1. The “AI will cover my weak spots” story
An insidiously comforting story is doing the rounds: powerful AI has done away with the gap between specialist and non‑specialist translators. In this story, subject‑matter expertise quietly moves from a must to an afterthought: you can work with inexperienced (and conveniently cheap) translators, run everything through an MTPE process, or even exclude translators altogether and ask monolingual staff to confirm that the AI’s output looks respectable enough to ship.
The promise is simple: if the machine “knows” chemistry, finance, or property law, then the only need for a professional is as a token “human in the loop”, someone to tidy up the target language, accept most suggestions, and – most importantly – take the blame if the project goes wrong. It’s a reassuring picture for buyers under pressure to cut costs. The problem is that it bears little resemblance to how translation decisions should be taken in high‑risk situations or for specialized content.
In my workflow – the one I outlined in the first article of this series – AI only comes in after a human draft is complete. I translate first in a CAT tool, using glossaries and whatever domain knowledge I’ve built up; I revise and run proper QA.
Only then do I ask an AI assistant to review terminology, register, and style, with a strict “no change if the translation is already correct” constraint. AI can suggest alternative phrasings and highlight relevant references, but it can’t decide what’s appropriate for a specific audience, legal system, or communicative purpose. That judgment still depends on subject‑matter expertise.
2. Where AI fits
The key point of my first article was simple: AI is helpful in review because it can propose alternative translations, catch possible mistranslations, or suggest options I hadn’t thought of. As a primary translator, though, it’s dangerous: it will always produce a translation, but there is no guarantee that what it offers is correct rather than an unthinking literal translation or a smooth, fluent hallucination. That’s why I have two basic rules in my prompts for the review step: if the segment is already correct, the AI should say so and not suggest changes; and whenever it suggests a change, it should explain why.
The chemistry example from Article 1 – where “third/fourth/fifth/sixth group” became “group 13/14/15/16 elements” – worked because I knew enough about the subject to question the suggestion, and could then find evidence that it made sense in context.
This article picks up that thread. If AI is going to review our work, we still need the subject‑matter knowledge to decide when its suggestions are helpful, not merely plausible.
To give a concrete example, let’s look at a term that appears all the time in Italian property documents: catasto.
3. When the “correct” term isn’t clear
3.1 From “cadastre” to “land registry”
In Italian real‑estate documents, the words catasto and catastale crop up everywhere: contracts for the purchase, sale, or lease of property, and notarial deeds that formalize those transactions. They refer to the Italian land registry system, which records properties, their location and boundaries, and various values used for tax purposes. I see these terms in work for US clients buying or renting property in Italy, who need to understand what they’re signing. For them, the question isn’t “What is the technically correct term?”, but “What wording gives me a clear picture of what this document actually does, and whether I understand it well enough to sign it?”.
3.2 AI or MT’s usual choices
AI systems and MT engines almost always go for “cadastre” or “cadastral” when they translate catasto / catastale, terms that sound satisfyingly technical and precise. Black’s Law Dictionary, for example, defines “cadastre” as “a survey and valuation of real estate […] compiled for tax purposes”. For US lay readers, the term may look impressive, but it is opaque, because it does nothing to help them understand how a “cadastral survey” fits into the process of buying an apartment.
In real‑estate contracts and notarial deeds, this leads to prose that looks precise but fails to convey its meaning. A clause mentioning a mappa catastale may emerge as “cadastral map”, a reference to rendita catastale as “cadastral income”, and a simple check at the land office as “cadastral survey”. US buyers reading such phrases understand little: they don’t know who keeps these records, what authority they carry, or how they relate to anything they’ve seen back home.
3.3 Why I usually prefer “land registry”
When I translate for US clients dealing with Italian property, my guiding principle is simple: the client must be able to understand the meaning. I therefore generally prefer “land register” or “land registry” over “cadastre”, even if the latter looks closer to the Italian term. To a US reader, “land registry” signals “official place where property records are kept”, which is close enough conceptually and easier to understand for a non‑specialist.
So visura catastale is more helpful as something like “extract from the land registry (visura catastale)” than as a bare “cadastral search”. Mappa catastale becomes “land registry map” or “land registry plan”, depending on the document. For rendita catastale, I normally use an explanatory translation: for example, “notional property income for tax purposes (rendita catastale)”, keeping the Italian term in brackets so that any later cross‑checking with Italian documents or forms is easier.
This doesn’t mean “cadastre” is always wrong. In a different context – for instance, a technical article explaining how the Italian land registry system works – I might very well keep “cadastre” in my translation.
The important distinction is audience and purpose: translations should serve their intended function in the target context rather than mechanically mirror the source. AI can tell you that catasto often maps to “cadastre”, and it can list definitions and examples. What it cannot do is weigh up whether your specific readers are better served by that term, or by a more accessible “land registry” plus a brief explanation.
That is where professional subject‑matter knowledge and judgment come into play. AI proposes strings of words; a professional understands what those words mean, and which are appropriate for the task at hand.
3.4 “But you can always add a glossary”
A fair objection is that terminological issues can be handled through glossaries or client‑specific termbases. If a client decides that catasto is always “land registry” in their English documents, an automatic system (whether AI‑ or MT‑based) can be instructed to translate it so.
That’s helpful as far as it goes, but it doesn’t solve the underlying problem: even a mandated term still has to make sense for a specific audience and in a specific situation, and someone must know when that glossary choice no longer serves that purpose. Glossaries encode past decisions; they don’t remove the need to make new ones when the context shifts.
3.5 “Private international law” or “conflict of laws”?
The same pattern appears in other fields: AI offers something that looks legally or technically correct, but only subject‑matter knowledge tells you whether it fits the system, the audience, and the purpose of the text.
Take a standard sentence that a law firm might use to describe some services it offers:
Lo studio assiste i clienti in materia di diritto internazionale privato, ad esempio per individuare la legge applicabile ai rapporti patrimoniali tra coniugi di diversa cittadinanza.
A first draft translation might be:
The firm assists clients in matters of private international law, for example, in determining the law applicable to property relations between spouses of different nationalities.
Linguistically, there’s nothing wrong with that. An MT system or a rushed human translator could easily stop there. During review, though, I might decide to recast it as:
The firm advises clients on conflict‑of‑laws matters, for example, in determining which law applies to matrimonial property regimes between spouses of different nationalities.
Here I’m not correcting a mistranslation; I’m choosing a formulation that makes more sense for the likely readership. In an academic context, “private international law” may be the natural field label, and I would keep it as is, perhaps adding “(conflict of laws)” on first mention for US‑trained lawyers. In a law‑firm profile aimed at lay clients, though, I favor the more common “conflict of laws” and make “matrimonial property regimes” explicit.
AI can give you “private international law” and even offer “conflict of laws” as an alternative. What it cannot do is decide which label fits this specific text for this specific readership. That decision rests on your knowledge of how lawyers talk about these issues and how much conceptual scaffolding your readers already have.
4. “Truth is always a defense”: when the words are fine but the content isn’t
Sometimes the problem is not the translation but what the source text is claiming. Imagine a global corporate policy drafted at US headquarters:
Under our global communications policy, truth is always a complete defense to any defamation claim.
A correct translation might be something like:
Secondo la nostra politica globale sulle comunicazioni, la verità è sempre una difesa completa contro qualsiasi causa per diffamazione.
The translation is correct and mirrors the English source text. An inexperienced translator might sign off on it; even an experienced one, under time pressure, might not give it a thought. Yet anyone who knows a bit of Italian defamation law will feel uneasy. In the Italian legal system, even truthful statements can be defamatory if they unjustifiably harm a person’s honor or reputation. The neat “truth is always a complete defense” formula is simply incorrect in Italian law.
In a case like this, I would leave the Italian sentence as it is, because it reflects what the policy says, but I would flag it to the client. My note might say that this rule is modeled on US law and may need to be reviewed or rewritten for Italy before the policy is rolled out locally. No AI system will ever volunteer that comment: it does not know where one legal system stops and another begins.
This is another way subject‑matter expertise shows up in our work. Sometimes it means choosing “conflict of laws” over “private international law”. Sometimes it also means not changing a single word of the translation, but warning the client that the underlying claim doesn’t travel as well as they may think.
5. Domain expertise, and how to build it
“Subject‑matter expertise” might sound like a grand label, something only actual experts like lawyers or engineers can claim, but it’s more modest and more attainable than that. It’s the cumulative result of studying, gaining experience, paying attention to how people in a given field communicate among themselves, and noticing where systems and expectations differ across languages.
Part of it is exposure: reading real‑world documents rather than only parallel texts and glossaries. For real‑estate work, that might mean notarial deed templates, tax guides, and English‑language information sites for foreign buyers. For legal work, it might mean court forms, statutes, bar association materials, and law‑firm blogs in both languages. Over time, you build a mental model of who the actors are, what institutions exist, which text types are standard, and how people phrase things when they’re not writing for translators.
Another part is building your own reference system. I like to keep a corpus for each niche: links to the Italian penal, civil, and procedural codes, to their US counterparts, PDFs of relevant articles, copies of online forms, client‑side FAQs, and a log of past decisions for tricky terms. My reasoning for the choice between “cadastre” and “land registry” goes into that log. So does the decision to use “conflict‑of‑laws matters” on a particular law‑firm website, or to flag a “truth is always a defense” policy line for legal review instead of quietly normalizing it.
Expertise shows in how you use AI, rather than whether you use it at all. If you already know what the catasto is, AI can help you compare “land registry” against “cadastre” in actual usage, or suggest alternative ways of phrasing an explanation for US buyers. If you already understand the difference between libel and slander in US law and Italian diffamazione, AI can help you find references and alternative wordings for a translator’s note. Without that base, the same tool is just as likely to lure you into mistakes, because it will give you authoritative‑sounding answers that fit the wrong system.
Subject‑matter knowledge is not an optional extra that AI makes less important. It is what allows you to decide when the machine’s output is good enough, when it needs adapting for your readers, and when the problem lies not in the words but in the underlying assumptions. That is the judgment clients should expect when they hire a professional translator, and it is the part of the job no model can take over for you.
6. A practical subject‑matter checklist
I keep a simple checklist in mind when I translate with AI in the loop; at the top of that list is always to start with the purpose of the translation.[1]
1. What is the purpose of this translation?
What job does this text need to do in the target context – help a US
buyer understand a contract, inform specialists, or support an academic
article? The same Italian text may need to be translated differently in
English for each purpose, as long as you don’t distort its legal effect
and underlying meaning. An academic article on Italian property law
might prefer “cadastre”, while the translation of a deed for a US buyer
won’t.
2. Who will actually read it?
Are we translating for lay US property buyers, Italian notaries,
in‑house counsel, or EU policy staff? Your choice between a specific
technical term and an explanatory phrase depends on this. A term that
works for a specialist audience may be pure noise for someone who just
wants to know what happens if the roof leaks.
3. What does your audience need to do with it?
Do they need to sign it, follow it, file it, or simply be informed?
Someone who has to sign a notarial deed needs more than a term that
looks technically correct; they need a fair chance of understanding what
they’re committing to. For a background briefing, on the other hand, a
more technical label might be perfectly acceptable if it’s explained
once.
4. Would this term mean anything to them?
If you showed your reader the term on its own, would they recognize it
and know what it involves? If not, you need a better solution, possibly
with the original term in brackets on first occurrence. That AI can
define the word doesn’t mean your readers will understand it when they
see it in a contract.
5. Is there a real‑world equivalent in the target
system?
Does your audience’s country or legal system have an institution that
plays a similar role, even if the match isn’t perfect? In many
English‑language contexts, “land registry” is closer to what a US buyer
expects than “cadastre”, even if the Italian catasto is not
identical to any US office. The goal is a responsible approximation, not
a one‑to‑one mirage.
6. How much risk is involved in this choice?
If a non‑specialist misunderstands this term, are we looking at mild
confusion, or potential legal or financial consequences? The higher the
risk, the less the translation can afford to hide behind a dictionary
gloss or an AI suggestion that “sounds legal”. In regulated or
high‑stakes domains, “defensible but potentially misleading or unclear”
is often the worst of both worlds.
7. How does this interact with your existing
terminology?
Will accepting the AI’s suggestion break consistency with your
glossaries or previous work for this client or domain? Sometimes the fix
is as simple as giving the AI a client glossary or termbase so that it
stops inventing alternatives. But a glossary doesn’t remove the need for
judgment; it just moves it upstream, to when terms are chosen and to the
cases where the approved term doesn’t fit the purpose. And when a client
questions an inconsistency, AI will not defend your choices for you.
8. What role do you want AI to play?
Are you asking AI to propose synonyms, retrieve definitions, check
consistency, or simply sanity‑check a decision you’ve already made? If
you catch yourself expecting it to decide which option fits your purpose
and audience, that’s a red flag. In my workflow, once I’m clear on these
questions, AI is there to challenge me and to point out blind spots, not
to choose the target term on my behalf.
If you can’t answer several of these questions with reasonable confidence, the problem isn’t your review prompt or your choice of AI tool. It’s that you don’t yet understand the subject – or the translation’s purpose – well enough to delegate any part of the decision‑making. That’s not a criticism; it’s a signal about where to invest your next hours of research.
7. Where this leaves human + AI workflows
Taken together, the picture is less dramatic than the “AI has done away with the gap” story, and much more useful. AI can help us check consistency, spot alternative phrasings, and compare how terms are used across sources, but it cannot decide which solution fits a given purpose, audience, and level of risk. That part depends on subject‑matter knowledge and on the professional responsibility you can’t outsource to a model.
In the workflow I described in Article 1, what matters most is when you use AI. I translate first, making concrete decisions about terms like catasto based on who will read the text and what they need to do with it; then I let AI challenge those decisions under strict “no unnecessary changes, always explain why” constraints. Used this way, AI doesn’t level the field between strong and weak translators; it widens the gap, because the more you know about your subject, the more the tool can help you.
The next articles in this series will pick up from here and get more practical. I’ll look at how to design prompts that reflect your existing decisions and glossaries, and at how to integrate AI review into a broader toolbench that includes CAT tools, QA checks, and your own reference material. The core principle, however, will stay the same: AI can speed up parts of the process, but it cannot think through our choices or take responsibility for them. That’s still our job – and it’s where a professional translator’s value lies.
Notes
[1] Vermeer and Reiss’ Skopos theory reminds us that the purpose of the translation should guide our choices, and Andrew Chesterman’s Memes of Translation is still one of the clearest accounts of how ideas like this shape everyday practice.
