What translation NDAs usually say, what ChatGPT, DeepL, Google Translate and other tools do with your text, and when AI is off limits.
I've translated between English and Spanish for years, and I've signed a lot of NDAs. They differ in length and tone, but most of them ask for the same few things. Don't share the files. Don't give them to third parties. Delete them after a set period. One client goes further: files come and go through Box, and nowhere else. Others say plainly that no AI may touch their content, because it includes patients' health information.
None of that is unusual. What changed is how easy it has become to break those rules without noticing. Select, copy, paste into a chatbot, and the text is on someone else's server. This post is what I check before I do that, and why some texts never go near an AI tool.
Policies change, so I checked every page linked here on 7 November 2026 (see Sources at the end).
An NDA doesn't need to mention AI to forbid it. If it says you may not disclose the material to third parties, then the company that runs the chatbot or the MT engine is a third party. Its servers receive the text, store it for some time, and in some cases use it to improve their models.
"Delete the files after 30 days" has the same problem. You can delete your copy. You can't delete the copy on a server you don't control, and you often don't know how long it stays there.
The American Translators Association says it directly in its 2025 statement on AI: AI tools "may store, process, or even repurpose the data", which "can lead to breaches of confidentiality agreements and violations of data protection laws". Its Code of Ethics and the ITI and CIOL codes already require us to keep client information confidential.
This isn't theory. In 2017, Norway's NRK reported that Statoil contracts and dismissal letters were findable online because staff had used the free site Translate.com. In 2023, Samsung restricted generative AI on company computers after staff uploaded sensitive code to ChatGPT.
The free consumer product and the paid business or API product of the same company often follow different rules. Here's what each company's own pages say.
ChatGPT. OpenAI's Data Controls FAQ says personal Free, Plus and Pro accounts can be used for training unless you turn off the option in Settings → Data Controls that lets OpenAI use your chats to improve its models. Business, Enterprise and Edu workspaces are excluded by default. Its API data page says API data "is not used to train or improve OpenAI models" unless you opt in, and abuse-monitoring logs are kept for up to 30 days.
DeepL. The privacy policy says that with the free translator, DeepL processes your texts "for a limited period of time" to "train and improve our neural networks". With Pro and the API, texts "will be deleted" and "will not be used to improve the quality of our services".
Google Translate. For the Cloud Translation API, Google says it "does not use the content you send to train and improve our Google Translation features" and holds it "briefly in-memory". For the free translate.google.com, I didn't find a statement either way. The privacy policy only says Translate works by "detecting common language patterns in phrases you ask it to translate".
Microsoft Translator. The Azure Translator documentation says it "doesn't persist customer data submitted for translation". Microsoft's No Trace page says that with the free apps, Translator for Bing and Edge, "text translations are recorded for service improvements purposes".
Gemini and Claude. Google's Gemini privacy hub says some chats are read by human reviewers and kept for up to three years, and chats are kept for 72 hours even with activity off. Anthropic lets consumer users choose whether chats train Claude: if you allow it, data is kept up to five years, otherwise 30 days. Work plans and the API aren't covered by that change.
A paid tier isn't automatically fine for client work. It just has terms you can show the client, and the client still decides.
Some texts can go through AI without any problem: a public product page, a press release that's already out, my own blog posts. Others can't, and the reason is usually the law, not the client's mood.
Health information. In the US, a hospital or insurer can share protected health information with a vendor only with a business associate agreement, a contract in which the vendor promises to safeguard it. HHS says even a cloud provider that stores only encrypted data is a business associate. If my client has a BAA with its translation vendor and I paste the file into a free chatbot, the chatbot has no BAA with anyone. That's why my medical clients forbid AI, and I think they're right.
Personal data in Europe. Under the GDPR, a company that has someone process personal data on its behalf must use processors with "sufficient guarantees" and a written contract (Article 28). A processor "shall not engage another processor without prior specific or general written authorisation". If I'm the processor, an AI service I add on my own is another processor nobody authorized.
Patents. This is the clearest case for me. A patent application has to be new. Under the European Patent Convention, the state of the art is everything "made available to the public" before the filing date (Article 54). The EPO guidelines say a disclosure under a secrecy agreement that hasn't been broken doesn't count (G-IV 7.2.2). That's exactly why patent work comes with an NDA. The US has a one-year grace period for the inventor's own disclosures (MPEP 2152), but that doesn't cover a leak through a translator.
The USPTO itself warned practitioners in 2024 that "AI systems may retain the information that is entered by users", and that if confidential information is used to train AI, parts of it "may filter into outputs from the AI system provided to others". An unpublished invention, pasted into a tool that trains on it, could end up shaping someone else's answer. I don't do that, and no client has ever asked me to.
This one is personal. I don't like the idea of a platform taking translators' work, training its MT or AI on it, and then selling that product back to us at a price. It's my opinion, not a legal point, and it's why I read the terms of every platform I use.
Matecat is the one I've thought about most. When I used it, Matecat machine-translated the text and I post-edited the segments. Translated, the company behind it, also sells its own MT: ModernMT, now turning into Lara. So I'd want to know whether my corrections end up improving the engine they sell back to us. I don't know that they do. Here's what their pages say:
So their terms allow it for MyMemory, and I found nothing that rules it out for ModernMT. If you use Matecat for confidential files, ask the client first, attach a private TM key and check the MT settings before you open the first segment.
The other platforms I checked:
So the picture is mixed, and it changes. Read the current terms, and if they're vague, ask the company in writing.
This isn't legal advice. If an NDA is unclear, ask the client or a lawyer.
I built the CAT in TraduQ with these rules in mind. It runs in your browser, keeps your files, memories and glossaries on your computer, and has no machine translation built in, so nothing is sent to an MT engine. TraduQ doesn't sell your data or use it to train AI, and an optional end-to-end encrypted cloud backup is coming. The Security page explains what we can and can't see.
Pages checked on 7 November 2026: