“AI makes everything more efficient.” True, mostly. We repeat it like a rule that holds everywhere. It doesn’t. Efficiency is one way to measure work and for a lot of the work I care about, it isn’t the one that counts.
I keep testing that against translation simply because we treat it as a solved problem.
Paste ... Translate ... Move on
And to be fair, the technology has earned some of that confidence. I recently came across a case where a professional needed a technical legal document translated. He ran it through DeepL (paid tier) and checked the output line by line against the original. The translation held up. 100%. So, let me say it plainly. These tools are good and they are getting better. Accuracy isn’t where this story gets interesting.
What gives me pause is a quieter question. The one we rarely ask before we paste:
Where does the document actually go?
The Tool You Think You Know Has Changed
The translation apps most of us reach for aren’t the narrow engines they used to be. Google moved Google Translate onto its Gemini models at the end of 2025. DeepL now runs on its own large language model. In plain terms the ordinary translate box is converging on the same class of technology as ChatGPT. That’s a real leap in quality. It’s also a real change in what happens to your words. Once they’re fed into a large language model, they aren’t simply processed and discarded. Depending on the tool and the tier, they may be stored, used to train the model, and in some cases seen by human reviewers. Microsoft states plainly that a portion of the text entered into its free consumer translation products may be kept to improve its models. DeepL’s free terms go further and tell you not to submit confidential or personal information at all.
And it isn’t only about pasting text.
Point your phone at a page and it translates in real time. Upload a photo and it does the same. Whole documents - a contract, a statement, an ID - all travel to the same servers, often with even less thought than a copy and paste. And these may not stop at the words. A photo carries where and when it was taken, on what device, and when you upload it. All that data travels with it.
Google doesn’t say whether it reads or keeps the location tucked inside your photos. What it did say in June 2026 is that images run through its consumer tools can be used to train its AI unless you go and switch that off. Apple at least offers a real private option. Set Translate to "Always" on-device and nothing leaves your phone. But, that isn’t shipped by default. By default, your text can still reach Apple’s servers and be kept up to two years.
Put simply, pasting a sensitive document into a free tool is close to handing it to an outside company with no agreement to keep it to itself. Not because anyone is acting in bad faith, but because you never set the terms.
For Regulated Work, This Isn’t a Preference. It’s the Law.
Here’s where it stops being a matter of taste. If you handle other people’s confidential information for a living, confidentiality isn’t a courtesy you extend. It’s a duty you can be punished for breaking and reaching for a convenient tool doesn’t shift the blame off you.
Under European data-protection law, the moment you send personal data to a third-party tool, that provider becomes a “data processor” and you remain responsible as the “controller.” GDPR Article 28 requires a binding written contract, a data processing agreement, before that transfer is lawful. No agreement, no lawful basis. Switzerland’s revised Federal Act on Data Protection, in force since September 2023, sets the same expectation. Article 9 requires a written agreement whenever you outsource processing to a third party.
In some fields it goes further. Swiss banking secrecy makes disclosing a client’s information a criminal offense. A machine can’t be held accountable for that.
Read all these rules together and one principle emerges. A confidentiality duty needs a party who can be bound to it and held responsible if it breaks. That’s an expectation a free (and sometimes paid) app cannot meet.
The Trap: Paid Doesn’t Automatically Mean Protected
Here’s the part most people get wrong. They assume that because they pay for a tool, they’re covered. Often they are not.
The paid and enterprise tiers do offer real protection. DeepL Pro deletes your text and will sign a data processing agreement. Other major providers offer zero-retention terms to business customers. But this protection only exists if someone actually signed the agreement, switched on the no-retention setting, and checked where the data is hosted.
So what does adequate protection actually require? The Swiss Bar Association spells out the checklist. These are:
- hosting in Switzerland or the EU
- zero data retention
- a written DPA
- no training on your data
- transparency about sub-processors
The reality is that convenience normally skips every one of those steps. The person pasting into the free or paid box out of habit rarely knows whether any of it is in place.
The Part That Stays Scarce
There’s a pattern I wrote about in Who Keeps the Gold?
When one layer of work becomes a commodity, the profit doesn’t vanish. It moves to the next layer wherever scarcity still lives. AI is doing that to translation now. The words are becoming cheap and near-instant. So what stays scarce?
Not accuracy
The machines have that. What stays scarce is accountability. A professional can be bound by a contract, a professional code, and personal liability. A non-disclosure agreement (NDA) is real. It is enforceable and there’s a name attached to it. You can’t sign an NDA with an app.
The value doesn’t disappear when the task is automated. It migrates to the layer AI can’t hand out and this includes trust, certification, and someone who stands behind the work.
The money doesn’t leave the sector. It moves from producing the words to vouching for them.
Risk Mitigation
In practice, I believe the tension between efficiency and safety can be mitigated and this all comes down to a few disciplined habits.
- Assume the free tier keeps your data. Treat anything you paste as if it may be stored and studied.
- Match the tool to the stakes. For ordinary, low-risk work, the app is fine. For anything that would hurt if it leaked, it is the wrong tool at any price.
- Ask one question before you paste. Would this cause harm if it got out and is there a duty attached to it? If the answer is anywhere near a "yes", it belongs with a person, not a tool.
None of these make AI the wrong choice. For casual, low-stakes translation, no one will ever care where the text went and convenience will keep winning. However, for regulated, sensitive, or client-confidential work, the calculation is not close. In those cases, the regulators clearly agree.
As for me, the position is simple. I’ll keep using AI every day for quick, routine tasks. But when my work touches anything sensitive or confidential, I’ll still call a human translator.
The views expressed here are the author’s own and do not constitute investment advice or a recommendation to buy or sell any security. See the full disclaimer below.