Interpreters and translators

Three parts of your work where AI is already doing real lifting, and what stays yours.

• Pen1815–1945

  Pen, dictionary, and consecutive note-taking (pre-simultaneous era)

  For the first century of professional language work, the interpreter's only tools were a notebook, pencil, and the reference dictionaries available in the working languages. Diplomatic and legal interpreters worked consecutively -- listening to a complete statement, taking shorthand notes using personal symbol systems, then rendering the speech in the target language while the original speaker waited. Literary and commercial translators worked with paper manuscripts, bilingual dictionaries, and personal glossaries. No technology mediated the language transfer itself; everything depended on the practitioner's bilingual fluency and memory.

• IBM simultaneous interpretation booth s…1945–1970

  IBM simultaneous interpretation booth system (Nuremberg 1945, UN standard 1946)

  The system IBM installed for the Nuremberg trials in October 1945 -- microphones, headphones, switching consoles, and glass-fronted booths delivering four simultaneous language channels -- redefined what professional interpretation meant. Consecutive interpretation had previously required the original speaker to pause for the interpreter; simultaneous interpretation allowed real-time multilingual communication at full speed. The trial would have taken at least four times as long without it. The UN General Assembly adopted simultaneous interpretation as its standard working method in 1946, and every international body that followed -- NATO, the OECD, the European Coal and Steel Community -- replicated the booth architecture. For the first time, a physical technology shaped the practitioner's work environment: the booth defined the shift structure (two interpreters per language pair, switching every half hour or so), the required cognitive skills (sustained split-attention at speed), and the ergonomics of the job.

• Early machine translation experiments1954–1990

  Early machine translation experiments (Georgetown-IBM 1954, ALPAC 1966, rule-based MT)

  On January 7, 1954, Georgetown University and IBM demonstrated the first public machine translation system in New York -- sixty Russian sentences rendered into English by a computer using six grammar rules and hundreds of vocabulary items. The demonstration made front-page news and sparked a decade of government-funded MT research. The US government commissioned the ALPAC report in 1964; it delivered its verdict in 1966: machine translation was slower, less accurate, and more expensive than human translation, and large-scale automated translation was not achievable in the near term. Funding dried up for a decade. The MT programs that survived through the 1970s and 1980s were narrow, rule-based systems (SYSTRAN, first used by the European Commission in 1976) that were useful for technical documentation in controlled language but could not handle unrestricted text at professional quality. The effect on practitioner employment was negligible: translators of the period considered MT a curiosity rather than a competitive threat.

• CAT tools1978–2000

  CAT tools and translation memory (TRADOS 1984, court certification programs 1978)

  Two developments in the late 1970s and 1980s reshaped the profession from different directions. The Court Interpreters Act of 1978, signed by President Carter on October 29, 1978, required certified human interpreters in federal proceedings -- establishing for the first time a legal mandate for professional qualification in US courts, which California, New York, New Mexico, and New Jersey quickly replicated at the state level. In the same period, Reinhard Schaler and Jochen Hummel founded Trados GmbH in Stuttgart in 1984, producing the first commercial computer-assisted translation (CAT) tool and translation memory system. CAT tools did not automate translation; they remembered previously approved translations and suggested them when a matching segment recurred, improving consistency and throughput on repetitive technical documents.

  Effect on the work

  CAT tools increased translator throughput on technical and software content, allowing the same workforce to handle greater volume as the PC software industry created massive demand for localized documentation through the 1990s. They did not reduce employment but they did create pressure on per-word rates, since buyers argued that leveraged (repeated) segments deserved lower rates than new content.

• Statistical machine translation2000–2016

  Statistical machine translation (Google Translate launch 2006, SMT era)

  Google Translate launched in April 2006, initially using a statistical machine translation engine trained on UN documents and the web. By 2010 it covered many language pairs. For most practitioners, the SMT era was an expansion of scope more than a threat: the output was obviously imperfect for professional use, buyers understood the quality gap, and the profession continued growing. The more consequential development was the expansion of remote interpretation via phone and video (video relay services for ASL interpretation grew rapidly through the 2000s, and telephone interpretation services like Language Line expanded into healthcare). For the general written translation market, SMT created downward rate pressure as buyers experimented with MT-then-post-edit workflows, but the quality gap remained large enough that most professional-quality assignments still required human translation from scratch.

• Neural machine translation -- Google NMT2016–2022

  Neural machine translation -- Google NMT (2016) and DeepL (2017)

  The shift from statistical to neural machine translation in 2016 was a genuine quality discontinuity. Google Translate switched to a neural architecture in September 2016 and reported an immediate reduction in translation errors on tested language pairs. DeepL launched its own neural MT engine on August 28, 2017, and quickly built a reputation for idiomatic quality that outperformed Google Translate on European language pairs. A 2025 study estimated that Google Translate alone had displaced tens of thousands of translator positions in the US over the period from 2010 to 2023. The certified segments -- court and medical interpretation, simultaneous conference work -- were structurally protected: you cannot post-edit a consecutive interpretation in a courtroom.

  Effect on the work

  Neural MT created the sharpest rate compression in the history of the profession. Freelance translators in commodity language pairs (Spanish-English, French-English, German-English) saw per-word rates collapse. Post-editing MT output became the dominant workflow for high-volume general content, paying less than full translation. The wage-and-salary employment series held steady because the institutional (court, medical, conference) workforce continued growing; the freelance written translation market absorbed the compression.

• Remote simultaneous interpretation plat…2019–2023

  Remote simultaneous interpretation platforms -- KUDO (2019), Interprefy AI assist

  KUDO launched in 2019 as a browser-based remote simultaneous interpretation platform, enabling conference interpreters to deliver real-time multilingual interpretation from home without physical booths. Interprefy followed with its own RSI platform and added AI-powered speech recognition as an assistive feature for interpreters. The COVID-19 pandemic from March 2020 accelerated adoption dramatically: international organizations, corporations, and governments that had relied on on-site booths switched to RSI virtually overnight. For conference interpreters, remote work was a genuine opportunity -- a Tokyo-based interpreter could service a Brussels meeting -- but also introduced new challenges: audio quality issues, screen fatigue, and reduced relay capability. The AI speech-to-text features in both platforms served as cognitive aids, not replacements: interpreters used real-time transcripts to catch missed phrases but still provided the full interpretation themselves.

• Generative AI translation -- GPT-42022–present

  Generative AI translation -- GPT-4, Claude, ElevenLabs dubbing (2022-present)

  GPT-4 and Claude handle translation at a level that broadly matches junior translator quality across major language pairs, and in some register-sensitive tasks approaches senior human performance. For general written translation, this is the most disruptive technology since writing itself: most of the demand for commodity written translation -- marketing copy, internal communications, basic documentation -- can now be met at near-zero marginal cost. ElevenLabs Multilingual v2 (2023) and HeyGen (2023) added AI video dubbing with voice cloning and lip-sync, collapsing the cost of multilingual video production. The profession has split into two distinct labor markets under the same SOC code: a general-content segment in structural decline, and a certified-institutional segment that retains its human mandate.

  Effect on the work

  The institutional (court, medical, conference) segment continued adding jobs. The net BLS OEWS effect is near-flat employment (53,360 in 2024) masking a genuine bimodal divergence.