AI translation: where is the line for using it without a professional translator
Contents
"AI translators have gotten a lot better" and "you still need a human to check it" can both be true. The real question is where the line falls.
Background — why this line matters now
As tools like DeepL, ChatGPT, and Gemini slip naturally into translation work — from email drafts to official letters with overseas partners — the decision to "have this translated" is disappearing as a conscious choice. The habit formed from translating messenger chats carries straight over into contracts and patient instructions.
The problem isn't that AI translation quality is bad. It's that the damage a mistranslation causes varies by content, and that difference gets ignored with a "it's AI output, so I'll just use it as-is" mindset. BLEND, a professional translation service, points to "not segmenting content by importance" as the most common failure (BLEND, September 2025).
Key data & current state
- Legal translation comparison study: A study comparing AI translation with human translation on legal text found that while AI handles word-level correspondence quickly, it diverges from human translation in context and concept translation for clauses that carry legal force (PMC, a National Institutes of Health resource).
- Structural risks in medical interpretation: A paper on AI-based medical interpretation notes that when AI is used alone, without oversight from a clinician or professional interpreter, a patient's informed consent itself can be undermined (PMC).
- Quality-gap figures are inconsistent. Numbers translation vendors cite — such as "AI translation accuracy 60–85%, professional translation 95%+" — vary even across different posts from the same vendor, and often don't disclose the underlying primary research (BLEND). It's safer to look at which content types show the widest gap than to fixate on the numbers themselves.
Takeaway: what matters more than competing figures is whether a mistranslation, if it happens, causes damage that can be undone.
The judgment criterion — split by the content's level of responsibility
1) Everyday and internal reference
Content like internal memos, summaries of foreign articles, or chat translations, where there's a chance to ask again or double-check if something's wrong. AI translation alone is enough.
2) Customer-facing and marketing
Brand messaging or announcements, where a wrong nuance dents credibility but doesn't directly create legal liability. An AI first draft followed by a light check of tone and context from a staff member is sufficient.
3) Legal and contracts
Sentences like clause numbers, amounts, deadlines, and scope of liability, where differing interpretations can cause major damage. Legal equivalence takes priority over natural phrasing, so a legal professional or a qualified translator must review the translation.
4) Medical and consent documents
Content tied directly to patient safety, like dosages, contraindications, side-effect warnings, and consent procedures. A study on discharge instructions found that the quality of AI-only translation varied widely by language, and that having an expert post-edit the AI's first draft was more reliable (PMC). Using AI-only translation without review from a licensed medical interpreter or clinician is not recommended.
How humans fit in — post-editing, not full retranslation
Needing human review doesn't mean retranslating everything from scratch each time. ISO 18587:2017, from the International Organization for Standardization, is a standard that defines requirements for "post-editing" — a human refining machine-translated output — and the qualification criteria for post-editors (ISO). The compromise in common use is: AI produces the first draft, and a qualified person focuses their edits only on the legally or safety-sensitive parts — faster than having a human translate everything, while still cutting risk.
Who ends up responsible
There aren't yet many legal precedents specifically disputing translation, but there are cases showing where responsibility for AI-generated output lands. In February 2024, British Columbia's Civil Resolution Tribunal (CRT) in Canada held Air Canada liable over a refund policy its chatbot had misstated, ruling that "it doesn't matter whether the information came from a static page or a chatbot" (DWW Law). It isn't a translation case, but the principle — that whoever delivers an AI tool's output to a customer or partner bears responsibility for it — applies just as well to translation.
Practical checklist
- ✅ Internal reference documents, everyday communication that can be corrected later — AI translation alone is enough
- ✅ Customer notices and marketing copy — AI first draft + a staff tone check
- ❌ Contracts, terms of service, documents carrying legal liability — a professional translator or legal review is required
- ❌ Medical documents tied directly to patient safety, like dosages or consent forms — confirmation from a qualified medical interpreter or clinician is required
- ❌ Documents where you can't immediately answer "who loses out if this is wrong" — add a review step before you start
Conclusion
- AI translation accuracy figures differ by vendor, so they're a weak basis for a standard on their own. It's more useful in practice to draw the line based on the weight of responsibility a mistranslation would create.
- For content that's hard to undo, like legal or medical material, a qualified person's post-editing must follow the AI's first draft.
- The principle that the organization delivering AI-generated output bears responsibility for it has already been confirmed in other AI services. Translation is no exception.
References
Contents
Related posts
In-house AI tools: build or buy? — a decision framework
Whether to build an in-house AI tool yourself or buy a SaaS product shouldn't be decided by price tag. Judging it on three criteria instead — how standard the requirements are, how tight your throughput and timeline are, and whether regulatory or data constraints apply — is what keeps you from regretting the choice down the road.
In-house AI chatbots: RAG or fine-tuning? — how to decide
It's easy to get confused about whether to try RAG or fine-tuning first when feeding internal documents into an AI. Using the distinction that RAG changes what the model knows while fine-tuning changes its behavior and tone, this piece lays out how to choose based on data volatility and purpose.
AI coding agents: local vs. cloud selection criteria
Whether to run an AI coding agent locally or in the cloud isn't a performance contest — it comes down to structural factors like data sensitivity, network environment, concurrent throughput, and access to the latest models. Here are the four comparison axes and a hybrid operating pattern.
AI design tools: replacing designers, or assisting them? A framework for judgment
As AI design tool adoption surges, the debate over whether designers will be replaced has flared up again. Using 2026 survey data, we separate real threats from hype by distinguishing templated production work from work that needs context and feedback.