Aiko has a blinking cursor and a problem. On her screen is a 600-word product launch announcement from a German consumer-electronics client, written in confident, idiomatic German, due in Brazilian Portuguese by tomorrow morning. The translation memory is almost no help; this is fresh marketing copy, and the TM (translation memory, the database of previously approved source-and-target pairs) returns nothing above a 40% fuzzy match. The termbase is thin. The source has three puns, a headline that rhymes, and a paragraph of warm, slightly playful brand voice she has spent two years learning to reproduce. The deadline assumes the speed of a machine. The content assumes the judgment of a human. For ten minutes she stares at the blank target column, the most expensive ten minutes of her day, the dreaded white space where a translator's mind churns through openings, discards them, starts again. Then she opens a second window, pastes the German into an LLM (large language model, a general text-prediction system that translates and rewrites as a side effect of its training), and types: "Give me five different Portuguese openings for this headline, each with a slightly different tone." Within seconds she has five options on the screen. None of them is her final answer. Every one of them is a door out of the blank page. This lesson is about exactly that moment: where AI genuinely helps a linguist generate a first draft and a spread of options, how to use that help without quietly surrendering authorship and accuracy, and where the line sits between draft-assist that makes you faster and unverified output that makes you liable. The speed can be yours. The judgment has to stay yours too.

The Blank Page Is the Real Enemy

To see why AI-assisted drafting is genuinely useful and not just a trendy shortcut, you have to be honest about where a translator's time actually goes. It does not all go into typing. A great deal of it goes into the silent, invisible work of starting: holding the source meaning in your head, considering an opening, rejecting it as too literal, considering another, rejecting it as off-tone, and slowly converging on a phrasing that feels right. That convergence is real cognitive labor, and on creative or unprecedented content it can eat more minutes than the typing ever will. The blank page is not empty time. It is expensive time, and it is the part of the work where an LLM can legitimately give you a lift.

This is a different use of AI than the one most of this program has dwelt on so far. In post-editing, the machine has already filled every cell, and your job is to distrust the smooth surface and verify it against the source. Here the cells are blank, the TM is silent, and you are reaching for the machine on purpose, not inheriting its output by default. The posture is different. You are not a reluctant verifier of something a TMS (translation-management system) pre-populated before you arrived; you are a working linguist who has decided, deliberately, that generating a few options fast is better than generating zero options slowly. That deliberateness is the whole game. The same model produces draft-assist when you own it and unverified output when you do not, and the only difference between those two outcomes is what you do after the text appears.

Consider what Aiko actually gains from those five openings. Not one of them will ship as written; she knows that before she reads them. What she gains is momentum and contrast. Seeing five attempts at the headline, even mediocre ones, surfaces the dimensions of the choice: this one is too formal for the brand, this one lands the rhythm but loses the pun, this one keeps the playfulness but drifts from the literal meaning, this one is close but uses a register that reads as European Portuguese rather than Brazilian. The options are not answers. They are a map of the trade-off space, drawn in seconds, that she would otherwise have had to draw slowly in her own head. The blank page gave her nothing to react to. The five drafts give her something to react to, and reacting is far faster than originating.

An AI first draft is not a translation you accept. It is a starting position you argue with. Its value is that it gives your judgment something to push against, faster than the blank page ever could.

Where the Lift Is Real, and Where It Is Imaginary

The lift is real on a specific kind of work and imaginary on another, and confusing the two is how linguists either underuse the tool or get burned by it. The lift is real when the bottleneck is starting, when you have many viable ways to render something and the cost is choosing among them. Marketing copy, internal communications, UI microcopy with several plausible phrasings, a paragraph where you know the meaning cold but cannot find the opening: these are option-generation problems, and an LLM is a fast option generator. It is genuinely good at producing variety on demand, at rephrasing the same idea in five registers, at breaking a logjam.

The lift is imaginary, and the risk severe, when the bottleneck is not starting but knowing. When the source contains a dosage, a contraindication, an indemnity clause, a regulatory claim, a number that must be exact, the hard part was never finding a phrasing. The hard part is being correct, and the model cannot be correct on your behalf, because it does not know your source's truth, it only knows what sentences tend to follow other sentences. Asking an LLM for a first draft of a drug interaction warning is not draft-assist. It is generating a fluent, confident, possibly inverted claim that you will now have to verify so thoroughly that the model saved you nothing and exposed you to the exact failure mode this whole program is built to prevent: the silent critical error that reads perfectly and means the opposite. The decision of whether to draft with AI at all is therefore the same risk-tiering decision you make everywhere else in the pipeline. On low-consequence, high-variability content, draft freely. On high-consequence content, the engine drafts nothing you would not have to rebuild from scratch anyway.

Generating Options Without Surrendering Authorship

Authorship is the word that matters most in this lesson, and it is worth defining precisely, because it is the thing you can lose without noticing. To be the author of a translation is to be able to say, of every meaningful choice in the target, why it is there: this word because the source meant that, this register because the brand demands it, this structure because the target language reads better this way, this term because the termbase approved it. Authorship is not about typing every character yourself. A translator who dictates to a typist is still the author. Authorship is about owning the reasons. You surrender it not when you let a machine produce text, but when you ship a choice you cannot explain, a phrasing you accepted because it looked fine rather than because you decided it was right.

This distinction dissolves the false dilemma that paralyzes a lot of careful linguists. The fear is that using AI to draft means the work is no longer "yours," that accepting a machine's sentence is a kind of fraud. But that fear misplaces where authorship lives. The machine that suggests an opening is no different in principle from a thesaurus that suggests a synonym, a colleague who offers a phrasing over coffee, or a fuzzy match that proposes a stored target. None of those sources writes your translation for you. They feed your judgment. What makes the output yours is that you took it through your judgment and emerged able to defend it. What would make it not yours, and dangerous, is taking it around your judgment, accepting it because the screen was full and the deadline was close.

The Three Things You Must Do to Every Generated Option

Owning a generated draft is not a vague attitude. It is three concrete operations performed on every option before it earns a place in your delivery, and they are the same three operations a disciplined post-editor performs on a pre-filled cell, applied here on purpose rather than by inheritance.

• Check it against the source, not against your taste. The first question is never "do I like this?" It is "does this mean what the German meant?" An LLM generating options will happily produce five fluent sentences, one of which quietly drops a qualifier, softens a claim, or invents a benefit the source never stated. The option that reads best is not exempt from this check; it is the one most likely to slip an alteration past you, because its fluency disarms suspicion. Every generated option is a claim about the source that you have to confirm, exactly as you would confirm a raw MT (machine translation) segment.
• Check it against the constraints that are not in the prompt. The termbase, the locale conventions, the length budget, the style guide, the placeholder integrity: the model does not know these unless you told it, and even when you tell it, it drifts. A generated headline might be lovely and three characters too long for the banner. A generated paragraph might use a perfectly good synonym for a term the client has mandated. Variety from the model is variety against the model's defaults, not against your client's rules, and the rules are yours to enforce.
• Decide, and be able to say why. The final operation is the one that restores authorship: you pick, or you blend, or you rewrite from the options into something none of them quite was, and you hold a reason for the choice. If you cannot articulate why the shipped version beats the four you rejected, you have not authored it. You have selected it, which is a weaker relationship to the text and a weaker thing to put your name on. Selection is what the model does. Decision, with a defensible reason, is what you do.

Run those three operations and the generated draft becomes yours in the only sense that matters professionally: you can defend every choice in it to a reviewer, a client, or an auditor, and "the engine wrote it" never has to be your answer, because the engine did not write it. It proposed, and you decided.

Aiko's Launch Announcement: A Worked Example

Let us follow Aiko all the way through, because the abstractions only mean something when you watch them touch a real file. Her source paragraph, in German, says something like: the new earbuds are so light you forget you are wearing them, the battery lasts a full working day, and the playful tagline puns on the German word for "ear" in a way that does not exist in Portuguese. Three challenges live here: a sensory claim that must stay accurate, a battery specification that must stay exact, and a pun that cannot survive a literal crossing.

Prompting for Options, Not for an Answer

Aiko does not type "translate this." That phrasing invites the model to hand her a single confident answer she will be tempted to accept whole, and a single answer is the opposite of what draft-assist is for. Instead she frames the request as option generation, and she loads the prompt with the constraints she already knows: the target is Brazilian Portuguese, not European; the register is warm and playful but not slangy; the brand never makes superlative claims it cannot back up; the battery figure is exactly one working day and must not be rounded, inflated, or dropped. She asks for several distinct renderings of each tricky element and, crucially, she asks the model to flag anything in the source it found ambiguous or could not render faithfully. That last instruction turns the model from a confident answer-machine into something closer to a junior colleague who says "I was not sure about this part," which is exactly the signal a verifier wants.

The model returns. For the sensory claim it offers a few options, all fluent. One of them, the prettiest, says the earbuds are "so light you will forget they exist." That is a small drift: the source said you forget you are wearing them, a claim about comfort during use, and the model's version says you forget they exist, a slightly different and stronger claim. A tired linguist accepts it because it sings. Aiko catches it because she is running operation one, checking against the source rather than against her ear, and the prettiest option is precisely the one she scrutinizes hardest. She keeps the structure she likes and corrects the meaning back to the source. The option gave her a good rhythm; her judgment kept the rhythm honest.

The Number the Model Must Not Touch Unwatched

For the battery specification, Aiko treats the model's output as untrusted by default, because a number is exactly the high-consequence, easily-corrupted element that machine output mangles in characteristic ways: a "full working day" can become "a full day," which is a different and larger claim, or "up to a day," which is a weaker hedge the brand did not authorize, or the figure can attach to the wrong feature entirely. She does not let the generated phrasing carry the number unexamined. She lifts the number out, confirms it against the source, and rebuilds the sentence around the verified figure rather than around the model's fluent guess. This is the discipline that separates draft-assist from liability: the model may shape the prose around a fact, but it never gets to be the source of the fact. The fact comes from the source segment, verified by the human, every time.

Let the engine shape the sentence. Never let it be the source of a number, a name, a negation, or a claim. The facts come from the source, verified by you; the model only helps with the words around them.

The Pun: Where the Engine Helps by Failing Out Loud

The pun is the most interesting case, because here the model genuinely cannot do the job and the value comes from a different direction. There is no faithful Portuguese rendering of a German ear-pun; the wordplay does not exist across the language pair. This is transcreation territory, the creative recasting where brand, culture, and intent require a human, and the next lesson in this chapter dwells on exactly why the engine cannot do it. But the LLM is still useful here, not as a translator but as a brainstorming partner. Aiko asks it for ten different Portuguese taglines that capture the playful spirit and the lightness theme without attempting the original pun, explicitly telling it to invent fresh wordplay native to Portuguese rather than translate the German joke.

The ten options come back. Eight are flat. One leans on a Portuguese idiom about feathers and lightness that is genuinely charming. One accidentally produces a double meaning that would read as slightly crude in Brazilian Portuguese, the kind of cultural landmine the model has no awareness of and Aiko catches instantly because she lives in the locale and it does not. She discards nine, takes the feather idiom as a seed, and crafts her own tagline that the model never wrote but helped her find. The engine did not transcreate. It generated raw material, including one option that would have embarrassed the brand, and her cultural judgment did the transcreation by selecting, rejecting, and building. The pun is a perfect illustration of the line: the model can stir the pot of options, but the human owns which option is safe, on-brand, and right, and the human is the only party in the room who could have known the crude reading was a problem.

The Line Between Draft-Assist and Unverified Output

Everything in this lesson turns on one line, and it is worth drawing it as sharply as possible, because a great deal of professional risk lives in the blur around it. Draft-assist and unverified output can be the literal same text from the literal same model. The difference is not in the words on the screen. The difference is entirely in what happened to those words between generation and delivery.

Output is draft-assist when it passed through a human's judgment and emerged owned: checked against the source, checked against the constraints, decided upon with a defensible reason, and now defensible as the linguist's own choice. The model accelerated the linguist; it did not replace the linguist's judgment, and if every trace of the model vanished, the linguist could still stand behind every word. Output is unverified output when it traveled around the judgment instead of through it: accepted because it read well, shipped because the deadline was close, delivered with a quiet hope that the fluent surface meant a sound interior. The danger of unverified output is not that it is always wrong. It is that you do not know whether it is wrong, and you have put your name on it anyway. You have converted a tool that should make you faster into a mechanism that makes you liable, and the conversion happened silently, in the gap where a verification step should have been and was not.

Why the Line Is Easy to Cross by Accident

Nobody decides to ship unverified output. They drift across the line under the same three pressures that drive the trust trap in post-editing, and naming the pressures is how you resist them. The first is fluency as a false signal: a generated draft reads beautifully, and your career-long instinct says beautiful prose is competent prose, an instinct that was reliable when only a competent human could produce beautiful prose and is now a liability because a model produces beautiful prose whether or not it understood anything. The second is the deadline that assumed machine speed: when the rate and the schedule are priced as if the machine did the thinking, every minute spent verifying feels like a minute you are not being paid for, and the cheapest action is to trust the surface. The third is the seductive completeness of a finished-looking draft: a blank page demands work, but a full page invites acceptance, and a full page produced in two seconds invites it most of all, because the effort that should have gone into authoring was never spent and its absence does not show.

The defense is not heroic vigilance, which fails under volume and fatigue exactly when you need it. The defense is to make the verification step non-optional and ritual, the same way a good post-editor checks numbers and negations on every segment by design rather than by suspicion. You decide, before you generate anything, that no generated text ships without passing the three operations, and you treat that rule as inviolable as the one-critical-fails delivery gate. The model is allowed to give you a running start. It is not allowed to give you a finished product, because it cannot, and the moment you treat its output as finished is the moment you stop being the author and become the liable party for something you never actually reviewed.

Idea Generation Beyond the First Draft

First drafts are the headline use, but the model's option-generation talent helps a working linguist in several quieter ways across a project, and they share the same discipline: generate freely, own the result.

When you are stuck on a single stubborn segment, the kind that has you staring for five minutes, asking the model for several alternative renderings can break the logjam the way a colleague's offhand suggestion does. You are not asking it to be right. You are asking it to be a wall to bounce off, and even a wrong option can reveal the shape of the right one by contrast. The verification discipline is unchanged: the option that frees your thinking still has to survive the source check before it ships.

When you are hunting for synonyms or register variants, the model is a fast, context-aware thesaurus that proposes a word in the actual sentence rather than in a dictionary's vacuum. This is genuinely useful and genuinely low-risk, because you are using your own linguistic competence to judge whether the proposed word fits, which is exactly the competence the ISO 18587 revision insists a post-editor hold: the full competence of a professional translator, applied to machine-proposed material. The model widens your shortlist; you remain the one who knows which word is right for this brand, this locale, this sentence.

When you are untangling an ambiguous source, you can ask the model to lay out the possible readings of a confusing sentence, not to pick one, but to make the ambiguity explicit so you can resolve it with knowledge the model lacks: the document's purpose, the client's prior decisions, a query you can send back to the source author. Here the model is a clarifier of the problem, not a producer of the answer, and that is one of its safest and most underrated uses. It is least dangerous precisely when you ask it to surface a question rather than assert a fact.

What You Still Have to Bring

Notice what every one of these uses requires from you and only you. The source truth, which the model approximates but does not know. The locale lived experience that catches the crude double meaning, the wrong-region register, the date format that would confuse a Brazilian reader. The brand voice you learned over two years and the model learned never. The client's specific approved terms and prior decisions. The judgment about which content is too consequential to draft with a machine at all. The model brings speed and variety. You bring everything that makes a translation correct, safe, on-brand, and defensible, and that is not a small remainder left over after automation. It is the entire substance of the profession, now concentrated onto the part the machine genuinely cannot do, with the drudgery of the blank page handed off to the part it genuinely can.

This is the same movement the whole program describes, seen from the drafting chair rather than the post-editing one: the engine takes the mechanical production of plausible text, and the human's contribution concentrates onto pure judgment. A linguist who uses AI to generate options and then owns every one of them is not a button-pusher who got cheaper. They are a faster author who kept the authorship. A linguist who generates options and ships them unverified is not faster. They are exposed, and they have traded the one thing that made them valuable, defensible judgment, for the one thing that makes them replaceable, raw throughput on text nobody verified. The tool is identical in both cases. The professional is not.

Key Takeaways

• The blank page is expensive cognitive labor, not empty time, and an LLM (large language model) is genuinely useful as a fast generator of options that gives your judgment something to push against, faster than originating from scratch. An AI first draft is a starting position you argue with, never a translation you accept.
• The lift is real when the bottleneck is starting (marketing copy, microcopy, creative content with many viable phrasings) and imaginary, even dangerous, when the bottleneck is knowing (dosages, contraindications, indemnity clauses, exact numbers), where the model cannot be correct on your behalf because it predicts likely text rather than knowing your source's truth.
• Authorship is owning the reasons for every meaningful choice, not typing every character. You keep it by taking generated text through your judgment and emerging able to defend it; you surrender it by shipping a choice you cannot explain because the screen was full and the deadline was close.
• Own every generated option with three operations: check it against the source rather than against your taste, check it against the constraints not in the prompt (termbase, locale, length budget, style guide, placeholders), and decide with a reason you can articulate. Selection is what the model does; decision with a defensible reason is what you do.
• Let the engine shape the sentence, but never let it be the source of a number, a name, a negation, or a claim. Facts come from the source segment, verified by the human, every time; the model only helps with the words around the verified fact.
• The prettiest generated option deserves the hardest scrutiny, because its fluency disarms suspicion and it is the one most likely to slip a softened claim, a dropped qualifier, or an invented benefit past you, exactly like a raw MT (machine translation) segment.
• Draft-assist and unverified output can be the identical text from the identical model; the only difference is whether it passed through your judgment or around it. Output is draft-assist when it is checked, owned, and defensible; it is unverified output when it was accepted because it read well, and the conversion to liability happens silently in the gap where verification should have been.
• Make verification non-optional and ritual rather than relying on heroic vigilance, which fails under volume and fatigue. Beyond first drafts, the model also helps break logjams on stubborn segments, suggest context-aware synonyms, and surface a source's ambiguity, and it is safest when you ask it to raise a question rather than assert a fact, while you supply the source truth, locale experience, brand voice, and risk judgment it cannot.