Key takeaways
- In a JAMA Internal Medicine study, Google Translate rendered 92% of discharge-instruction sentences accurately into Spanish but only 81% into Chinese.
- A study in Pediatrics found machine tools close to professional translation for Spanish and Portuguese but weaker for Haitian Creole, a language with less training data.
- DeepL's privacy policy states that free-service texts are processed to train its models, while DeepL Pro texts are deleted after translation and not used to improve the service.
- OpenAI says it may use content from individual ChatGPT plans for training unless the user opts out, but not business or API data by default.
- Use AI translation only after deciding the risk level of the content, the data tier of the tool and who will review the output.
This article is for managers who already use, or are being asked to approve, AI tools such as DeepL and ChatGPT for translating business content. It looks at what independent studies show about accuracy, what the tools' own terms say about your data, and how to decide which content can go through AI and which needs a professional.
The workflow in which a linguist corrects machine output is covered separately in what machine translation post-editing (MTPE) is.
How accurate are DeepL and ChatGPT for business translation?
AI translation tools are now accurate enough for many simple sentences in major language pairs, but error rates rise sharply with less common languages and specialist content, and the errors are not flagged. That second point matters most: the output reads fluently whether or not it is correct.
Medical research gives the clearest published evidence, because the consequences of errors are measured. A study in JAMA Internal Medicine by Khoong and colleagues tested Google Translate on emergency department discharge instructions: “Overall, 594 (92%) and 522 (81%) sentences were accurately translated into Spanish and Chinese, respectively”. Of all sentences, 2% of the Spanish and 8% of the Chinese translations had the potential for clinically significant harm.
A later study of ChatGPT and Google Translate, published in Pediatrics and summarised by the American Academy of Pediatrics, found that “machine translation platforms have comparable performance to professional translations for Spanish and Portuguese”, but that “shortcomings in quality, accuracy, and preference persist for Haitian Creole”. The lesson for business buyers is that results from one language pair cannot be assumed for another.
Why does DeepL or ChatGPT get business text wrong?
AI tools translate sentence patterns, not your business context, so they fail where meaning depends on information outside the sentence. Typical failure points are:
- Company terminology. Product names, internal job titles and defined contract terms are translated inconsistently unless a glossary is enforced.
- Legal and regulatory concepts. UK terms such as “without prejudice” or “reasonable endeavours” may have no direct equivalent.
- Omissions and additions. Large language models can drop a negation or add content that sounds plausible.
- Tone and audience. Formal and informal address, which matters in German, French or Polish, may switch mid-document.
- Numbers, units and dates. Decimal separators and date formats are sometimes left in UK format.
Our older article Humans vs Machines discusses why nuance and idiom remain difficult for machines.
What happens to confidential data you paste into DeepL or ChatGPT?
What happens to your text depends on the tool and the licence tier, so the free version and the business version of the same product can behave very differently.
| Tool and tier | Used to train models? | What the provider states |
|---|---|---|
| DeepL Translator (free) | Yes, for a limited period | “We process the content you upload and their translations or improvements for a limited period of time to train and improve our neural networks and algorithms.” |
| DeepL Pro | No | Texts are deleted after the service is performed, and “your texts will not be used to improve the quality of our services.” |
| ChatGPT Free, Plus and Pro | Yes, unless the user opts out | “we may use your content to train our models.” |
| ChatGPT Business, Enterprise, Edu and API | No, by default | “By default, we don’t use inputs or outputs from ChatGPT Business, ChatGPT Enterprise, ChatGPT Edu, or our API to improve our models.” |
Sources: DeepL privacy policy and OpenAI help centre. DeepL's policy also states that its free service may not be used for texts “containing personal data of any kind”.
UK guidance for the judiciary shows how seriously public tools are treated. The AI guidance for judicial office holders states: “Do not enter any information into a public AI chatbot that is not already in the public domain.” For the wider data-protection picture, read data security when outsourcing translation.
Which business content can you safely translate with AI?
Decide by risk: the higher the cost of an error, the more human involvement the content needs. The table gives a practical starting point.
| Content | Risk if wrong | Suggested approach |
|---|---|---|
| Incoming emails and documents, for your own understanding | Low | AI on a business-tier tool is usually fine |
| Internal notes and drafts | Low to medium | AI plus a check by a fluent colleague |
| Product data, support articles | Medium | Machine translation with full post-editing |
| Marketing, website and brand copy | High (reputation) | Human translation or transcreation |
| Contracts, regulatory, medical, safety content | High (legal or safety) | Specialist human translation and revision |
How should a business set rules for AI translation?
A short internal policy prevents most problems. Include these points:
- List approved tools and the licence tier, and ban free versions for anything confidential or personal.
- Classify content into the risk levels above, with named examples from your organisation.
- Require a qualified reviewer for anything external, published or contractual.
- Maintain a glossary of company terms and supply it to both tools and translators.
- Record which documents were machine-translated, so errors can be traced.
Common mistakes when using DeepL or ChatGPT for translation
The most damaging mistake is publishing AI output that nobody in the organisation can read. Others to avoid:
- Back-checking by re-translating into English. A round trip can hide errors that cancel out.
- Trusting fluent output. Fluency is not accuracy; check numbers, negations and names.
- Assuming results transfer between languages. Good French output says nothing about Romanian or Haitian Creole.
- Pasting client data into personal accounts. This may breach your contracts and data-protection obligations.
For legal and clinical content, see machine translation for legal documents and machine translation in healthcare; for the full buying process, see how to choose a translation agency.
How TranslationWorks can help
TranslationWorks is a UK-based translation agency that provides professional human translation by native-speaking linguists in 150+ languages, for content where machine output is not enough. It holds ISO 17100 (translation services) certification, ISO 27001 (information security) and ISO 9001 (quality management) registration and is Cyber Essentials certified, which is relevant when sharing confidential files. If you are unsure whether a document can go through AI or needs a professional, tell us what it is and how it will be used, and we will confirm what we can provide in your free quote. See our translation services or get a free quote within 24 hours.
