This is the opening article in a series about how I combine human translation with controlled AI review. I’m writing mainly for translators who don’t want to outsource their judgment to a black‑box tool, and for organizations that care more about process and reliability than “magic prompts”.
In this first article, I want to answer a simple question before we dive into scripts and workflows: why do I use AI almost only for review, not to produce the first draft?
My baseline: human first, tools as helpers
On a typical project, I follow a familiar workflow:
I translate segment by segment in a CAT tool (usually Trados or memoQ).
I draw on glossaries I’ve built over the years, and sometimes glossaries provided by the client.
I revise each segment in the CAT tool as I go.
At the end, I run a QA pass, usually with Xbench, to catch inconsistent translations and errors in spelling, numbers, tags, terminology, and similar issues.
AI comes in after the first human draft. I send selected segments or paragraphs to an AI assistant (Perplexity Pro) with a narrow brief: check terminology and register, suggest improvements where needed, but say “no change necessary” if the translation is already correct. I’m not asking it to invent content; I’m asking it to challenge my choices.
That distinction—AI for drafting vs. AI for review—is the point of this article.
What AI is actually doing to translation work
AI and MT have already taken over the bottom of the translation market and are chewing through the middle. Much low‑end work has either disappeared or returned as “human in the loop” post‑editing, often at rates that don’t reflect the risk.
At the same time, many translators are using the same tools to improve quality and consistency, especially on specialized material. There’s a big difference between:
“Let’s have AI churn out a first draft and you do a quick skim”, and
“You produce a solid human draft, then use AI to check terminology and style in a structured way.”
This series is about the second scenario.
Where AI actually helps in review
In one recent project, I translated Italian‑into‑English engineering and chemistry course descriptions written in the 1980s, for an application being submitted now. The source reflected older terminology; the translation had to match current usage.
One syllabus section included:
Parte descrittiva: idrogeno; metalli alcalini; elementi del terzo gruppo; elementi del quarto gruppo; elementi del quinto gruppo; elementi del sesto gruppo; alogeni; elementi di transizione; chimica organica.
My first English draft was:
Descriptive section: Hydrogen; alkali metals; third‑group elements; fourth‑group elements; fifth‑group elements; sixth‑group elements; halogens; transition elements; organic chemistry.
With a “check and only suggest changes if needed” prompt, AI proposed:
Descriptive section: Hydrogen; alkali metals; group 13 elements; group 14 elements; group 15 elements; group 16 elements; halogens; transition elements; organic chemistry.
On the face of it, that’s a bold change: terzo gruppo becomes “group 13”. If the Italian says “third group”, “group 3” is the obvious reading.
The key step was to look at the surrounding content. The same section goes on to discuss:
third‑group elements and the industrial production of aluminium
fourth‑group elements, focusing on carbon and silicon
fifth‑group elements, then nitrogen, phosphorus, nitric and phosphoric acids, fertilizers
sixth‑group elements, then oxygen, sulfur, sulfur oxides, sulfuric acid
Aligned with modern periodic‑table families, this clearly follows an older main‑group convention: aluminium family → Group 13, carbon/silicon → Group 14, nitrogen/phosphorus → Group 15, oxygen/sulfur → Group 16.
In this context, “group 13/14/15/16 elements” is the right modern phrasing in English. AI’s suggestion pointed in the correct direction, but I still had to read the syllabus, know the chemistry, and confirm the mapping against current references. The useful part was speed: Perplexity also surfaced relevant reference material, so checking the group numbers took minutes rather than a small research session.
A second example from the same materials is simpler. One course title read Automazione e Regolazione. My first draft was “Automation and Regulation”, which is literal but slightly off in engineering. When I asked AI to review the title in context (“US English, engineering university course titles”), it suggested “Automation and Control” instead and noted that in control engineering regolazione here is about automatic control systems and control theory, and “Control” is the standard term.
It’s a small change, but it makes the course title sound like something an engineer would actually write today, not a direct echo of the Italian.
Both examples show the pattern. I draft the translation, then use AI to ask: “Is this how a current English‑language textbook or course catalog would phrase it?” Sometimes the answer is “no change necessary”, which is helpful in itself. Sometimes I get a better term (“Control” instead of “Regulation”). Sometimes I get a suggestion that only holds up once you check it against the underlying discipline. In all cases, I decide what goes into the final draft; AI just helps me interrogate my own work.
Why I don’t use AI for the first draft
If AI can do all that in review, why not let it produce the draft and just “fix things up” afterwards?
I resist that for a few reasons:
Hallucinations and smooth nonsense
Modern systems can produce fluent, plausible text, but they’re not good
at signalling uncertainty. In technical or academic work, that’s risky.
I prefer to start from a translation whose meaning I know, rather than
from a confident text that may be wrong in subtle ways.
Terminology drift and inconsistency
Over a long document, AI can drift in terminology, use different phrases
for the same concept, or shift definitions. Cleaning that up after the
fact is often harder than keeping terms under control while writing.
The “expert as janitor” problem
“We’ll have AI do the first draft and you just review it” usually means
“take on the liability of spotting errors, but at a discount”. An
AI‑first draft often has a higher error rate than a careful human draft,
and responsible review takes time. It’s not quick; it’s just different
work.
Control over voice and argument
In some content, structure and tone are part of the meaning. If I let a
tool produce the first version, I still have to rebuild the argument,
nuance, and hedging afterwards.
In short: I’d rather think through the text once as I write it, then use AI to check and refine, than spend the same or more time trying to infer what a system “meant”.
How I keep AI in its lane
A lot of this comes down to how you frame the task.
For review, my prompts usually say, in effect:
Focus on terminology, register, and clarity.
Suggest changes only where something is incorrect, unclear, or clearly suboptimal.
If the translation is already correct, respond with “No change necessary.”
If you’re uncertain about a term, say so and give alternatives.
This reduces noise: I don’t want the tool rewriting clean sentences. It also makes it clearer which suggestions are genuine corrections vs. preferences.
There are times when I “spar” with the system. It suggests something that doesn’t quite fit; I push back, adjust, and sometimes end up with a third option that’s better than either my original or its first attempt. But the direction is clear: I have the brief and the responsibility. AI is there to catch blind spots and propose options.
I’ll dig into prompt design in a later article. For now, the important point is that the prompt mirrors the workflow: human first, AI second, human last.
Practical takeaway (for colleagues and clients)
If you’re a translator, before asking “How can I get AI to translate this for me?”, try “How could AI help me review my own translation more effectively?” Start human, then use the tool to:
check that your terminology matches current usage in the field,
challenge titles, headings, and boilerplate that may have aged, and
reach reference material quickly when something looks like an old convention.
If you’re a client or project manager, the safer setup for serious content is still human translation plus AI‑assisted review, not AI‑first with “quick human post‑editing”. You want someone who understands both the domain and the tools, and who uses AI to support their judgment, not replace it.
In the next article, I’ll look at why subject‑matter expertise is still non‑negotiable in an AI world. Without that, AI review quickly turns into curating style instead of checking meaning.
No comments:
Post a Comment
Thank you for your comment!
Unfortunately, comment spam has grown to the point that all comments need to be moderated. All legitimate comments will be published as soon as possible.