The file landed in her inbox at 8:42 on a Monday: forty thousand words of a German medical-device user manual, target French, due Thursday. She opened it in her CAT tool, the computer-assisted translation environment where she has spent twenty years living segment by segment, and every single segment was already filled. Not with fuzzy matches from the translation memory, the database of her own past translations, but with crisp, confident French that the project manager's note called "MT pre-translation." She read the first three segments and felt the old reflex relax: this was good French. Grammatical, natural, the right register for a clinical audience. The fourth segment described a contraindication. The German warned that the device must not be used on patients with an active pacemaker. The French in front of her, fluent and clean and ready to accept with a single keystroke, said the device should be used on patients with an active pacemaker. The negation was gone. Nothing on the screen flagged it. The sentence read perfectly. And if she had been moving at the speed the deadline demanded, trusting the smooth surface the way her eye wanted to trust it, she would have pressed the key, banked the segment, and shipped a sentence that could kill someone, with her name on the delivery. This lesson is about why that happened, what kind of machine produced it, and why the words "the machine translates as well as a human" are not just wrong but dangerous in the exact place they are most often said.

Three Machines Wearing One Costume

When a client, a journalist, or a panicked colleague says "AI can translate now," they are collapsing at least three distinct technologies into one word. Inside a localization shop, that collapse is the source of most bad decisions: bad pricing, bad routing, bad trust. Before you can own quality on top of any engine, you have to be able to name which engine you are standing on. There are three, and they behave differently enough that treating them as interchangeable is the professional equivalent of treating a scalpel, a bread knife, and a chainsaw as "blades."

Machine translation (MT) is the umbrella term, and it is old. It means any system that converts text from a source language to a target language without a human writing the words. The phrase covers everything from the clunky rule-based and statistical systems of the 1990s and 2000s to today's engines. When someone says "MT" without qualification in 2026, they almost always mean the current dominant flavor, which has its own name.

Neural machine translation (NMT) is the engine that has run the industry's high-volume pipelines since roughly 2016. It is a neural network trained specifically and only on the task of translation: feed it millions of source-and-target sentence pairs, and it learns to map one language onto another. The well-known production engines that localization teams have wired into their translation-management systems (the TMS, the platform that routes files, applies translation memory, and now pre-populates segments) are NMT systems at their core. An NMT engine is a translation specialist. It does one job. It does not chat, it does not summarize, it does not reason about your instructions. You give it a source segment, it returns a target segment. That narrowness is a feature: it is fast, it is cheap at scale, and its failure modes, while real, are comparatively predictable.

A large language model (LLM) is a different animal that arrived in force around 2022 and 2023. An LLM is a general-purpose text-prediction system trained on an enormous, broad corpus of human writing to predict the next word given everything before it. It was not built to translate. Translation is one of many things it can do because translation is well represented in its training data and because predicting plausible next words in another language is, mechanically, something it can do. When you ask a general chat model to translate a paragraph, you are asking a generalist to perform a specialist task by improvisation. Sometimes that improvisation is brilliant, more fluent and more context-aware than a narrow NMT engine. Sometimes it is confidently, catastrophically wrong in ways an NMT engine would not be, because the LLM is following the gravity of plausible prose rather than the discipline of faithful transfer.

"AI can translate" is not a fact about one machine. It is a slogan that hides three machines, each with a different job, a different failure mode, and a different price.

Why the Costume Fools People

All three of these systems produce the same artifact: fluent target-language text on a screen. The output looks identical. A clean French sentence from a 2018 NMT engine, a clean French sentence from a 2026 NMT engine, and a clean French sentence improvised by a general LLM are indistinguishable to the eye. They have the same fonts, the same grammar, the same confident absence of any uncertainty marker. The costume is the fluency, and the fluency is universal. What differs is everything underneath: how the words were chosen, how faithful they are to the source, how the system behaves when it does not know, and what it costs. A linguist who cannot see past the costume is a linguist who prices, routes, and trusts blind.

What Neural MT Actually Does Inside the Shop

Strip away the marketing and a neural machine-translation engine is a pattern-completion machine pointed at one task. During training, it saw vast quantities of parallel text: sentences in the source language paired with their human translations in the target. It learned, statistically, that when this pattern of source words appears, this pattern of target words tends to follow. It does not understand the German sentence about the pacemaker. It has no model of pacemakers, of patients, of harm. It has a model of which French word-sequences are probable given that German word-sequence, learned from the millions of sentence pairs it ingested.

This matters enormously for how you read its output. The engine is optimizing for fluency and plausibility in the target language, calibrated against the patterns it learned. Faithfulness to the specific source segment in front of it is an emergent property, not a guaranteed one. Most of the time the most probable French sequence genuinely is the correct translation, which is why NMT is good enough to be the baseline first pass for high-volume work. But "most of the time" is doing heavy lifting in that sentence, and the gap between "most of the time" and "every time" is exactly where your job now lives.

The Predictable Failure Modes of NMT

An NMT engine fails in characteristic, almost diagnosable ways. Learning them is the start of reading output skeptically.

• Dropped negation. The pacemaker example. Negation is linguistically light: a single "not," "no," "ne...pas," "kein" can flip the meaning of a sentence, but it carries very little statistical weight. An engine optimizing for fluent, probable prose can smooth right over it, producing a grammatical sentence that asserts the opposite of the source. The output reads beautifully. It means the reverse.
• Number and unit corruption. Engines have been observed to transpose digits, change a decimal, or silently convert or fail to convert a unit. A dosage of 2.5 mg can surface as 25 mg. A confident, clean sentence wrapped around a wrong number is the number error you will not catch by reading for flow.
• Terminology drift. The client approved one specific term for their device or feature, recorded in the termbase (the controlled glossary of approved terms). The engine, trained on general text, "prefers" a more common synonym and quietly substitutes it, segment after segment, propagating the wrong term across the whole file.
• Omission. A clause, a qualifier, a whole phrase can simply not appear in the target, because the engine found a fluent path that skipped it. Nothing tells you a word was lost. The sentence is complete and grammatical; it is just missing part of the source's meaning.
• Hallucination on thin input. Fed a fragment, a heading, a list item with no context, or garbled source, an NMT engine can generate plausible target text that corresponds to nothing in the source at all. It abhors a vacuum and fills it with fluent invention.

Notice the through-line: every one of these failures produces output that reads correctly. That is not a coincidence. The engine is built to produce fluent text. Fluent text is its entire competence. Accuracy is a separate property that the engine approximates but does not verify, and the cases where fluency and accuracy diverge are precisely the cases where the engine has no idea anything is wrong.

What an LLM Actually Does, and Why It Is More Dangerous

An LLM translating a paragraph is doing something subtly but importantly different from an NMT engine. The NMT engine was trained on the translation task and only that task. The LLM was trained to continue text plausibly across every domain of human writing, and translation is a side effect of that general competence. When you hand an LLM a source segment, it is not running a dedicated translation routine. It is predicting what plausible text comes next given a prompt that, in effect, asks for a translation. The difference sounds academic. In the file, it is the difference between a tool that drifts and a tool that improvises.

Because the LLM has absorbed an enormous range of human writing, it is often more fluent than a narrow NMT engine, more sensitive to register, idiom, and surrounding context, better at making a sentence sound like a native speaker wrote it. This is genuinely useful, and it is exactly why LLMs are now woven into localization workflows for drafting, for stylistic polish, for handling context that a sentence-by-sentence NMT engine misses. But the same generative freedom that produces beautiful prose produces a more confident and more creative class of error. An LLM is more willing to "fix" a source it finds confusing by guessing your intent. It is more willing to add a clause that "should" be there. It is more willing to produce a fluent, authoritative-sounding rendering of something it has effectively made up. The industry shorthand for this is the same word borrowed from the NMT world, hallucination, but with an LLM the hallucinations tend to be larger, smoother, and more persuasive.

An NMT engine drifts off the source. An LLM improvises around it. The smoother the prose, the harder the error is to see, and the LLM is the smoother machine.

The Medical Evidence: Fluent and Wrong

This is not a theoretical worry. When researchers measured LLM output on medical content, the error rates were not small and they were not in the grammar. Studies of LLM-generated medical content found error rates of roughly 59% on drug names, 60% on dates and times, and 66% on adverse events. Read those numbers again with the costume in mind: every one of those errors arrived in grammatically perfect prose. The model named a wrong drug fluently. It stated a wrong time fluently. It described an adverse event, the documented harm a treatment can cause, incorrectly and fluently. There was no asterisk, no hedge, no change in tone between the correct outputs and the catastrophic ones. The thing that makes the error a Critical error in a localization sense, an error severe enough to fail the entire file, is not its rarity. It is its invisibility. A clumsy mistranslation trips your eye and you fix it. A fluent one slides past, gets accepted, gets shipped, and surfaces as harm.

Generative Drafting Is a Third Job, Not Translation

There is a further distinction that working linguists must hold separate, because clients constantly blur it. Using an LLM to translate a specific source segment is one task: there is a source, there is a right answer, and faithfulness to that source is the measure. Using an LLM to generate content, to draft marketing copy from a brief, to summarize, to expand a bullet into a paragraph, to invent a tagline, is a completely different task with no source to be faithful to. That is generative drafting, and it is creative writing assistance, not translation. The danger comes when the two get mixed in one workflow. An engine asked to "translate and make it punchier" is being asked to be faithful and inventive in the same breath, and it will resolve the tension by inventing. The moment a request slides from "render this source" to "improve this content," the standard against which you judge the output changes from accuracy to appropriateness, and the failure modes change with it. Keeping translation, LLM-assisted translation, and generative drafting in three separate mental boxes is not pedantry. It is how you know what question to ask of the output: "Is this faithful?" or "Is this good content?" Those are not the same question, and answering the wrong one is how errors ship.

Killing the "Translates as Well as a Human" Framing

Now we can take apart the slogan directly, because we have the pieces. "The machine translates as well as a human" treats translation as a single skill measured on a single axis, where the machine has now caught up to or passed the human. That framing is built on a category error. Human translation and machine translation are not two scores on the same test. They are two different processes that happen to produce text in the same target language.

A human translator reads the source, understands what it means, knows who it is for, recognizes when a sentence is a contraindication and that contraindications must never be inverted, notices when a number looks wrong, remembers that this client approved a specific term, and is accountable for the result. A machine produces the most probable target text given its training and the input, with no model of meaning, no awareness of consequence, no memory of the client's approved term unless engineered to have one, and no accountability. When the machine's most-probable output coincides with the correct translation, which is often, the outputs look equivalent. The equivalence is in the artifact, not in the process, and the process is what protects you when the stakes are real.

What "Fluent, Not Correct" Means Operationally

The core discipline of this entire field can be compressed into four words: fluent is not correct. Fluency is a property of the prose. It is how natural, grammatical, and confident the target text sounds. Correctness is a relationship between the target text and two external things: the meaning of the source segment, and the approved terminology and rules for this content. A machine optimizes hard for fluency, because fluency is what its training rewards. Correctness is not something the machine can check, because checking it would require comparing the output back against the source's meaning and the client's rules, which the generation process does not do.

This is why the fluent error is the dangerous one and the clumsy error is the safe one. A clumsy mistranslation is self-announcing: it reads wrong, your eye catches on it, you stop and fix it. A fluent mistranslation is camouflaged in the exact quality, smooth prose, that you have been trained your whole career to read as a sign of competence. The engine's greatest strength, fluency, is precisely the thing that hides its most expensive failures. The pacemaker sentence was not dangerous despite being well-written. It was dangerous because it was well-written.

Fluency is a property of the prose. Correctness is a relationship to the source and the approved terms. The machine guarantees the first and cannot verify the second.

MT-First Is the 2026 Baseline, Not a Pilot

If you are reading this expecting a debate about whether AI is "coming" to localization, you are a few years late. It is not coming. It arrived, it is in your CAT tool, and it filled your segments before you opened the file. The professional question is no longer whether to use MT. It is how to own quality on top of an MT-first pipeline that already exists. The numbers make this concrete.

Machine-translation post-editing (MTPE), the workflow in which a human edits machine output rather than translating from scratch, has crossed from niche to default. According to Nimdzi industry research, MTPE adoption rose from 26% in 2022 to roughly 46% in 2024, and 81.1% of language-service providers (LSPs) now offer MTPE. Roughly 70% offer subtitling, much of it now machine-first. Post-editing, abbreviated PE, is no longer a specialty line item. It is the shape of the work.

The economics drove the shift and explain the pressure you feel. MTPE typically prices at 50 to 75% of full human translation, around $0.05 to $0.15 per word, with light post-editing sometimes as low as $0.02 per word. A hybrid MT-plus-human workflow lifts a linguist's throughput from roughly 2,000 words a day to 5,000 or more. From the client's chair, that looks like cheaper, faster translation. From your chair, it looks like a file that arrives pre-filled, a deadline that assumes the machine already did the thinking, and a per-word rate that assumes you are correcting rather than creating. The temptation built into that arrangement is to trust the clean segments and move fast. The whole point of this program is that trusting the clean segments is exactly the trap, because the clean segments are where the fluent error hides.

What MT-First Changes About Your Job

When every segment is pre-populated, your work inverts. You are no longer primarily a producer of target text. You are primarily a verifier and quality owner of machine-produced target text, and the rare creator of the segments the machine could not handle. That is a different skill, and in many ways a harder one. Producing a translation engages your active competence; you build the sentence and you know what you put in it. Verifying a translation engages a more vigilant competence: you must catch what is wrong in text that someone, or something, else produced fluently, under time pressure, on a per-word rate that did not budget for deep scrutiny. The engine handed you speed. It also handed you a new and more dangerous failure mode and called it your responsibility. Recognizing that trade is the first act of the AI-aware linguist.

When Each Engine Is the Right Tool

None of this is an argument against MT or LLMs. It is an argument for matching the engine to the job and the job to the consequence. Used with their natures understood, each of these machines earns its place.

Where Narrow NMT Shines

Neural MT is the workhorse for high-volume, repetitive, lower-creativity content where the source is clean and the consequence of an error is recoverable: product catalogs, support articles, internal documentation, user-generated content, the bulk of a large software product's strings. The engine is fast, cheap, and its failure modes are known and checkable. Wrapped in a proper post-editing workflow with a verification step, NMT is genuinely transformative for throughput. The mistake is not using it. The mistake is using it on content where its failure modes are unrecoverable, or skipping the verification step that converts its speed into safety.

Where LLMs Shine

LLMs earn their keep where context, register, and fluency matter more than narrow sentence-by-sentence fidelity, and where a human is going to own the output anyway. They are strong at drafting, at adapting tone, at handling longer context that a sentence-bound NMT engine fragments, at suggesting terminology candidates for a human to approve, at first-pass transcreation that a human then shapes. Their generative freedom, dangerous in a contraindication, is an asset in a tagline. The skill is using that freedom where invention is welcome and locking it down where invention is a liability.

Where the Machine Must Not Touch

And some content the machine must never touch on its own, no matter how good it has become: drug labels and clinical instructions where a flipped negation or a corrupted dosage is a death, legal clauses where an inverted obligation or a dropped "not" is a lawsuit, financial disclosures, life-safety warnings. This is the principle of risk-tiered intake: you classify content by consequence before a single segment is post-edited, and high-liability content gets full human translation or full, careful post-editing by a fully qualified linguist, never a cheap light pass and a trusting keystroke. The revised post-editing standard, ISO 18587, codifies this thinking by insisting that the post-editor hold the same professional competence as a translator, precisely because catching the silent fluent error in high-stakes content is a translator's job, not a button-pusher's. Knowing which content the machine is forbidden to touch is not a limitation on your use of AI. It is the most senior judgment you bring to it.

Match the engine to the job and the job to the consequence. The most valuable thing you know about an engine is the content it must never be allowed to translate alone.

Key Takeaways

• "AI can translate" collapses three distinct machines into one word. Machine translation (MT) is the umbrella; neural MT (NMT) is a narrow engine trained only to translate; a large language model (LLM) is a general text predictor that translates as a side effect. Generative drafting is a fourth, separate job with no source to be faithful to. Name the machine before you trust, price, or route the output.
• All three engines produce the same artifact, fluent target text, so the output is indistinguishable to the eye. The costume is fluency; what differs is faithfulness, behavior under uncertainty, and cost. A linguist who cannot see past the costume trusts blind.
• An NMT engine optimizes for probable, fluent target text and only approximates faithfulness, so it fails in characteristic ways: dropped negations, corrupted numbers, terminology drift, silent omissions, and hallucination on thin input. Every one of those failures reads correctly.
• An LLM is often more fluent and context-aware than narrow NMT, but its generative freedom makes its errors larger, smoother, and more persuasive. It improvises around a confusing source rather than drifting off it, which makes its hallucinations harder to catch.
• Fluent is not correct. Fluency is a property of the prose; correctness is a relationship between the output, the source segment's meaning, and the approved terminology. The machine guarantees fluency and cannot verify correctness, which is why the fluent error is the dangerous one.
• The evidence is concrete: LLM medical-content error rates of roughly 59% on drug names, 60% on dates and times, and 66% on adverse events, every error delivered in grammatically perfect prose with no warning marker.
• MT-first is the 2026 baseline, not a pilot. MTPE adoption rose from 26% in 2022 to about 46% in 2024, 81.1% of LSPs now offer it, it prices at 50 to 75% of full human translation, and it lifts throughput from about 2,000 to 5,000-plus words a day. Your job inverts from producing target text to owning the quality of machine-produced target text.
• Match the engine to the job and the job to the consequence. NMT for high-volume recoverable content, LLMs where context and register matter and a human owns the output, and full human translation or full qualified post-editing for high-liability content the machine must never translate alone.