AI in Translation and Localisation: The Quality Gap Is Nearly Closed

Professional translators have been hearing "machine translation will replace you" for twenty years. For most of that time, they could laugh it off. Google Translate was useful for getting the gist of a foreign language website but the output was clearly robotic, often comically wrong, and nowhere near publishable quality. The gap between machine translation and human translation was enormous.

Then LLMs arrived and the gap nearly closed overnight.

i'm not a translator but i worked with localisation data and multilingual content projects, and the change in output quality from 2022 to now is staggering. What used to require a professional translator working for hours can now be produced at near-publishable quality in seconds. Not for everything. Not perfectly. But for a huge swath of content that the translation industry used to charge premium rates for.

This isn't Google Translate getting incrementally better. This is a fundamentally different technology that understands context, nuance, tone, and cultural references in ways that previous machine translation simply could not. And it's changed the economics of the entire industry.

What changed with LLMs

Previous machine translation (statistical MT and neural MT systems like Google Translate or DeepL) worked by learning patterns from parallel texts — documents that existed in both the source and target language. They were good at common language pairs with lots of training data (English-French, English-Spanish) and poor at less common pairs. They translated sentence by sentence, often losing context across paragraphs. They struggled with idioms, cultural references, tone, and anything that required understanding the meaning rather than matching patterns.

LLMs translate differently. They understand the content. They maintain context across paragraphs and documents. They can match tone — formal, casual, technical, marketing, literary. They can be instructed to adapt cultural references rather than translate them literally. They can handle idioms, humour, and nuance in ways that previous MT could not.

The quality for general content — business documents, technical documentation, website content, news articles, product descriptions — is now genuinely close to professional human translation. Not identical. A skilled human translator will still produce slightly better output. But the gap has narrowed to the point where most readers wouldn't notice the difference, and many clients don't consider the remaining gap worth the premium.

For specialised content, the gap is larger. Legal contracts, literary fiction, marketing copy that needs to resonate culturally, medical texts where precision is life-or-death. These still benefit significantly from human expertise. But they're a fraction of the total translation market by volume.

The bulk content revolution

The translation industry's bread and butter has always been bulk content. Technical manuals, product documentation, software strings, website content, corporate communications, e-commerce product descriptions. This work isn't glamorous but it's been a reliable revenue source for translation agencies and freelance translators for decades.

AI has decimated this segment. A company that used to pay translators to translate its product catalogue into 20 languages can now generate those translations with AI at a fraction of the cost and time. The quality is good enough for the purpose. The product description for a pair of trainers doesn't need literary translation. It needs to be accurate, readable, and SEO-friendly. AI delivers that.

Software localisation — translating user interfaces, help documentation, error messages — is similarly affected. These are typically short, standardised strings with limited context requirements. AI handles them well, especially with proper prompting and glossary management.

Website localisation for informational content, corporate communications for internal audiences, routine business correspondence — all increasingly AI-translated with minimal or no human review.

The volume of professional translation work in these bulk categories has dropped significantly. Some agencies report 30-50% revenue declines in bulk translation. The work hasn't disappeared — companies are still translating content. They're just not paying human translators to do it.

Machine Translation Post-Editing: bridge or dead end?

MTPE — Machine Translation Post-Editing — has emerged as the industry's compromise position. AI translates the content. A human translator reviews and edits it. The idea is that it's faster and cheaper than translating from scratch while maintaining quality through human oversight.

The reality is more complicated.

MTPE pays significantly less than translation. Typically 30-60% less per word. The reasoning is that the translator is "only" editing rather than translating, so it should be faster and therefore cheaper. But experienced translators report that editing poor MT output can take nearly as long as translating from scratch, because you have to read the source, read the MT output, assess whether it's correct, fix what's wrong, and ensure consistency. When the MT quality is good, MTPE is efficient. When it's bad, it's painful.

The psychological aspect matters too. Many professional translators find MTPE deeply unsatisfying. You're no longer translating — you're cleaning up after a machine. Your creativity, your craft, your years of language study are reduced to fixing errors in AI output. The work is tedious in a way that translation isn't. Several translators i've spoken to describe it as demoralising.

As AI quality improves, the MTPE role narrows further. When the machine output is 95% acceptable, there's less for the post-editor to do but they still need to read every word carefully to catch the 5% that's wrong. The effort-to-output ratio makes less and less economic sense for the client, who starts asking: "Do we really need post-editing for this content?"

MTPE is probably a bridge rather than a destination. It's providing income for translators during the transition but the economic logic points towards either full AI translation (for content where near-enough is good enough) or full human translation (for content where quality is critical). The middle ground of AI-plus-light-editing may not sustain a large workforce long term.

What still needs humans

Legal translation. Contracts, legislation, court documents, patent filings. The precision requirements are extreme. A mistranslation can change the meaning of a contract, invalidate a patent, or compromise a legal proceeding. Translators working in this space need deep legal knowledge in both jurisdictions. AI can assist but the stakes are too high for unsupervised AI translation.

Literary translation. Novels, poetry, creative non-fiction. This is translation as art. The translator isn't just converting words; they're recreating a literary work in another language. The choices involved — how to handle voice, rhythm, cultural references, wordplay, ambiguity — require creative talent that AI doesn't have. Literary translation is a small market but it's about as safe from AI as any translation work gets.

Marketing localisation (transcreation). Taking a marketing campaign and adapting it for a different market isn't translation. It's transcreation — recreating the creative concept in a way that resonates culturally. A pun that works in English needs a different pun in French. A cultural reference that resonates in the US might mean nothing in Japan. This requires cultural fluency and creative skill.

Medical and pharmaceutical translation. Clinical trial documentation, patient information leaflets, regulatory submissions. The regulatory requirements and life-or-death accuracy requirements mean human translators are still essential, though AI is increasingly used for first drafts that humans review.

Audiovisual translation. Subtitling and dubbing require fitting translations to time constraints, lip movements, and visual context. AI is getting better at this but the creative and technical constraints make full automation difficult.

Rare language pairs. AI performs best for well-resourced language pairs. Translation involving less common languages — many African languages, indigenous languages, smaller Asian languages — still requires human expertise. The training data for AI in these languages is limited and the quality reflects that.

The agency landscape

Translation agencies are restructuring rapidly. The traditional model — receive project, assign to translator, QA check, deliver — is being replaced by technology-mediated workflows where AI does the heavy lifting and humans handle quality control and complex content.

The agencies that are surviving are:

Technology-forward agencies that have integrated AI into their workflows, reduced costs, and passed some savings to clients while maintaining margins through efficiency. They position themselves as language technology companies rather than translation companies.

Specialist agencies focused on high-stakes content — legal, medical, regulatory — where human quality assurance is non-negotiable and clients will pay premium rates for it.

Creative agencies that focus on transcreation, marketing localisation, and brand content. These compete on creative quality rather than price.

The generalist translation agency that competed on volume for bulk content is struggling badly. The economics of paying human translators to translate content that AI can handle is not sustainable when clients can see the alternative.

What to do if you're a translator

Specialise ruthlessly. The generalist translator handling a mix of business, technical, and marketing content is the most exposed. The legal translator, the medical translator, the literary translator — they have defensible niches. Pick an area where accuracy is critical and stakes are high, and build deep expertise.

Learn to work with AI, not against it. The translators who use AI to enhance their productivity — using it for first drafts, terminology research, consistency checking — can deliver better work faster. The translators who refuse to use AI on principle are producing the same output at higher cost. Principled resistance is understandable but economically damaging.

Move up the value chain. From translation to consultancy. Linguistic consulting, cultural advisory, localisation strategy, quality assurance programme design. If you understand language, culture, and the target market at a deep level, there's consultative value that goes beyond translating individual documents.

Consider adjacent roles. AI training data curation, AI quality evaluation, prompt engineering for multilingual AI systems, cross-cultural communication consulting. Your linguistic skills are valuable in contexts beyond traditional translation.

Be realistic about rates. Bulk translation rates are not coming back. The market rate for content that AI can handle acceptably is set by AI's cost, not by the value of human expertise. You can either compete at those rates (unsustainable) or focus on content where the human premium is justified and clients will pay it.

Build direct client relationships. Translation agencies are struggling, and their struggles become your struggles if you depend on them for work. Direct relationships with clients who value your specific expertise give you more control over rates and work quality.

The translation industry isn't disappearing. Language barriers still exist and culturally competent communication across languages still requires human expertise for important content. But the volume of work that requires human translation is shrinking, and the economic value of bulk translation has collapsed. The translators who thrive will be the ones who provide something AI genuinely cannot — deep specialist knowledge, creative talent, or cultural insight that goes beyond language conversion.

The one thing to do today: take a piece of your recent work and run the source text through an LLM translator with careful prompting. Compare the output honestly with your own. Where is the AI close? Where does it fall short? That gap — and only that gap — is where your value lies. Make sure it's a gap clients care about.