Learning in the Age of Good-Enough Translation

AI translation reshapes the labor market, eroding low-skill roles while rewarding domain expertise
Education must shift toward “language plus” skills—pairing translation with data, law, or health
Policy should teach students to work with machines, not against them

Over the ten years following the emergence of neural machine translation from research labs to everyday use on our devices, a quiet transformation occurred in the hiring of translators. A recent study examining the spread of machine translation across various sectors suggests that for every one-percentage-point rise in usage, translator job growth slowed down by approximately 0.7 percentage points—leading to an estimated 28,000 lost new translator jobs between 2010 and 2023 compared to what would have happened otherwise. The same research indicates significant drops in job postings requiring foreign language proficiency, particularly in industries most affected by automated translation, even for widely spoken language pairs like English and Spanish. These figures are not indicative of a catastrophic outcome; instead, they represent consistent, incremental changes that alter the profession's trajectory. If we continue to teach languages and translation as though no significant changes have occurred, we are preparing students for a job market that is rapidly evolving. Conversely, if we recognize “good-enough” machine translation as a foundational tool and educate individuals to build upon it, we can create new opportunities on a more stable foundation.

We are reframing the familiar “translation jobs are dead” lament into a policy argument about complements. When the baseline cost of getting a passable translation falls toward zero, value migrates to what machines still do badly or cannot be trusted to do alone. That includes domain-specific precision, liability-bearing use cases, quality assurance at scale, and multilingual data stewardship. The practical implication for education is not to abandon language learning or to double down on a nostalgic defense of artisanal translation. It is to couple language with domain expertise and with the tooling, methods, and judgment needed to supervise automated systems—what we might call “language operations.” The goal is neither to resist automation nor to yield to it, but to teach students how to place the machine in the loop and keep themselves in charge of outcomes.

The timing matters. In the last two years, research groups and vendors have demoed real-time, expressive speech-to-speech systems that preserve tone, and large language models increasingly match or surpass dedicated MT systems on specific WMT evaluations. At the same time, human translators report material income losses in parts of the market most exposed to automation. The pattern is classic technological diffusion: rapid capability gains at the frontier, uneven enterprise integration, and sharp price pressure in commoditized segments long before “replacement” is complete. Educators who wait for the dust to settle will find their graduates already competing on the wrong margin.

Shifting the Focus from 'Translation' to 'Language Operations'

Across the commercial language industry, the work has already been unbundled and re-bundled, showcasing the industry's adaptability. Post-edited machine translation (MTPE) is now the default production baseline for a large share of vendors, with industry surveys showing adoption rates rising steeply from 2022 to 2024. Shared tasks at WMT have folded LLM systems into head-to-head evaluations with traditional MT providers, while research programs like Meta’s Seamless line push low-latency voice translation into everyday tools. In other words, production workflows now assume that a machine produces the first draft and that humans specialize in triage: deciding when the draft is safe, what to escalate, and how to achieve publication-grade accuracy in contexts where errors carry legal, financial, or clinical risk. The pedagogy that matches this world does not start with replacing bilingual drafting; it begins with training students to measure quality, select and tune systems, maintain domain glossaries, and document the line between acceptable and unacceptable risk.

This reconfiguration is evident in labor data that, at first glance, appears contradictory. The U.S. Bureau of Labor Statistics reports a 2024 median annual wage of $59,440 for interpreters and translators, with employment projected to grow 2% from 2024 to 2034—slower than average but not collapsing. Meanwhile, industry and union surveys in publishing report that over a third of translators have already lost work or income due to the use of generative AI. Both can be true when aggregate employment is propped up by growth in specialized and public-sector roles (e.g., courts, hospitals, government), even as prices in commoditized segments compress. That is precisely why education policy should pivot away from preparing generalists for one-off document work and toward roles that attach language competence to regulated domains, enterprise systems, and accountability frameworks.

What the data say—and what they miss

We can quantify aspects of the shift while acknowledging uncertainties, particularly regarding price signals. Freelance translator rates hover around $20/hour or $0.10–$0.12/word, but crowd-sourced and vendor sites also show $0.06/word tiers, with MTPE priced at 50–70% of human rates. Using conservative throughput estimates—400–600 words/hour for human translation and 800–1,000 for light post-editing—implied hourly revenues can fall below a living wage in commodity segments. At the same time, premium rates exist in legal, medical, and technical fields. These estimates, derived from industry guidance and empirical studies, reveal critical differences by language pair, client type, and risk profile, highlighting that automation drops prices where quality tolerance is high and liability is low.

A broader education system question sits upstream of wages: Are students still choosing to study languages at all? In the United States, higher-education enrollments in languages other than English fell 16.6% between 2016 and 2021, and roughly 29% since 2009’s peak. In Europe, the picture differs: in 2023, 60% of students in upper-secondary general education studied two or more foreign languages. However, primary-level multi-language study remains rare outside a few countries. If we teach as if every graduate will work in cross-border teams, Europe’s data justify continued investment. If we teach as if automated translation removes the need for language learning, America’s enrollment declines suggest students already believe that story. Policy must rebuild the case for language study on the ground where value now accrues: language plus something, with explicit attention to supervising AI.

Teach to the complement: a compact for schools, universities, and employers

Designing curricula in a world with free and instant machine translation (MT) requires a focus on key skills. First and foremost, quality evaluation must be taught as a core competency. Students should conduct human assessments, interpret automated metrics with caution, and utilize error analysis to evaluate systems—skills that reflect current research in MT meta-evaluation. Literacy should encompass the ability to judge the appropriateness of translation models for specific applications, reinforcing the crucial role of human judgment in the translation process.

Second, post-editing should be formalized as a professional skill rather than a fallback option. Evidence suggests post-editing can be faster than starting from scratch, but quality can vary significantly based on approach and familiarity with content. Structured training in post-editing helps students avoid monotonous tasks and enhances their decision-making capabilities.

Third, early integration of domain specialization—such as Language + Law or Language + Biomedicine—can align education with job market demands, emphasizing terminology management and regulatory understanding. Capstone projects should involve real-world partnerships for practical experience.

Fourth, students must learn data stewardship, which includes ethical curation of parallel data and understanding legal considerations in translations. These skills are increasingly relevant to emerging roles in localization and multilingual content management.

For K-12 and general education, machine translation should be seen as a tool rather than a threat to language classes. Students should learn to critically engage with machine translation to verify information and enhance their media literacy.

Employers must clearly define when expert human translation is necessary versus when post-editing or machine-only processes are acceptable. Transparent standards can optimize language work and help educators prepare students for realistic negotiation around language service quality.

Critics may argue that advancements in AI translation will render language roles obsolete; however, evidence suggests that adoption often lags behind capability. Additionally, as translation systems evolve, the need for judgment and error detection will remain crucial, ensuring durable opportunities in the field.

This perspective is supported by data on translator employment trends, labor statistics, enrollment figures, and industry practices, each with its limitations, yet collectively outlining the landscape of evolving language needs.

The 28,000 “missing” translator jobs are a blunt measure of a subtler transformation. We do not live in a world that has abolished translation; we live in a world where mediocre translation is abundant. That abundance drains value from generalist workflows and restores it to the edges: to the legal contract whose ambiguity matters, to the clinical discharge summary that cannot be wrong, to the cross-border dataset whose labels must be audited, and to the editorial line where voice and meaning, not just words, carry the freight. Education policy should meet the world as it is. Keep languages, but teach them with systems. Build majors and minors that yoke language to law, finance, health, and energy. Make post-editing, evaluation, and data stewardship as familiar as verb conjugations. Public institutions that deploy MT should publish human-oversight thresholds and train their staff to meet them. If we do, we stop arguing about whether translation is dead and start preparing graduates for the work that only they can do: the work of deciding what counts as accurate, safe, and fit-for-purpose in every language we use.

The views expressed in this article are those of the author(s) and do not necessarily reflect the official position of the Swiss Institute of Artificial Intelligence (SIAI) or its affiliates.

References

Acemoglu, D. (2024). Interview in NPR Planet Money newsletter (media clip). Massachusetts Institute of Technology.
Bureau of Labor Statistics (U.S.). (2025). Interpreters and Translators (OOH updated 2024 wage/outlook).
European Commission (Eurostat). (2025). Upper secondary education: 60% study 2 or more languages (2023 data).
Lightcast / OECD. (2024). Artificial intelligence and the changing demand for skills in the labour market. OECD Publishing.
Meta AI. (2023). Seamless: Multilingual, expressive and streaming speech translation.
Modern Language Association. (2023). Enrollments in Languages Other Than English in U.S. Higher Education, Fall 2021.
Nimdzi. (2025). The MTPE Efficiency Gap (industry survey).
Society of Authors (UK). (2024). AI survey: impacts on translators.
Upwork. (2025). Hire Translators (median rates).
WMT (Conference on Machine Translation). (2024). General MT task overview and results. Association for Computational Linguistics.
X. Peng (2024). The Impacts of MT quality on post-editing effort (SAGE Open).
J. Algaraady et al. (2025). ChatGPT’s potential for augmenting post-editing (PMC review).