Translating documents with AI. The machine writes the first version, the glossary decides the terms, and a person reads every line before it goes out.
ChatGPT and DeepL translate a PDF in seconds, and get things wrong where it costs most: technical terms, figures, text inside images, layout. Here is the glossary in Excel, with the review checklist.
A work document is translated reliably with AI by starting from the right file, settling the terms in a glossary first and having a person read the whole text. The machine writes the first version, and no line goes out without a signature.
Human review has a standard. ISO 18587 sets the requirements for full post-editing of machine translation carried out by a person. The WMT25 study, across 60 systems, found that the best ones still struggle with domain terminology.
The plan matters more than the tool. DeepL keeps the file’s formatting; ChatGPT reads the PDF, but OpenAI makes no promise of a laid-out PDF in return. On free plans your texts can be used to train the models.
The glossary and the checklist can be downloaded below, in one Excel file with prompts and a worked example.
Written by: Mattia Esposito, AI and marketing consultant, founder of Itria AI. From Puglia, based in Turin since 2023, he works remotely in Italy and abroad. LinkedIn profile.
The translation glossary template and review checklist, to download
An Excel file with four sheets: the glossary of terms that must not change, the checklist to fill in for every translated document, the data sheet showing what can be uploaded and where, and the prompts. The summaries calculate themselves, and outcomes are picked from a menu. At the end, the glossary and the checklist are filled in for a made-up olive oil mill.

| File | Contents | Link |
|---|---|---|
| Excelto fill in, in Italian | A 40-row glossary, a 12-point review checklist, the data sheet, prompts and a worked example. | |
| PDFworked example, in Italian | Eight glossary entries and the review of a technical data sheet translated into English, with the outcomes counted. |
| Sheet | What you enter | Why |
|---|---|---|
| Glossary | Italian term, type, approved translation per language, whether it is translated, banned translations, who decided. | The best systems still get domain terminology wrong |
| Review | Twelve checks per document, from the source file to the full read-through, with outcome and note. | The scope of ISO 18587: a person, the whole text |
| Data | For each document: personal data, confidential information, tool, plan, agreement signed. | Free plans can use your texts for training |
| Prompt | Full translation, long document in blocks, glossary check, figures check. | For ChatGPT and Claude, which have no glossary to upload |
itria.io has 223 pages in English. None of them is translated word for word: each is rewritten for an English reader, with a fixed glossary, so the same term stays the same from one page to the next.
So the file starts with the glossary and the checklist ends with a person reading the whole text. No tool does those two parts for you, and they are what holds one document together with the next.
223 pages in English, counted on the site on 13 November 2026.
Rewritten, not translated word for word, for readers in English.
A fixed glossary, the same from one page to the next. This guide has its Italian original too.
How to translate a PDF with ChatGPT, step by step
Upload the PDF to the chat, ask for a full translation with no summarising, paste the glossary into the prompt and, if the document is long, go one section at a time. ChatGPT reads the text and images in PDFs, charts and diagrams included, but OpenAI warns that embedded images may not be included automatically.
Uploads can be large: 512 MB and up to 2 million tokens per document, according to OpenAI. The real limit is how much text the model handles at once. The ChatGPT pricing page, read on 13 November 2026, gives about 12 pages for the fast model on the free plan and about 40 on Go, Plus and Business. Limits change: check them on the official page.
Translate the attached document from [language] into [language] for [the reader]. Translate everything, without summarising and without skipping sections. Use these terms as written: [glossary, term = translation].
Do not translate [brands and product names]. Leave figures, units, dates and codes unchanged. Put any sentence you are unsure of in square brackets.
The free plan allows three uploads a day. Accepted formats run from Word to PDF and from Excel to PowerPoint, and a Google Doc has to be exported to PDF or Word first. The file holds three more prompts: the long document in blocks, the glossary check and the figures check, which tables every figure in the original next to its translated version.
How to translate a PDF and keep the formatting
To keep the layout you need a tool that gives you the translated file back, such as DeepL or Google Translate, and the best result comes from the original Word or PowerPoint file. DeepL says its file translation works on the whole document “while preserving the format of the original file”.
OpenAI’s help pages describe how ChatGPT reads a PDF, not how it returns one with its layout intact. If you use ChatGPT, plan on putting the text back into the layout yourself. If you use DeepL and the translated PDF disappoints, DeepL itself suggests uploading the original as .docx or .pptx.
Scans are the weak spot for every tool. For DeepL, the quality of the scan affects the quality of the translation, and PDFs exported from CAD software, such as floor plans, are not supported. Google Translate leaves the text in images and scanned pages in the output file without translating it. Before translating a scan, the text has to be extracted: the guide on how to extract data from a PDF with AI and the entry on OCR explain how.
DeepL vs ChatGPT vs Google Translate for work documents
DeepL suits a file that has to come out laid out and with a glossary; ChatGPT suits adapting the text to the reader or checking a translation; Google Translate suits a long PDF you need to read, not send. The differences lie in formats, limits and what happens to images.
| Tool | Files and stated limits | Layout and images | Fixed terms |
|---|---|---|---|
| ChatGPT | Word, PDF, Excel, PowerPoint and more; 512 MB per file; three uploads a day on the free plan. | Reads text and images; no promise about the PDF that comes back. | The glossary goes in the prompt |
| DeepL free | Word, PowerPoint, PDF up to 5 MB and 100,000 characters; no Excel. | Keeps the formatting; translated files cannot be edited. | Not stated for the free plan: see DeepL’s plans page |
| DeepL Pro | Up to 30 MB and 1 million characters; Business 100 MB; Excel and InDesign (IDML) as well. | Keeps the formatting; scans and custom fonts come out worse. | Glossary, translation memories, formality |
| Google Translate | Word, PDF, PowerPoint, Excel up to 10 MB; PDFs up to 300 pages; desktop only. | Returns the document; text in images and scans is not translated. | Not covered in the help page |
One practical difference concerns Excel: free DeepL does not accept it, Google Translate does. For multilingual price lists and catalogues, the format comes before the tool, and the guide to the product catalogue in English compares one file per language with two columns. Which assistant to use for the rest of office work is covered in the guide to ChatGPT, Claude or Gemini for work.
A translation glossary for technical terms: why it comes first
Because domain terminology is where even the best systems still go wrong. The WMT25 study from the Association for Computational Linguistics, the yearly benchmark of the machine translation conference, says so: the leading systems struggle with non-standard input, domain terminology and gender agreement.
A glossary solves the problem before it shows up. It holds product names and brands that are never translated, legal wording to copy as it stands, technical terms with their approved translation, trade terms such as Incoterms, and the courtesy formulas of your register. Next to each entry go the banned translations: the ones the tool tends to suggest and you never want to see.
In DeepL the glossary is uploaded and the tool applies it by itself. In ChatGPT and Claude it is pasted into the prompt and then checked with the dedicated prompt, because a language model can switch to a synonym halfway through a document. The entries that most often end up in a food exporter’s glossary are in the guide to the technical data sheet in English.
Human review and ISO 18587 post-editing
ISO 18587 is the international standard for post-editing: a person revising a machine translation. According to the ISO page, the standard “provides requirements for the process of full, human post-editing of machine translation output and post-editors’ competences”.
The standard dates from 2017 and ISO lists it as due for revision. The new version, in preparation as ISO/DIS 18587, changes the title from “machine translation” to “non-human translation output”: any translation not done by a person. For translation in general, the standard refers to ISO 17100. The full text is paid for, and the public page does not list the requirements one by one.
What the public page says is enough to set two rules: the review is complete, covering the whole text, and it is done by a person with the skills for it. The checklist in the file starts there and adds the points where tools stop short: scans, text in images, figures, layout. The reviewer and the signatory are the part a system must never skip, as the entry on human approval explains.
How reliable is AI translation?
On hard texts, the best systems of 2025 matched or beat the human reference translation in nine language pairs out of fifteen, and English to Italian was the hardest. Those are the WMT25 findings: 60 systems, 30 language pairs, human evaluation on 15 by professional annotators.
The authors put it this way: “Human references are in the winning cluster for only six out of 15 language pairs”. This does not mean the machine always beats the person: in English to Icelandic, the human translation beat the best system by 20 points. And on Italian the authors are blunt: English to Italian was the hardest pair for the leading systems, along with English to Egyptian Arabic.
The study assesses systems from mid-2025, and today’s models are newer. Another lesson from the same work still holds: automatic metrics got the ranking wrong, and the system ranked first on computed scores sat much lower in human judgement. A quality score does not replace reading, any more than the confidence of a model replaces it when the model makes things up.
What language labels and documents must be in: Italy, EU, UK, US
Mandatory information goes in the language of the country where the product is sold, and that country’s law decides it. For food, Article 15 of Regulation (EU) 1169/2011 requires that “mandatory food information shall appear in a language easily understood by the consumers of the Member States where a food is marketed”.
| Where | What the rule says | Source |
|---|---|---|
| European Union | Mandatory food information in a language easily understood; each Member State may require its own official languages; several languages allowed. | Reg. (EU) 1169/2011, Art. 15 |
| Italy, consumers | All consumer information at least in Italian; where other languages appear, Italian in characters no less visible. | Italian Consumer Code, Art. 9 |
| Italy, loose food and vending machines | Italian required in the decree itself; 1,000 to 8,000 euros for vending machines. | Legislative Decree 231/2017, Arts 18, 19 and 22 |
| United Kingdom | In the UK text of the regulation, in force since 31 December 2020, mandatory information must be in English; several languages allowed. | Reg. 1169/2011, Art. 15, UK version |
| United States | Everything the label must carry goes in English, except products sold only in Puerto Rico or in territories where another language predominates. | 21 CFR 101.15(c)(1) |
The Italian Consumer Code makes no exception by sector: all information intended for consumers and users must be given at least in Italian. That covers instructions, warnings and data sheets meant for the consumer, as well as labels.
For prepacked food, Decree 231/2017 has no article on language; according to a note from ICQRF, the Italian food fraud inspectorate, dated 10 May 2018, missing Italian is penalised under Article 3, now 4,000 to 32,000 euros or 3% of turnover.
So legal wording goes into the glossary as “Legal wording” and is copied from the text of the destination country, never translated. Country by country, the full picture is in the guide to the export label by country. This page is about work documents: deeds and certificates filed with a public office follow the rules of the office receiving them.
Confidential documents: what to upload, and on which plan
A document with personal data or confidential information does not go on a free or personal plan: there, texts can be used to train the models. On ChatGPT business plans and on DeepL Pro, by default, they are not. The Data sheet in the file applies this rule.
| Tool and plan | Are your texts used for training? | Worth knowing |
|---|---|---|
| ChatGPT Free, Plus, Pro | They can be, until you switch off “Improve the model for everyone” in data controls. | With a thumbs up or down, the whole conversation can be used |
| ChatGPT Business, Enterprise, API | No, by default. | OpenAI signs a data processing agreement; admins can see conversations |
| DeepL free | Yes, for a limited period, corrections included. | Documents processed in the European Economic Area |
| DeepL Pro | No: texts deleted after translation. | Data processing agreement for personal data |
OpenAI says it plainly: “By default, we don’t use inputs or outputs from ChatGPT Business, ChatGPT Enterprise, ChatGPT Edu, or our API to improve our models”. On personal plans, temporary chats are not used while they stay temporary. Where the data physically ends up, and what the US CLOUD Act changes, is in the guide to ChatGPT, GDPR and the CLOUD Act.
How to translate a PDF for free
A PDF can be translated for free with Google Translate, free DeepL or free ChatGPT, within their limits on size and uploads. To read a document you have received, that works. To send a document of your own, two things count: uploaded texts can end up in training, and a file translated by free DeepL cannot be edited.
The free version helps you understand what a contract or a data sheet from a foreign supplier says. When the document goes out with your name on it, the sums change: a mistake in a data sheet or a price list reaches the customer, who reads it as yours.
When documents pile up, the work moves from translating to managing: the system that drafts in the customer’s language does that, and the other guides in the series are in the marketing for small businesses index.
Questions and answers
Can AI translate a PDF?
Yes. ChatGPT reads uploaded PDFs, text and images included, and DeepL translates the whole file while keeping the original formatting. There are two limits: scanned pages translate worse or not at all, and the translated text has to be read in full by a person before it is used for work.
How do I translate a PDF file into English?
If you have the source file in Word or PowerPoint, translate that: DeepL itself recommends it when a translated PDF disappoints. Otherwise upload the PDF to DeepL or Google Translate, which give you a file back, or to ChatGPT with a prompt asking it to translate everything without summarising, one section at a time if the document is long.
How can I translate a PDF for free?
Google Translate takes PDFs up to 10 MB and 300 pages, free DeepL up to 5 MB and 100,000 characters, free ChatGPT three uploads a day. For work documents there is a hidden cost: the free version of DeepL keeps texts to train its models, and ChatGPT on personal plans can do the same unless you switch it off.
What is the best PDF translator?
To keep the layout, one that gives you the file back: DeepL states it preserves the original formatting, and Google Translate returns a translated document. For technical terms, what matters is whether the tool accepts a glossary. For confidential documents, the plan matters: business plans do not use content for training and sign a data processing agreement.
Does ChatGPT translate technical documents well?
Often well, though not on its own. The WMT25 study, covering 60 systems on hard texts, found that the best still struggle with domain terminology, and that English to Italian was the hardest pair. So technical terms are settled first in a glossary, and a person reads the whole text.
Notes on sources
- ISO, ISO 18587:2017, Post-editing of machine translation output: public abstract, status “to be revised”; the new version is the ISO/DIS 18587 project. Read on 13 November 2026.
- Regulation (EU) No 1169/2011, Art. 15; Italian Consumer Code, Art. 9; Legislative Decree of 15 December 2017, No 231, Arts 3, 18, 19 and 22 as in force. The reading of Article 3 on language comes from the ICQRF note of 10 May 2018, reference 0156934. Read on 13 November 2026.
- Regulation (EU) No 1169/2011, Article 15, UK text as amended by the Food (Amendment) (EU Exit) Regulations 2019; 21 CFR 101.15, paragraph (c)(1). Read on 13 November 2026.
- OpenAI, Uploading files and audio to ChatGPT, updated on 12 November 2026; ChatGPT pricing, read on 13 November 2026; How your data is used to improve model performance; Enterprise privacy at OpenAI, updated on 8 January 2026.
- DeepL, File formats, updated on 9 November 2026; About file translation; Translate PDF files; Privacy Policy, October 2026, sections 3, 4 and 7.1.
- Google, Translate documents & websites, Google Translate Help. Read on 13 November 2026.
- Tom Kocmi et al., Findings of the WMT25 General Machine Translation Shared Task, Association for Computational Linguistics, November 2025, CC BY 4.0: 60 systems, 30 language pairs, professional human evaluation on 15, texts chosen to be hard. It assesses systems from mid-2025.
- The number of English pages on itria.io can be counted on the site. The glossary and the review in the example use a made-up company, document and outcomes.
The glossary and the checklist handle one document at a time. When documents run to a hundred in five languages, you need a system that drafts them and lines them up for the person who signs. The system first, then growth.
Mattia Esposito, who wrote this guide, builds tailored AI systems for small businesses with Itria: automations connected to the software the business already uses, and a Dashboard showing what has happened and what is waiting for a decision. Nothing goes to a customer without a person’s yes. On a test bench on 4 September 2026, with 30 documents including orders, delivery notes and invoices, the system extracted 234 fields out of 237 correctly, and the unreadable document deliberately placed in the batch was set aside without a single invented field.