Will AI Replace Translators?

What the employment figures say

The US Bureau of Labor Statistics counts 75,300 people in Interpreters and Translators, SOC 27-3091, for 2024, and projects 2% employment growth through 2034 with roughly 6,900 openings a year, attributing none of it to AI ^{(US Bureau of Labor Statistics 2024)}. That is not a contradiction of a score of 26, because the two figures measure different things. The official category bundles oral interpreting and sign-language work with written translation, so the segment under most pressure is diluted by segments that need a person in the room, and the agency's own OEWS series for the same code reports far lower employment — 51,560 in May 2023 — on a basis that should not be spliced with the Handbook number ^{(US Bureau of Labor Statistics 2023)}. The task exposure is real and the official headcount is holding, largely because the category is broader than the job.

What the research says about translators

Now

Two measurements sit directly on top of each other. Interpreters and Translators record the highest AI applicability of any occupation in the Microsoft data, at 0.492 ^{(Microsoft Research 2025)}. Official US employment for the same code barely moved across the most recent comparable pair of years, from 52,160 in May 2022 to 51,560 in May 2023, while the median wage rose from $53,640 to $57,090 ^{(US Bureau of Labor Statistics 2023)}. The European base is small to begin with: Eurostat counted about 180,000 translators and interpreters across the EU in 2018, some 0.1% of total EU employment ^{(Eurostat 2019)}. The sharpest present-day signal is not in the statistics at all but in practitioner accounts, where an Italian-to-English translator in Rome went from 50-60 hours a week to essentially no work requests in June 2025, and another estimated demand had fallen more than 70% since 2023 as companies routed websites, contracts, blogs and internal documents through AI ^{(Blood in the Machine 2025)}.

Buyer-side accounts point the same way, though they come from parties with something to sell. Smartling, which sells translation-management software and AI translation, reports customer results including a 60% cut in translation costs at Therabody and a Fortune 500 company processing more than 50 million words annually with $3.4 million saved in year one and quality scores above 99 — vendor case studies rather than independent audits, and none of them report translator layoffs ^{(Smartling 2025)}. A 2025 case study of Netflix localization, published by an industry institute with its own commercial stake in the field, describes neural machine translation across dozens of languages alongside synthesized voices and automated subtitling, with professional linguists reviewing the raw output as translators shift toward editor and quality-controller roles ^{(The Localization Institute 2025)}.

The next few years

Employer intent is running ahead of measured employment. Duolingo said in April 2025 that it would gradually stop using contractors for work AI could handle and expect employees to focus on creative, non-repetitive work, disclosing no realized headcount reduction ^{(The Verge 2025)}. A 2025 survey of translators by Acolad — a company that sells the managed language services and AI products under discussion, published without a stated sample size — found 42% using AI tools daily, 59% using neural machine translation, 53% seriously concerned about the professional impact and 84% expecting lower demand for human translation alongside higher demand for post-editing ^{(Acolad 2025)}. Canada's Job Bank, with no product to sell, rates several regional outlooks Moderate for 2025-2027 and states that advanced translation software may moderate demand for some services over the long term ^{(Government of Canada Job Bank 2027)}.

Longer term

The official long-range view is undramatic: 2% projected growth for Interpreters and Translators from 2024 to 2034, about 6,900 openings a year, and no AI attribution from the agency ^{(US Bureau of Labor Statistics 2024)}. Against that sits an explicit regulatory floor. The HHS Section 1557 rule published on May 6, 2024 defines machine translation as automated text-based translation without qualified human review, and requires review by a qualified human translator when accuracy is essential, when the source is complex, non-literal or technical, or when the text is critical to a limited-English-proficient person's rights, benefits or meaningful access ^{(US Department of Health and Human Services 2024)}. The shape that follows is a shrinking pool of general document work and a durable, credential-gated remainder.

What is exposed, and what protects it

● Where the exposure is

The exposed portion is the larger portion. Interpreters and Translators post the highest AI applicability score of any occupation in the Microsoft measurement, at 0.492 ^{(Microsoft Research 2025)}. Straight document work — commercial websites, contracts, blogs, internal material — is where buyers have already switched, and at least one working translator has put the fall in demand since 2023 at more than 70% ^{(Blood in the Machine 2025)}. Volume buyers describe the economics behind that switch, with one vendor's customer case studies claiming 60% cost reductions and 50% faster delivery, although the vendor sells the software and the numbers are not independently audited ^{(Smartling 2025)}. Media localization shows the same pattern at scale, with machine translation across dozens of languages and human linguists moved into review and quality control ^{(The Localization Institute 2025)}.

○ What holds

Regulation is the firmest moat here, and it is written down. Under the HHS Section 1557 rule, machine translation without qualified human review is not sufficient where accuracy is essential, where material is complex, non-literal or technical, or where the text governs a limited-English-proficient person's rights and access to benefits ^{(US Department of Health and Human Services 2024)}. The American Translators Association, which represents the profession and has a representational interest in defending the value of expert work, warns that output which reads as accurate can carry errors a non-expert cannot detect, and flags confidentiality alongside healthcare, legal, financial and public-sector risk ^{(American Translators Association 2025)}. Peer-reviewed work supports part of that case: a 2025 comparison of professional post-editing with ChatGPT-4o on Arabic translations found the human post-editors ahead on accuracy, terminology, consistency, coherence, grammar and cultural appropriateness ^{(Frontiers in Artificial Intelligence 2025)}. Quality also compounds downstream, since with the weakest machine system in one controlled study roughly one sentence in four was post-edited below the acceptable quality threshold, and worse raw output produced worse finished work ^{(Marina Sanchez-Torron and Philipp Koehn 2016)}.

Which tasks go first

First-pass translation of websites, contracts, blogs and internal documents is where substitution has already occurred ^{(Blood in the Machine 2025)}. The exposure indices behind this score originally asked whether a model could speed up a task with a person prompting and reviewing each step; agentic systems that plan, use tools and run multi-step work unattended change the question to whether the work can be completed without you ^{(Nous Research 2026)}. For drafting, the answer is largely yes, with writing and proofreading among the highest-applicability occupations and final editorial judgment left to a human ^{(Microsoft Research 2025)}. What remains human is the work with a name attached to it: certified and high-risk material in healthcare, legal, financial and public-sector settings, where fluent output can still hide errors a lay reader will not catch ^{(American Translators Association 2025)}. Post-editing is the other survivor, and trained post-editors still outperform the model on terminology, consistency, coherence and cultural fit ^{(Frontiers in Artificial Intelligence 2025)}. That work is paid for the hours it takes — measured at roughly 716 to 887 words an hour in one controlled study — and it is not the same job as translating from scratch ^{(Marina Sanchez-Torron and Philipp Koehn 2016)}.

What to do about it

Move toward work that legally requires a human

Certified, sworn and regulated-sector translation is the segment with a written requirement behind it, since the Section 1557 rule obliges qualified human review where accuracy is essential or the source material is complex, non-literal or technical ^{(US Department of Health and Human Services 2024)}. Healthcare, legal, financial and public-sector content is where the trade body also locates the risk of undetectable error ^{(American Translators Association 2025)}. Credentials in those domains are the part of the job that a buyer cannot quietly route around.

Price and position as a post-editor, not a word-count supplier

Survey evidence from within the industry points to falling demand for human translation and rising demand for post-editing, even allowing for the seller's interest in that framing ^{(Acolad 2025)}. Post-editing throughput has been measured at 716 to 887 words an hour, which is a rate you can quote and defend rather than a per-word price competing with a machine ^{(Marina Sanchez-Torron and Philipp Koehn 2016)}. Quality reviewing also holds up where the model does not, on terminology, consistency and cultural appropriateness ^{(Frontiers in Artificial Intelligence 2025)}.

Test the adjacent quality-control roles before you need them

Large localization operations are keeping human linguists in the loop while translators and dubbing directors move toward editor and quality-controller roles ^{(The Localization Institute 2025)}. Canada's Job Bank rates regional outlooks Moderate for 2025-2027 while noting that AI may change some tasks in the occupation ^{(Government of Canada Job Bank 2027)}. Task change of that kind arrives before headcount change, which makes the current period a reasonable one in which to try the adjacent work.

Tools and courses that fit this situation

Below are paid tools and courses that fit the situation this page describes — a role where the drafting layer is delegable and the review, domain and credential layers are where the remaining value sits. None of them removes the exposure.

Some links above are affiliate links. If you buy through them, Jobs After AI may earn a commission at no extra cost to you. It does not change which tools appear here.

Common questions

Which parts of a translator's job can AI already handle?

General document translation is the clearest case: websites, contracts, blogs and internal material have already moved, on one translator's account driving demand down more than 70% since 2023 ^{(Blood in the Machine 2025)}. Media localization runs neural machine translation across dozens of languages with human linguists reviewing the output rather than producing it ^{(The Localization Institute 2025)}. Applicability for this occupation is the highest recorded in the Microsoft data, at 0.492 ^{(Microsoft Research 2025)}.

How is the 26 out of 100 score for translators calculated?

It follows the applicability ranking directly, and this occupation holds the top position in the Microsoft measurement at 0.492 ^{(Microsoft Research 2025)}. Eloundou's task exposure work separately marks translation fully exposed ^{(Eloundou et al. 2023)}. Confidence in the placement is high, which is why the role sits 10th of 159 rather than somewhere softer.

What is AI still bad at in this job?

Professional post-editors outperformed ChatGPT-4o on accuracy, terminology, consistency, coherence, grammar and cultural appropriateness in a 2025 study of Arabic translations ^{(Frontiers in Artificial Intelligence 2025)}. Weak machine output also degrades the finished product, with one sentence in four falling below the acceptable quality threshold for the poorest system tested ^{(Marina Sanchez-Torron and Philipp Koehn 2016)}. The trade body's specific warning is that fluent-looking output can contain errors a non-expert cannot see ^{(American Translators Association 2025)}.

How fast is this actually happening?

Faster in freelance document work than in the official statistics: employment in the US category moved only from 52,160 in May 2022 to 51,560 in May 2023, while median wages rose ^{(US Bureau of Labor Statistics 2023)}. Practitioner accounts describe a much steeper fall, including a translator who went from 50-60 hours a week to essentially no requests in June 2025 ^{(Blood in the Machine 2025)}. Employer announcements such as Duolingo's April 2025 plan to phase out contractor work AI can handle set intent without yet disclosing realized reductions ^{(The Verge 2025)}.

Should I still train as a translator?

The official projection is 2% growth from 2024 to 2034 with about 6,900 openings a year, and the agency attributes nothing to AI ^{(US Bureau of Labor Statistics 2024)}. Those openings are concentrated in work that is harder to delegate, including interpreting and the regulated review that Section 1557 requires ^{(US Department of Health and Human Services 2024)}. Entering on general document translation alone runs against the strongest exposure reading in the ranking ^{(Microsoft Research 2025)}.

Related occupations

The evidence behind this page

How this score was built

The AI-Proof Score is an estimate of how exposed an occupation's core tasks are to AI over roughly the next three to five years, on a 0–100 scale where higher is safer. It measures the exposure of tasks, which sits upstream of, and is softer than, job loss — it is not a probability that this job disappears, and not a claim about any individual doing it. The full derivation, the four research families behind it and its limitations are set out in the methodology, and every occupation is ranked in the AI Job Impact Report 2026.