This guide is for anyone translating emails, work documents or texts with a personal tone who is unsure whether to use DeepL, Google Translate or a chatbot such as ChatGPT, Gemini or Claude. The short answer: a conventional translator for volume and layout, a chatbot when register matters and you can give context, and no free service when the document is confidential. Most users overlook that last point, and it costs most.
- Key Point: Quality is no longer the main problem: at WMT 2025, human raters ranked a general language model first, not a dedicated translation engine.
- Key Point: Gemini 2.5 Pro reached the top group in 14 of 16 human-evaluated language pairs and matched or beat reference human translation in 10.
- Key Point: WMT25 caveats: Gemini used reasoning mode, commercial services were anonymised, Spanish was not human-scored.
- Key Point: Free tiers may reuse text for training; DeepL’s free terms bar personal data, so confidential work needs paid terms or anonymisation.
What WMT25 showed — and hid
The WMT25 results report, published in November 2025, assessed 60 systems. Professional human translations made the winning group in only 6 of 15 pairs; the authors cite translation difficulty but accept that rater style preferences may also matter.
The organisers give three warnings. Gemini’s output was collected with reasoning mode active, and they cannot quantify its effect. The three commercial services appear as ONLINE-B, ONLINE-G and ONLINE-W, so the report cannot say whether DeepL or Google Translate led or trailed. Spanish was not among the human-evaluated pairs.
The report concludes that leading systems still fail on specialised terminology, unconventional texts and gender agreement in some languages. A chatbot can address that if you explain the audience; a boxed translator cannot. My reading is that the
“chatbot contra traductor”
debate is badly framed: the context you supply is the differentiator.
Where each tool earns its keep
Google Translate is the choice for formatted files and offline use. Its help says it translates .docx, .pdf, .pptx and .xlsx files up to 10 MB, with a 300-page PDF limit, but only in a desktop browser; document translation does not work on mobile. Android users can download languages for offline translation, useful on trips and showing on-device versus cloud AI. For sign or menu photos, Google warns that small, blurry or decorative lettering is less accurate.
DeepL also preserves document layout. Its privacy policy says the free version accepts .docx, .pptx and .pdf, while paid adds .html and .txt, with processing in European Economic Area infrastructure.
A chatbot is right when text has intent: a delicate email, a slogan, a message that should sound close without being informal. You can ask it to keep tuteo, adapt an idiom or offer two register variants. Good prompts follow techniques from OpenAI, Anthropic and Google. A useful trick is to translate with one tool and ask another to flag forced phrases. It does not replace a professional translator for a contract.
Privacy after you paste
DeepL is direct: on the free version it processes uploaded content and translations for a limited period to train and improve its neural networks, corrections included. It also bars texts containing personal data of any kind from the free version. DeepL Pro contracts otherwise: texts are not stored permanently, are deleted when translation ends and are not used to improve the service, though document translations, free or paid, keep certain technical metadata for 14 days.
Google separates consumer and enterprise products. For Cloud Translation API, it says submitted content is not used to train translation features and stays in memory only as long as needed. The free Translator help pages consulted here make no equivalent commitment, so it should not be assumed.
Personal chatbot plans may use conversations for training unless you disable it or use temporary modes. ChatGPT’s data-controls help says training is on by default for Free, Plus and Pro; it is disabled with
“Improve the model for everyone” (mejorar el modelo para todos)
. Temporary chats are excluded from training but may be kept up to 30 days for safety. Business, Enterprise and Edu plans are not used for training by default. Gemini’s apps privacy centre sets out its own rules.
Translation quality now suffices for much daily work, and WMT25 supports that. What lags is user awareness. The costliest mistake is rarely an awkward sentence; it is pasting a confidential document full of names and figures into a free tool whose terms forbid it. Buy the professional tier before an employee seeks a shortcut.