AI translates very well, up to a point. There are four specific scenarios where letting it run unsupervised can cost you dearly: contracts and legal documents, marketing and branding, technical and safety manuals, and anything containing confidential information. We explain why it falls short in each one, with real cases, and what we do at Overseas so you don’t end up paying twice for the same translation.
AI isn’t the problem; using it without a safety net is
At Overseas we use AI every day. We integrate it with professional human post-editing (we explain this in our article on AI and human review) and combine it with the translator’s judgement depending on the document, as we discussed in DeepL, ChatGPT or a human translator. But there are areas where letting it work alone is simply bad business. Not out of ideology, but from experience.
Scenario 1: contracts and legal documents
The problem here isn’t that AI translates badly, but that it translates plausibly. A legal sentence can read perfectly well and, at the same time, have changed who is liable for what. An obligation becomes a recommendation, an absolute deadline becomes relative, an indemnity clause falls short. And you don’t notice until the dispute arrives.
A documented case: in a machine translation programme for courts in the United States, the system rendered the English word court (as in a court of law) as a sports court in several contexts. It is recorded in a report by the National Center for State Courts, and it caused enough confusion among staff and users for the whole workflow to be rethought.
For legal texts, the bare minimum is a human translator specialising in law and, if the destination requires it, a sworn translation backed by professional accountability.
Scenario 2: marketing, branding and emotional copy
AI translates words; your marketing needs its intent translated. That’s not the same thing. A slogan, an advertising claim or campaign copy works because it has rhythm, double meanings and a specific promise. That’s something a person who knows your industry and your customer will pick up on, not a model trained on billions of sentences that have nothing to do with yours.
The textbook example of how not to do it is still KFC’s entry into the Chinese market: its iconic Finger-lickin’ good came out in Mandarin as something like “eat your fingers off”. It’s an old story, but the pattern still crops up every week on websites and campaigns translated with DeepL and never reviewed.
And there’s a measurable cost. Nimdzi estimates that brands that skip professional localisation of slogans, images and user experience lose up to 25% in engagement and conversion.
Scenario 3: technical manuals and safety documentation
In an installation manual, a safety data sheet or a technical specification, a single word can mean two very different things depending on the industry. AI doesn’t know which one you’re in.
A couple of examples from the industrial and medical sectors: the English word clearance can mean “regulatory approval” or “free space”. In the manual for a medical device, mistranslating this term can suggest that the product isn’t approved when it is, or vice versa. In mechanical manuals, nuts and washers should be rendered in Spanish as “tuercas” and “arandelas”, but AI, lacking context, translated them as “frutos secos” (edible nuts) and “lavadoras” (washing machines).
On a construction site, in an industrial installation or on a safety label, this is no longer a typo. It’s a safety risk, a compliance issue or a costly correction. For the industrial sector, one of our specialisms (engineering, construction, materials, transport), the rule is a specialised human translator or, at the very least, AI with full post-editing by a translator from the sector.
Scenario 4: confidential or regulated information
This is the one least talked about and, in many cases, the most serious. When you paste a text into a public AI service, that data travels to the provider’s servers. For internal drafts it doesn’t matter; for a signed contract, personal data covered by the GDPR, a medical record or non-public financial information, it does.
Even if the translation comes out flawless, you have moved sensitive information outside your control. That could land you with a compliance problem (GDPR, sector regulations) or, depending on the client, a breach of the confidentiality agreement you signed with them. The major tools offer “enterprise” plans with policies against training on your data, but you have to activate them and read the small print.
A professional agency handles this with closed workflows, translators who have signed NDAs and controlled environments. It’s not a minor detail: it’s part of the service.
A word on “hallucinations”
Generative models sometimes make up what they don’t know. A study on Whisper, OpenAI’s transcription model, found invented sentences in between 38% and 80% of the transcriptions analysed, particularly in medical settings. Translation with large language models isn’t immune to this problem: if an unusual word doesn’t fit what the model expects, it may replace it with another that sounds right but isn’t. The “worst” part is that it slips these inventions in so subtly that it completely convinces you.
That’s why, according to Slator (2025), between 90% and 98% of those using machine translation or LLMs apply some level of human review. The entire industry, including those pushing AI the hardest, treats it as a supervised tool.
What AI does do well
So it doesn’t look like we’re against technology, here’s the other side: AI is great for getting the gist of a text in a language you don’t speak, for internal drafts, for high volumes where publishable quality is achieved through post-editing, and for speeding up the professional translator’s work. It would be absurd not to make the most of it. The trap is assuming that because it works well here, it will work well everywhere.
How we do it at Overseas
When a project comes in, the first step is to decide which route it needs based on risk, industry and confidentiality. For volume without nuance, AI with post-editing, reviewed by a native translator from the sector. For websites, marketing or legal texts, a specialised human translator, and the same goes for sensitive data, with controlled environments and NDAs. Always backed by the ISO 9001:2015 and ISO 17100:2015 certifications we have held since 2017, and applying the criteria of ISO 18587:2020 as best practice on post-editing projects.
If you have a translation on your desk and aren’t sure which group it falls into, drop us a line and we’ll advise you, no strings attached. End-to-end service in every language combination.
Frequently asked questions
Can I translate a contract with ChatGPT just to understand it myself?
As your own draft, yes. As a final document to be signed or delivered, no. And bear in mind that pasting it in there could be a confidentiality leak if it’s a real contract with a third party.
What if it’s just for internal company use?
It depends on the document. An internal note, AI is fine. An operating procedure or an internal technical manual that people will follow, better with human review.
Is it safe to paste confidential documents into DeepL or ChatGPT?
In the free and standard versions, not entirely. Business plans usually offer more safeguards, but you need to activate and configure them. For projects involving the GDPR or NDAs, play it safe with a professional provider.
Will AI improve and make these problems go away?
Quality will keep rising. The thing is, errors become more subtle, not more visible. That calls for more human review, not less.
Sources
- National Center for State Courts — Machine Translation: Considerations and Cautions for Courts.
- Nimdzi Insights — data on localisation adoption and losses for brands that don’t localise.
- Slator — Five Ways AI Reshaped the Translation Industry in 2025.
- Studies on Whisper (OpenAI) hallucinations in medical transcription.
