AI can now translate documents, subtitle videos and support multilingual conversations in seconds. Yet newer research suggests that fluency is only one part of translation. As machines take over more routine language work, human translators are increasingly being relied on for what AI still struggles to guarantee: context, cultural judgement, consistency and accountability.
For anyone who has used an AI translation tool recently, the question is understandable: why would we still need translators?
A product description written in English can be converted into multiple languages almost instantly. Meetings can be captioned in real time. Videos can be subtitled for different markets. A marketer can ask a large language model to translate a campaign into Hindi, French or Japanese, then make it more formal, conversational or suitable for a younger audience.
Tasks that once required hours can now take seconds.
The technology is also moving beyond simple word-for-word translation. Generative AI systems can rewrite sentences, offer alternative phrases and adapt tone. For businesses managing large volumes of multilingual content, the appeal is clear: more languages, faster turnaround and potentially lower costs.
The European Language Industry Survey 2026, which drew responses from 1,058 participants across 45 countries, shows just how deeply automated translation has entered professional workflows. Among independent language professionals, the survey calculated 63% overall use of automated translation, including post-editing text provided by clients and translators choosing machine translation or generative AI themselves.
Businesses are becoming comfortable with the technology too.
A 2026 survey commissioned by AI translation company Wordly and conducted by Dimensional Research found that among 205 enterprise event leaders in the US and UK, 66% believed AI translation and captioning delivered better quality than human interpreters, while 95% said AI was more affordable and easier to deploy.
The findings need context because the study was commissioned by a company selling AI translation services. They are not an independent verdict on whether machines have surpassed professional interpreters. But they do point to a broader shift in how organisations view automated language technology, particularly where speed, scale and accessibility are priorities.
Lakshman Rathnam, Wordly’s founder and CEO, said organisations were increasingly “making decisions based on outcomes, not assumptions.”
AI, then, is clearly capable of doing translation work.
The more complicated question is whether translating words is the same as understanding what those words need to accomplish.
Fluent does not always mean right
One of the biggest advances in AI translation has also made its mistakes more difficult to detect.
Earlier machine translation frequently sounded mechanical. Sentences were awkward, grammar could be unreliable and literal translations often revealed that a machine had produced the text.
Generative AI can sound considerably more natural.
That creates a new problem. A sentence can be fluent, grammatically correct and convincing while still being the wrong translation for its audience or purpose.
The ELIS 2026 findings illustrate that gap. Despite widespread adoption of automated translation, only 23% of independent language professionals rated its quality as high or very high, down from 40% in the previous year’s survey.
Between 27% and 35% of the different groups surveyed also believed machine translation might never reach the quality of what the report describes as average human translation. The survey itself acknowledges that such comparisons are difficult because translation quality varies by language, content and context.
That distinction becomes important in marketing.
Consider a retailer translating “delivery available tomorrow”. There is relatively little room for interpretation. The message primarily needs to remain accurate.
Now consider a beauty brand launching a campaign built around a pun. Or an entertainment company translating dialogue where humour depends on a cultural reference. Or a financial services company adapting customer communication that needs to remain legally accurate without becoming difficult to understand.
The challenge is no longer simply finding equivalent words.
A translator may have to decide whether a joke should be changed, whether a phrase sounds offensive in another market, whether a literal translation weakens the emotional meaning or whether the language fits the way consumers actually speak.
This is also why localisation and transcreation have become important parts of international marketing. The most effective version of a campaign in another country may not be a literal translation at all. Examples, expressions, humour, sentence structures and cultural references may need to change while the original communication objective remains intact.
Carol Bereuter, a specialist health translator and lecturer, described the distinction in an interview with Le Monde: “Translation isn’t simply converting words from one language to another.”
The point becomes clearer in research comparing human and machine-produced text.
A 2026 study published in Humanities and Social Sciences Communications compared the human translation of a Chinese literary autobiography with outputs from Google Translate, ChatGPT-4o and OpenAI-o1. Researchers examined 83 aligned sections using 106 measures covering vocabulary, syntax, cohesion and readability.
Seventy-nine of those 106 measures showed statistically significant differences between the human and AI-based translations.
This did not mean every machine-generated sentence was poor. ChatGPT-4o came closest to the human version among the systems tested. But the human translation showed advantages in areas including coherence and how information was connected across the wider text.
It points to a distinction that is becoming increasingly important as AI improves. The question is no longer whether a machine can produce a translation. It is whether it can consistently make the appropriate translation decision for the audience, purpose and consequences involved.
When AI sounds human, readers can still prefer the human version
Literary translation offers one of the clearest tests because there is rarely only one technically correct way to translate a sentence.
A 2026 experiment by researchers from Simon Fraser University, Université du Québec à Montréal and Microsoft compared professional human translations with an advanced AI translation workflow across 15 recently published novels originally written in French, Polish and Japanese.
Importantly, the AI was not simply given a book and asked for a one-shot translation. The researchers used multiple AI systems to translate, review and revise the material.
Fifteen avid readers evaluated excerpts of roughly 8,000 words. Across 30 whole-excerpt comparisons, readers correctly identified the machine-produced version only 17 times.
In other words, AI translation had become convincing enough that readers often could not reliably tell which version had been produced by a machine.
But that was not the end of the experiment.
When participants examined 772 sections side by side, they preferred the professional human translation in 522 cases, or around 68%. Human versions were particularly favoured for qualities including clarity, dialogue, word choice, ease of reading and immersion. The researchers also found greater variation in machine translation quality within individual books.
The research is a preprint rather than a peer-reviewed published study, and its reader sample was small, so the findings should not be generalised to all forms of translation. But they highlight a useful distinction.
Passing as human is not necessarily the same as being preferred by humans.
For marketers, that matters.
A consumer does not need to recognise that an advertisement was translated by AI for the translation to underperform. A slogan can be grammatically correct and still feel unnatural. A joke can make sense but fail to be funny. A luxury brand can preserve the meaning of a sentence while losing the tone that made it feel premium.
The danger may increasingly be less about obviously bad machine translation and more about language that appears correct while being slightly wrong for the people receiving it.
Professor Janice Pan of Hong Kong Baptist University’s Academy of Language and Culture has described human interpretation as the “beauty of translation”, arguing that AI can be useful for generating drafts that people then assess for nuance, cultural sensitivity and bias.
This makes the human role less visible, but not necessarily less important.
One language can be very different from another
There is another reason why broad claims about AI replacing translators become difficult: AI does not perform equally across every language.
Models learn from data, and the world’s languages are not represented equally in digital datasets.
English, Spanish and other widely digitised languages have large amounts of online text, translated material and parallel language data. Many other languages have significantly fewer digital resources.
Recent healthcare research demonstrates what that difference can mean when translation carries consequences beyond convenience.
A 2025 study published in npj Digital Medicine examined AI translation of hospital discharge instructions across Arabic, Armenian, Bengali, simplified Chinese, Somali and Spanish. Forty-two linguists, clinicians and family caregivers assessed translations produced by ChatGPT-4o, professional linguists and a hybrid process in which professionals edited AI-generated translations.
AI performance varied considerably between languages.
ChatGPT-4o performed particularly poorly relative to professional translations for Armenian and Somali, which the researchers described as digitally underrepresented languages. Its performance was closer to professional translation for Bengali and Spanish.
The most notable finding, however, came from the hybrid workflow.
Human-edited AI translations were the most frequently preferred overall and took an average of 7.1 minutes to complete, compared with 16.8 minutes for translations created entirely by professional linguists.
Rather than proving that either humans or machines should perform the entire task, the study pointed towards another model. AI handled much of the initial work while people checked terminology, corrected weaknesses and applied contextual judgement.
The machine made the process faster. Human review made the output more dependable.
For India, the uneven-language problem is particularly relevant.
A company communicating across India is not simply translating “into Indian languages”. It may be working across Hindi, Bengali, Tamil, Telugu, Malayalam, Marathi, Urdu and multiple other languages, each with different levels of digital representation, regional variation and cultural context.
Research presented at Ashoka University’s Bhashavaad 3.0 translation conference in New Delhi in August 2026 examined AI-assisted literary translation from Malayalam. Researchers found that AI could improve when humans supplied targeted contextual information, but the resulting translation still required substantial human intervention and editorial judgement.
This means the question facing a marketer is not simply whether the company should use AI translation.
It is where, for which language, for what type of content and with what level of human review.
A frequently asked question on a website carries a different level of risk from the terms of an insurance policy. Internal meeting notes are different from a national advertising campaign. A rough translation for understanding a customer comment is different from communication that will appear publicly under the brand’s name.
AI may be perfectly adequate for one and require specialist human oversight for another.
The translator is not disappearing, but the job is changing
None of this means translators are protected from automation.
The profession is already being reshaped by it.
ELIS 2026 found that traditional human translation’s share of reported language-company revenue fell from 37% to 29%. Among language-company respondents, 47% reported lower staffing levels in 2025, compared with just 7% reporting an increase.
The survey also found that 59% of language companies reported a direct negative impact from AI, up 20 percentage points from the previous year.
At the same time, companies implementing AI reported productivity and efficiency gains.
Both things can be true.
AI can make translation businesses more productive while reducing the amount of routine work that previously required a person. The economic pressure on translators is therefore real even if the technology cannot perform every part of their job.
What appears to be changing is where human expertise enters the process.
Instead of beginning every project with a blank page, translators may increasingly receive an AI-generated first draft. Their work then becomes identifying what the system has misunderstood, checking specialist terminology, maintaining consistency, adapting cultural references and determining whether the translation serves its intended purpose.
In some cases, the professional becomes less of a first-draft translator and more of an editor, localiser, subject specialist and quality controller.
Lucile Munch of France’s National Chamber of Translation Companies captured the shift simply: “The key is knowing when to use AI.”
That may ultimately prove more useful than trying to divide translation into a competition between humans and machines.
For high-volume and predictable material, automated translation can offer enormous advantages. Help-centre articles, internal documents, product information and routine communication can increasingly be translated quickly and at a scale that would be difficult to achieve economically with human teams alone.
For creative, culturally sensitive, specialist or high-risk material, the case for human involvement remains stronger.
And between those two categories sits a growing middle ground where AI produces the first version and people decide whether it is ready to be trusted.
That distinction is likely to become more important as the technology improves.
The easiest machine translations to reject were those that sounded obviously robotic. The harder ones are fluent, polished and completely plausible, but miss a joke, change an implication, flatten a character’s voice or use words that people in that market would never naturally choose.
AI is getting remarkably good at moving words from one language into another.
What it has not eliminated is the need to decide what those words should mean when they arrive.
Disclaimer: All data points and statistics are attributed to published research studies and verified market research. All quotes are either sourced directly or attributed to public statements.