Will AI Replace Translators? What the Evidence Says, and How to Ride the Wave

If you translate for a living, someone has probably asked you by now whether AI is about to take your job. It is a fair question. In the latest international evaluation of machine translation, the best AI system matched or beat the professional reference translations in 10 of the 16 language pairs that people evaluated, though in one of them, English to Maasai, both scored under 10 out of 100.¹ Real-time speech translation now ships in phones and earbuds.^2,3

The fuller picture is more useful than the headline. The evidence from 2024 to 2026 shows AI splitting the profession rather than ending it. Routine, generic work is shrinking and getting cheaper. Work that carries risk, needs a signature, crosses cultures or happens live is holding up, and in places growing. This guide brings together the most reliable surveys, research and UK rules we could find, with every figure linked to its original source, and turns them into a practical plan.

What has actually changed

Start with the money. Nimdzi, one of the industry's main analysts, estimates the language industry at USD 72.6 billion in 2025 and, in its realistic forecast, now projects growth of under 1% a year, reaching USD 76.1 billion by 2030.⁵ In 2024 it expected the industry to pass USD 100 billion by 2029. It cut that forecast in 2025, pointing to buyers holding back while they worked out what generative AI meant for them, and to tighter budgets,¹¹ then cut it again in 2026.⁵ Growth has not disappeared; it has moved. Interpreting makes up a third of the top 100 providers' revenue, new AI and data services were the biggest source of new revenue in 2025, and desktop publishing and linguist staffing lag behind.⁵

The job evidence tells the same story. An Oxford study of 695 US local labour markets estimates that about 28,000 fewer translator positions were created between 2010 and 2023 than the profession's growth path implied, as machine translation spread. Wages dipped, then recovered from around 2016, which the authors link to rising demand for higher-skilled translation that machines cannot easily replace.¹² On one large online freelance platform, demand for writing and translation fell by 20 to 50% against the expected trend after ChatGPT, with the sharpest fall for short jobs of one to three weeks.⁴ The US Bureau of Labor Statistics now projects employment of interpreters and translators to grow 2% from 2025 to 2035, slower than average, while noting that many of these jobs cannot be entirely automated.¹³

Two widely shared findings deserve a second look. A Microsoft Research study gave interpreters and translators the highest "AI applicability" score of any occupation, but its authors state plainly that reading this as a forecast of job or wage loss would be a mistake.¹⁴ And Duolingo's much-reported cut at the end of 2023 affected about 10% of its contractor workforce overall, while the company said human experts still validate its AI translations.¹⁵

What translators themselves are reporting

The mood is anxious, and the numbers explain why. In a January 2024 Society of Authors survey, over a third of translators (36%) said they had already lost work to generative AI, 43% said the income from their work had fallen in value because of generative AI, and 77% expected it to hurt their future income. The survey went to Society of Authors members and other authors, so its translators are not a sample of legal or commercial translators.¹⁶ ITI's Autumn 2024 Pulse survey found 42% of members very or extremely concerned about AI's effect on earnings, and 51% said work from existing clients had fallen.⁸

The same surveys also show people adapting, and some doing well. In the 2026 European Language Industry Survey, 17% of independent professionals said they were considering ending their freelance work, yet almost one in two were learning new skills or building another main activity.⁶ One in four raised their prices in 2025, while only 12% cut them.⁶ In ITI's survey, slightly more members reported rate improvements (26%) than reductions (23%), and 63% had already invested in training.⁸ A 600-person international survey in spring 2024 found that 62% saw generative AI as a threat to their profession's finances and 35.5% reported less work over the previous few years, yet 24% reported more work and 28% reported higher pay rates.²⁶

Where AI is strong, and where it still fails

It is worth being honest about how good the machines have become. In WMT25, the main annual evaluation of machine translation, Gemini 2.5 Pro was in the top group for 14 of 16 language pairs and on par with or better than human translation in 10 of them. The organisers add three important caveats: their human references were prepared by professional translators but were not necessarily free of errors or checked to the same level as sworn or published translations, the models were weak on Egyptian Arabic and largely unusable into Maasai, and human evaluation "should remain the final arbiter of translation quality".¹

The gaps show up most where mistakes are costly. In a 2019 study of emergency department instructions, Google Translate rendered 92% of sentences accurately into Spanish but 81% into Chinese, and 8% of the Chinese sentences had the potential for clinically significant harm. In one case, the Chinese version of "Hold the kidney medicine until you have a chance to speak with your kidney doctor" translated back as "Keep taking kidney medicine until you talk to your kidney doctor".²⁷ Newer tools have narrowed the gap in major languages but not everywhere: a 2025 study found AI translation of real hospital discharge instructions performed similarly to professionals for Spanish, but worse for Chinese, Vietnamese and Somali,²⁸ and a 2024 study found ChatGPT and Google Translate made more potentially serious errors in Haitian Creole than professional translators did.²⁹

The errors are also getting harder to spot. The Chartered Institute of Linguists warns that AI mistakes with numbers, names, terms and who did what to whom are more common in less well-resourced languages and may become harder to notice,³⁰ and the American Translators Association points out that AI translations can look accurate to anyone without linguistic expertise.³¹ In 2023, volunteers told the Guardian of US asylum applications where "mi jefe", used by a woman to describe her father, was translated as "my boss", and where "I" became "we".³²

Where a human is still required in the UK

For official documents, UK guidance asks for a person. GOV.UK says the translator should confirm in writing that the translation is a "true and accurate translation of the original document", with the date, their full name and contact details.⁹ Home Office visa guidance adds the translator's signature and says the translation must be one the Home Office can independently verify.¹⁰ Even Meta's own open translation model, NLLB-200, states in its documentation that its translations "can not be used as certified translations".³³

Healthcare, courts and public bodies point the same way. NHS England's 2018 guidance for primary care says automated online translation such as Google Translate "should be avoided", and its 2025 improvement framework warns that AI translation apps, now widely used in the NHS, carry risks for accuracy and patient safety.^34,35 Interpreters support around 600 court and tribunal hearings a day,³⁶ and an independent review for the Ministry of Justice recommended degree-level (RQF Level 6) qualifications as the default for professional interpreting bookings.³⁷ Here in Wales, the Welsh Language Commissioner advises that machine-translated text should not be published without thorough editing by a human translator.³⁸

The public sector is not anti-AI. The Ministry of Justice translated its own AI Action Plan into Welsh with an AI engine, then had it quality-assured by the HMCTS Language Services Team.³⁹ That is the model in miniature: the machine drafts, a qualified human checks and takes responsibility. Confidentiality matters too. The judiciary's AI guidance tells judges to treat anything typed into a public AI chatbot as "published to all the world".⁴⁰

Seven moves to ride the wave

The professional bodies cited in this article, ATA, CIOL, ITI and FIT, have landed in the same place: do not fight the tools, and do not become a passive editor of them either. The American Translators Association says growing reliance on AI makes expert human linguists more necessary, not less,³¹ and the Chartered Institute of Linguists expects skills in conveying true meaning, quality assurance, post-editing and cross-cultural judgement to become more valuable.³⁰ Here is what that means in practice.

1. 1

   Specialise where mistakes are expensive

   Generic text is where AI is strongest, and Nimdzi reports that linguists left with basic post-editing face significant downward pricing pressure.⁵ Specialisation is the clearest defence in the data: in ELIS 2026, most of the few freelancers who reported no negative trend got more than half their work from a single client sector, and the report calls high specialisation a way to carve out a defensible niche.⁶ Nimdzi reports that experienced specialists who move into auditing, cultural editing and managing AI workflows are commanding higher premiums.⁵

   Do this: choose one or two fields where errors carry legal, medical or financial risk, learn their terminology and rules, and say so clearly on your CV and profiles.

2. 2

   Be the expert in the lead, not the button presser

   The UK's Institute of Translation and Interpreting backs what translator Michael Bailey calls an "expert in the lead" approach, in which skilled professionals decide how and when technology is used, and the International Federation of Translators prefers "human at the core" to "human in the loop".^41,42 In comments to the Chartered Institute of Linguists' 2026 poll, research and terminology was by some distance the most common use of generative AI, and very few linguists described using it to produce translations.⁴³ ITI members who felt positive about the future were more likely to have invested in training: 72%, against 53% of those who felt negative.⁸

   Do this: learn two or three AI tools properly, including where they fail, and keep your own judgement as the final check. If you work with EU clients, note that the EU AI Act already expects organisations using AI to support their staff's AI literacy, and the European Commission gives translating text with ChatGPT as an example.⁴⁴

3. 3

   Treat post-editing as a skilled service, and price it that way

   ISO 18587 draws a clear line between full post-editing, which aims for a result comparable to human translation, and light post-editing, which only aims for text you can understand.⁴⁵ Know which one you are being asked for. ELIS 2026 found post-editing discounts cluster around 20%, not the 90% of forum horror stories,⁶ and the International Federation of Translators notes that post-editing can even be slower than translating from scratch, depending on the language pair and text.⁴² The industry itself is moving away from per-word pricing towards hourly and platform models.⁵

   Do this: ask for a sample before agreeing a discount, quote full post-editing by the hour or by the job when the machine output is poor, and turn down light post-editing for anything that will be published or signed.

4. 4

   Sell quality, review and risk removal

   Nimdzi reports that some buyers who switched to AI-only translation suffered high-profile quality failures and came back to language providers for expert validation.⁵ ELIS 2026 found that language companies' growth plans focus strongly on post-editing but give considerably more attention than last year to quality evaluation, transcreation and localisation, and that independent professionals now rate machine and AI translation quality lower: only 23% scored it high or very high, down from 40% a year earlier.⁶ ISO 5060:2024 now gives guidance on evaluating human, post-edited and raw machine translation.⁴⁶

   Do this: add revision, language quality assessment and review of AI-translated content to your services, and describe them as risk removal for the client.

5. 5

   Move closer to live, spoken work

   Interpreting is one of the industry's strongest growth areas: Nimdzi estimates it at about USD 11.7 billion in 2024, projects around 8% growth a year to 2029,¹⁷ and expects interpreters' hourly and daily rates to increase.⁴⁷ In ELIS 2026, income satisfaction fell for independent professionals, except for those working only as interpreters, whose satisfaction rose considerably.⁶ Be clear-eyed, though: Oxford researchers expect the effects of machine translation to intensify, particularly for simultaneous interpreting.¹²

   Do this: if you have strong spoken skills, look at public service interpreting qualifications such as the DPSI and registration with the NRPSI, which NHS England's primary care guidance expects of spoken language interpreters.³⁴

6. 6

   Own the languages and cultures AI finds hardest

   AI is weakest where training data is thin. The WMT25 evaluation found models weak on Egyptian Arabic and output into Maasai largely unusable,¹ and researchers report weaker, sometimes much weaker, performance even for very large languages such as Hindi, Bengali, Urdu, Tamil, Swahili and Farsi.²⁶ Culture matters too: after Google's neural machine translation launched, regular translation jobs on one platform fell, but transcreation did not.⁴⁸

   Do this: if you work in a lower-resourced language or dialect, make it the centre of your offer. If you work in a major language, build the cultural and creative skills machines still miss.

7. 7

   Prove it, and get closer to your clients

   In the UK, translation and interpreting are unregulated, so anyone can call themselves a translator.⁴¹ That makes recognised credentials worth more, not less: the CIOL Level 7 Diploma in Translation is set at Master's level and regulated by Ofqual, and leads towards Chartered Linguist status;⁴⁹ ITI's Qualified (MITI) membership is for translators with at least three years' professional experience.⁵⁰ Direct relationships matter as much. Translators with more direct-client work reported significantly better job satisfaction and fairer pay,⁷ and the share of freelance work coming from direct clients rose to 48% in ELIS 2026, from 42% in 2024.⁶

   Do this: work towards one recognised credential, and aim for part of your income to come from clients who know you by name.

What this means if you buy translation

Use AI tools freely to understand a text for yourself. For anything you will submit, sign, publish or act on, use a qualified professional and ask how their work is checked. Certified translations for official use at Lingo Service Translations are produced by a qualified human translator from the original, and we do not certify AI translations. If you already have an AI draft for non-official use, our AI vs human translation guide explains your options, including our AI Translation Check, which reviews and corrects AI drafts for non-official use. For official documents, see our certified translation service.

Frequently asked questions