The Industry That Trained Its Replacement

Almost nobody outside it can name a translation company, and almost nothing in global commerce moves without one. The language-services industry — worth on the order of sixty to seventy billion dollars a year by the mid-2020s — is the plumbing of globalisation: every drug label in twenty-four EU languages, every aircraft manual, every subtitle, every patent filed abroad, every contract that crosses a border. Its structure is distinctive: a handful of large firms coordinating work that is actually performed by hundreds of thousands of freelancers, paid by the word, invisible even to the clients whose voices they become.

One corner of the industry is more dramatic than the rest, and worth meeting properly: over-the-phone and video interpreting. Hospitals, courts, police forces and emergency lines in immigrant societies cannot function monolingually — in the United States, access to an interpreter in healthcare is a legal right — so specialised firms keep thousands of interpreters on call around the clock, reachable in seconds, covering two hundred languages and more. An interpreter on such a line might carry a heart-attack call at nine, an asylum interview at ten, and a child-custody hearing before lunch — pure performance under pressure, where a mistranslated word can cost a case, a diagnosis, or a life. Keep this corner of the industry in mind; it behaves differently in every act of the story that follows.

And for decades, the whole trade had one quiet habit that becomes the case's hinge: it kept its work. Every translated sentence was stored beside its original in "translation memories," so that repeated content need never be paid for twice. Public institutions did the same at monumental scale — the European Parliament alone published decades of proceedings, professionally translated, in over twenty languages. By the 2010s, humanity had accumulated the largest parallel corpus of carefully matched sentences ever created: hundreds of millions of examples of exactly how an expert converts one language into another. Nobody thought of it as anything but an archive.

Machine translation was, for fifty years, the future that reliably failed to arrive. Rule-based systems produced word salad; the statistical systems of the 2000s — including early Google Translate — produced something better than nothing and worse than embarrassing, good enough for gist and a running joke among professionals. Then, in late 2016, Google switched its translation service to a neural network trained on parallel text, and the quality jump was so abrupt that it made global headlines: overnight, in major language pairs, output went from broken to frequently fluent. The training data that made this possible was, to a very large degree, the accumulated archive — the memories, the public corpora, the decades of human sentences. The industry had spent thirty years, without knowing it, writing its replacement's textbook.

Two ironies followed, each worth savouring. The first: in 2017, Google researchers looking for a better translation architecture published a paper called "Attention Is All You Need," introducing the transformer — the invention that later powered ChatGPT and the entire modern AI wave. The technology that would unsettle every knowledge profession was built, first and specifically, to translate. The second irony was European: that same summer, a small company in Cologne — DeepL, grown out of the translation-search engine Linguee — launched a neural translator that independent evaluations repeatedly ranked above the American giants' in its core languages. One of the finest translation machines on earth came from a mid-sized German firm; the disrupted industry's own continent produced the sharpest disruptor, later valued in the billions.

Inside the profession, the ground shifted within a few years. Agencies re-priced: why pay full rate for translation when the machine's draft was nearly right? The working offer became post-editing — clean up the machine, at half the rate or less — and the per-word price of ordinary commercial translation began a slide it never reversed. When large language models arrived after 2022, the pressure jumped again: now the machine could match tone, rewrite, localise idiom. Industry surveys filled with freelancers reporting collapsing volumes in "easy" segments; veteran translators of manuals and marketing copy watched decades of expertise re-priced as review work. And in 2025 a well-known language-learning company became the public face of the shift when its "AI-first" strategy memo — with contractors replaced where the machine sufficed — leaked into a week-long consumer backlash. The argument had left the trade press and reached the general public.

Yet the industry did not shrink — and that paradox is the case's second half. Total language-services revenue kept growing through the disruption, because demand exploded: at machine prices, the world began translating things that would never have justified human rates — every product page, every support ticket, every video's subtitles in thirty languages. The pie grew while the per-word slice collapsed. The companies that thrived changed what they sold: not words but assurance — workflows in which machines produce volume and humans carry judgement, with the price attached to the risk of the content rather than the length of it. A tourist menu goes through raw; a pacemaker manual, a court bundle, a drug label goes through people, because when those are wrong, someone must be accountable — and a model cannot be sued, sworn in, or struck off.

Which is why the spoken corner of the industry — the interpreting lines — became the fortress. Demand for medical and legal interpreting kept rising through the AI decade: migration did not pause, regulations still required qualified humans, and no hospital's lawyers would put an unaccountable system alone between a doctor and a consent form. The interpreter's product was never really words-per-minute; it was a credentialed human being, present in the moment, responsible for what both sides understood — closer to a nurse or a court officer than to a document. Machines entered as assistants — transcribing, suggesting, covering languages no one could staff — while the accountable human stayed on the line. The pattern generalises, and by the mid-2020s every profession watching AI was studying it: automation flows fastest where errors are cheap and slowest where someone must answer for them.

The case remains live, and honest telling requires the open questions. Model quality keeps rising; the boundary of "errors are cheap" keeps moving; some interpreting itself — routine retail calls, basic support — is already crossing to machines, and the fortress holds only its high ground. Meanwhile the profession's pipeline is quietly cracking: if the entry-level work that once trained juniors is automated, where does the next generation of experts come from — the same succession problem Novo Nordisk faces with molecules, arriving here as people. The question this industry answered first, every other profession now faces in translation — so to speak: when the machine can produce your output, what exactly is it that you sell? The translators who are still standing have an answer, and it fits on an invoice: not the words — the responsibility for them.