pairs that people have already translated. The standard design is the encoder–decoder: one part of the network reads the source sentence and turns it into numbers that capture its meaning in context, and the other part generates the target sentence word by word from those numbers. In 2017 Google researchers introduced the Transformer, an encoder–decoder built entirely on "attention", a mechanism that lets each word weigh every other word in the sentence. On the WMT 2014 English-to-German test it scored 28.4 BLEU, more than 2 points above the previous best results, and it trained in a fraction of the time.11 BLEU is an automatic score of how closely machine output matches human reference translations.
2. LLM translation: the same idea, a bigger model
A large language model (LLM) is a Transformer trained on vast amounts of general text, which can also translate when asked. Vendors now sell both kinds side by side. Google Cloud Translation, for example, offers its neural MT (NMT) model at $20.00 per million characters above a free 500,000 a month, and a "Translation LLM" at $10 per million input characters plus $10 per million output characters, which Google describes as "cost equivalent with NMT" (vendor list prices, retrieved October 13, 2026).8 LLMs can also be steered with examples: in Google's adaptive translation, the customer supplies example sentence pairs and the service picks the five most similar ones to tailor each translation.7 The examples still come from a person, which is why the scorecard rates "compile information on … intended audience" as not yet done by AI.
3. Quality estimation: scoring a translation with no answer key
Because MT output varies from fluent to wrong, systems try to predict their own quality. Quality estimation means scoring a translation automatically "at run-time, without relying on reference translations", in the words of the research community's shared task at the WMT 2024 conference. The task scores whole sentences and single words, and in 2024 added a subtask in which systems must explain the errors they detect and suggest corrections.12 In a translation workflow, a score like this can be used to decide which segments a human needs to look at first.
4. Post-editing: where the translator comes back in
Post-editing is a person correcting machine output instead of translating from scratch. It has its own international standard, ISO 18587:2017 (edition 1), which "provides requirements for the process of full, human post-editing of machine translation output and post-editors' competences" and applies only to content processed by MT systems. ISO lists it as due for revision; the draft that will replace it, ISO/DIS 18587, is titled "Post-editing of non-human translation output", a wording that no longer assumes classic MT.13 How much time post-editing saves depends on the text and the engine. In one peer-reviewed experiment, six professional literary translators working on a fantasy-novel chapter from English into Catalan were 36% more productive when post-editing neural MT, and 18% more productive with older statistical MT, than when translating from scratch.10 That is one language pair and one genre, so it isn't a general rate.
5. AI interpreting: speech in, speech out
Spoken interpreting chains three steps. Microsoft documents its Teams Interpreter agent this way: speech recognition converts the spoken language into English text, machine translation translates that text into the selected languages, and text-to-speech produces the translated audio, either in a preset voice or in a simulation of the speaker's own voice. It supports nine languages, and Microsoft states that voice samples and biometric data are never stored.2 Google made speech translation in Meet generally available to Workspace business customers on January 27, 2026, at launch for one language pair per meeting, between English and Spanish, French, German, Portuguese or Italian, with the translation dubbed over the original speech.14 Both are vendor products; the evidence grade for a deployment needs an operator. One operator-confirmed case is Beijing Daxing International Airport, which in 2025 introduced AI translation devices at its "Beijing Service" counter for face-to-face voice translation. The same airport also expanded a hotline on which translation students from Beijing Foreign Studies University interpret between passengers and staff in eight languages.3
6. A production workflow at scale
The European Commission's Directorate-General for Translation shows how the pieces fit in one large operation. Its translators have used a computer-assisted translation (CAT) tool with shared translation memories since 1997, retrieving earlier translations of similar passages, which also keeps terminology consistent. The Commission's own neural MT service, eTranslation, is trained on that translation data; the directorate says its translators "can also use machine translation results as an optional resource". eTranslation delivered 890 million pages in 2025, up from 19 million in 2017, while the directorate itself produced 2,499,000 pages in 2025, 36% of them outsourced. Its terminologists use "terminology extraction tools, termbases and AI" to prepare entries for IATE, the EU terminology database.4 eTranslation also lets users upload their own glossary to control how terms are translated.5
7. What the US job data shows
The BLS counts wage-and-salary jobs, not the self-employed. Its estimates for interpreters and translators were 58,870 in May 2019 (median wage $51,830)15, 56,920 in May 202016, 52,170 in May 202117, 52,160 in May 202218, 51,560 in May 2023 (median $57,090)19 and 52,060 in May 202520, with a May 2025 median of $60,170 (wages are nominal).9 The series doesn't show why employment fell between 2019 and 2021, a span that includes the COVID-19 pandemic, so it can't be read as an AI effect. Counting the self-employed, the BLS Occupational Outlook Handbook puts the occupation at about 73,900 jobs in 2025 and projects 2% growth from 2025 to 2035, with about 6,000 openings a year. It says AI and other computer tools "are making the work of translators and localization specialists more efficient", but that many jobs "cannot be entirely automated" because "computers cannot yet produce work comparable to what human translators do in most cases."9
8. What exposure and survey studies measure
Three studies are often cited, and each measures something different. An Oxford Martin School working paper (March 2025) compared 695 US local labor markets after Google Translate's app launch: a 1 percentage point rise in Google Translate use went with 0.71 percentage points slower growth in translator employment, which the authors translate into about 28,000 fewer translator positions created in 2010–2023 than in a scenario without machine translation, with no statistically significant effect on wages over the full period. It studies pre-LLM machine translation.21 A Microsoft Research paper (version of December 22, 2025; Microsoft sells the AI it studied) analyzed 200,000 Bing Copilot conversations and gave interpreters and translators the highest "AI applicability" score of any occupation, 0.492, with a coverage of 0.980, meaning that work activities making up 98% of the occupation's importance-weighted activities each showed up in at least 0.05% of those conversations. That is a measure of overlap, not job loss. The authors write that it is "tempting to conclude" that occupations scoring high on what the AI itself does "will be automated and thus experience job or wage loss", and add: "This would be a mistake, as downstream consequences of new technologies are very hard to predict and often counterintuitive."22 Finally, a January 2024 Society of Authors survey in the UK (787 responses from members and other authors who chose to take part, across all types of author) found 36% of translators said they had already lost work due to generative AI, and 43% said their income had decreased in value.23
Sources
- O*NET OnLine (O*NET 31.0 data), 27-3091.00 Interpreters and Translators: Tasks, retrieved 2026-10-13.
- Microsoft Learn, Manage Interpreter agent for your organization, updated October 8, 2026.
- Beijing Municipal Government (english.beijing.gov.cn, from Beijing Daily App), Travel Easier with Daxing Airport Expanding Language Support, November 22, 2025.
- European Commission, Directorate-General for Translation, Translation at the European Commission, updated July 24, 2026.
- European Commission, Directorate-General for Translation, Resources for language professionals, updated July 7, 2026.
- University of Edinburgh researchers, Towards Effective Disambiguation for Machine Translation with Large Language Models (WMT 2023), arXiv, September 20, 2023.
- Google Cloud, Translate text by using adaptive translation, updated October 9, 2026.
- Google Cloud, Cloud Translation pricing, retrieved 2026-10-13.
- U.S. Bureau of Labor Statistics, Occupational Outlook Handbook: Interpreters and Translators, last modified August 27, 2026.
- Frontiers in Digital Humanities (University of Groningen and Dublin City University researchers), Post-editing Effort of a Novel With Statistical and Neural Machine Translation, May 15, 2018.
- Google researchers, Attention Is All You Need, arXiv, June 12, 2017.
- Conference on Machine Translation (WMT 2024), Shared Task: Quality Estimation, retrieved 2026-10-13.
- ISO, ISO 18587:2017 Translation services — Post-editing of machine translation output — Requirements, retrieved 2026-10-13.
- Google Workspace Updates, Speech translation in Google Meet now generally available for businesses, February 4, 2026.
- U.S. Bureau of Labor Statistics, Occupational Employment and Wages, May 2019: 27-3091 Interpreters and Translators, retrieved 2026-10-13.
- U.S. Bureau of Labor Statistics, Occupational Employment and Wages, May 2020: 27-3091 Interpreters and Translators, retrieved 2026-10-13.
- U.S. Bureau of Labor Statistics, Occupational Employment and Wages, May 2021: 27-3091 Interpreters and Translators, retrieved 2026-10-13.
- U.S. Bureau of Labor Statistics, Occupational Employment and Wages, May 2022: 27-3091 Interpreters and Translators, retrieved 2026-10-13.
- U.S. Bureau of Labor Statistics, Occupational Employment and Wages, May 2023: 27-3091 Interpreters and Translators, retrieved 2026-10-13.
- U.S. Bureau of Labor Statistics, OEWS series OEUN000000000000027309101 (national employment, 27-3091), May 2025, retrieved 2026-10-13.
- Oxford Martin School, University of Oxford, Lost in Translation: Artificial Intelligence and the Demand for Foreign Language Skills (working paper), March 7, 2025.
- Microsoft Research, Working with AI: Measuring the Applicability of Generative AI to Occupations (v6), arXiv, December 22, 2025.
- European Writers' Council, SoA Survey UK: A third of translators and quarter of illustrators losing work to AI (reporting the Society of Authors' survey), May 9, 2024.