Translator · The profession that carries meaning between languages — from the Septuagint and Jerome's Vulgate to Toledo, Nuremberg's booths and the machine-translation age.
No profession on this site faces the machine question more directly: translation is the task machine learning was practically built around, from the Georgetown–IBM demonstration of 1954 to the transformer architecture — invented for translation — that powers today's language models. Neural systems have drafted a large and growing share of the world's commercial translation since 2016, and pretending otherwise would insult the reader.
But the honest answer has a second half. What the machine took first is the middle of the market — routine text where a fluent-sounding approximation is acceptable. What it has not taken is the work where translation carries liability, voice or life-and-death stakes: the sworn court document, the clinical protocol, the novel, the brand. There, humans still sign, and someone must be answerable for every fluent-sounding error the machine hides.
Most of the profession's traditional task — producing a competent first draft of routine text — is already automated, and post-editing machine output is now a standardized service with its own ISO norm. What resists is task-shaped, not title-shaped: legally sworn translation, high-stakes medical and legal content, literary and marketing voice, and low-resource languages the models handle poorly. The realistic future is a smaller profession concentrated at the accountable, creative top of the market.
Scored from the tasks, not the job title. Lower is safer.
Jobs AI cannot take →A novel's translation is authored, not retrieved: prize juries, publishers and readers can tell. The International Booker splits its prize money equally between author and translator precisely because the English book is partly the translator's writing.
Courts, ministries and immigration systems accept translations only from humans who can swear an oath, stamp a document and be sued. No machine can carry legal responsibility, and no state currently proposes to let one.
A mistranslated dosage, contract clause or safety warning costs lives and fortunes — and machine errors are precisely the fluent-sounding kind that only an expert bilingual reader catches. Where the cost of error is high, the human check is the product.
Marketing that must land in a culture — names, humor, taboos, register — is rewritten, not transferred. The judgment about what will work in Osaka or São Paulo draws on lived cultural knowledge models only imitate.
Machine quality collapses outside the top few dozen data-rich languages. For hundreds of languages — including many with millions of speakers — human translators remain the only serviceable option, and training data often does not exist to change that soon.
Manuals, product listings, support content and internal documents are now routinely machine-drafted, with humans post-editing to a defined quality level under ISO 18587 workflows — at a fraction of from-scratch rates.
CAT tools automatically flag inconsistent terms, wrong numbers, missing tags and deviations from the termbase — mechanical checking that once consumed a large share of a reviser's day.
Reviews, support tickets, social posts and internal email are consumed as raw machine output with no human in the loop at all — a vast volume of translation that simply never reaches a professional anymore.
Speech recognition plus machine translation drafts subtitles at industrial scale for streaming platforms; human subtitlers increasingly time, condense and correct rather than translate from zero.
Machine-translation post-editing became a standardized service with ISO 18587 in 2017, and it now anchors the bulk market. The skill shifts from drafting to diagnosing — catching the confident, fluent error — and pricing shifts from per-word toward hourly and per-task.
Bulk work races toward machine prices while premium work — sworn, medical, literary, brand — holds or rises, hollowing out the generalist middle where most twentieth-century freelancers lived. Specialization has gone from career advice to survival condition.
As the anonymous middle automates, the human end professionalizes its visibility: cover-credit campaigns like #TranslatorsOnTheCover (2021), the International Booker's equal split, and sworn stamps and certifications that function as a mark of accountable human work.
Clients who know a machine draft exists expect delivery in hours, not days. Workflows compress: the translator is consulted later, on less text, with the machine's version already on the table — changing negotiation, pricing and what 'urgent' means.
Running multilingual product launches across dozens of markets — coordinating translators, engineers, machine pipelines and release deadlines. The management layer of the language industry has grown as fast as its automation.
Recreating advertising and brand voice in another culture, briefed like a copywriter and paid like one — the corner of the trade furthest from per-word pricing and closest to pure writing.
Designing tests, scoring machine output and auditing post-edited work against defined quality frameworks — the profession's craft knowledge repackaged as the quality-control layer of the machine pipeline.
Curating training data, writing and rating multilingual model output, and building terminology and evaluation sets for language models — bilingual expertise sold to the systems that disrupted the old market.
The tasks already lost are the ones that could be specified in advance: routine text, tolerant of approximation, in data-rich languages. The tasks that remain are the ones defined by accountability and voice — where a human must swear to the court, sign off on a dosage, or write English good enough that García Márquez would call it better than the original. That boundary will keep moving, but it moves task by task, not all at once.
The numbers tell a double story: the language industry around the machine keeps growing — analysts size it around seventy billion dollars in the mid-2020s — even as traditional per-word translation shrinks inside it. Employment forecasts like the US BLS's low-single-digit growth mask the reallocation underneath: fewer generalist translators, more post-editors, localization managers, quality evaluators and language-data specialists.
For anyone entering now, the honest advice follows from the tasks: pair the languages with a field where errors are expensive, work into your native language at a level no machine imitates, and treat the machine as the industrial base of the trade rather than its rival. The profession that survived the printing press and the telegraph will survive this too — smaller, stranger, and concentrated where the words matter most.
The scholar paid to profess a subject in public — twenty-five centuries of teaching and discovery, from Plato's grove to the tenure track and the AI tutor.
AI-resistant 76 🏫The world's largest profession: from Sumerian tablet houses to AI-era classrooms, one adult tasked with turning thirty strangers into a room that learns.
AI-resistant 78 📜The trained interrogator of the past — from Herodotus and Sima Qian to the digital archive, turning fragile documents into checkable accounts of what happened.
AI-resistant 70 🧩The scientist of mind and behavior, from Wundt's 1879 Leipzig laboratory to the therapy room — trained to hear what a person cannot yet say themselves.
AI-resistant 82 📖The keeper and guide of recorded knowledge, from Nineveh's clay tablets to the e-lending fight — organizing what humanity knows so that anyone can find it.
AI-resistant 55