Why You Should NEVER Use ChatGPT for Sensitive Translations (It's Not Just Privacy)

Everyone focuses on the privacy angle with ChatGPT. And yeah, pasting your NDAs into a chatbot is a terrible idea. We've covered that.

But honestly? That's the easy part to fix. You get an enterprise license, you toggle "don't train," and you feel safe.

The real problem isn't what ChatGPT keeps. It's what it creates.

I was playing around with some legal text the other day—just trying to break things, as one does—and I realized something terrifying. LLMs don't translate meaning. They predict tokens. Most of the time, those two things look the same. But when they don't, you get silent, fluent, confident lies.

If you're using ChatGPT for sensitive docs, privacy is just the first hurdle. The tripwire is accuracy.

The allure of 'free and instant' (and why it's a trap)

I get it. You have a 50-page contract in German. You need to know what it says now. You don't want to email a vendor, get a quote, wait three days, and pay $500.

So you paste it into ChatGPT.

It comes back instantly. It reads perfectly. The English is smooth, maybe even better than what a human translator would write.

That smoothness is the trap.

Traditional machine translation (like old Google Translate) used to sound robotic. When it failed, it sounded broken. You knew to double-check it.

LLMs are designed to sound convincing. When they fail, they don't sound broken—they sound like a confident lawyer who happens to be completely wrong.

The 3 Silent Killers

If you're translating casual emails, who cares. But if you're dealing with sensitive documents, these three things will burn you.

1. Hallucinations (making up facts)

I'm not talking about it inventing a unicorn. I'm talking about it inventing a clause.

There's this thing called the "Translation Barrier Hypothesis." Basically, because LLMs are predicting the next word based on probability, they can get "bored" with a dry legal sentence and spice it up with a phrase that sounds likely but isn't there.

I've seen LLMs:

• Add "not" to a sentence, reversing the liability.
• Invent specific dollar amounts in damages that weren't in the source.
• Hallucinate entire regulatory references that look real (e.g., "according to GDPR Article 99") but don't exist.

2. Inconsistency (changing terminology mid-doc)

Legal documents rely on defined terms. "The Company" means something specific. "The Contractor" means something else.

LLMs have a context window, but they don't have a "termbase" (a database of approved translations) unless you force-feed them one.

In paragraph 1, it might translate Kündigungsfrist as "notice period." In paragraph 40, it decides "termination window" sounds better. To a lawyer, those might be legally distinct concepts. To an LLM, they're just synonyms.

3. Data Leakage (the classic)

Okay, I said I wouldn't dwell on this, but we have to touch on it.

If you are on the free tier, you are training the model. Why not Google or DeepL? goes into the gory details, but essentially: OpenAI's terms allow them to use your inputs to improve the model.

Imagine your M&A strategy appearing as a suggestion in someone else's prompt completion next week. Unlikely? Maybe. Impossible? No.

Why 'human in the loop' isn't enough

"But Yash," you say, "we have a human review the output!"

Do you?

Here's the problem with "human in the loop" for LLMs: The anchoring bias.

When the output looks 99% perfect, the human reviewer's brain switches off. They enter "proofreading mode" (checking for typos) instead of "translation mode" (checking for meaning).

Because the LLM is so fluent, the errors are subtle. A human skims right past the hallucinated "not" or the slightly wrong legal term because the sentence flows so well.

To actually catch LLM errors, you basically have to re-translate the document from scratch in your head and compare it. At that point, you haven't saved any time.

The Liability Gap: Who do you sue?

Let's say the LLM messes up. It translates "indemnify" incorrectly, and you sign a contract that exposes you to millions in damages.

Who is liable?

• OpenAI? Their terms clearly state the service is provided "as is." Good luck suing them for a bad translation.
• The junior associate who pasted it in? Maybe, but that doesn't save the company.
• You.

When you use a specialized translation agency or a dedicated translation service for sensitive documents, there are SLAs. There are warranties. There is a chain of custody.

With an LLM, you are taking 100% of the risk for 0% of the guarantee.

When to use LLMs vs specialized tools

I'm not saying LLMs are useless for languages. I use them all the time. But you have to know what tool to use for what job.

Use ChatGPT when:

• You're brainstorming marketing copy in another language.
• You need to summarize a long foreign article to see if it's relevant.
• You're learning a language and want to practice conversation.
• The stakes of being wrong are zero.

Use specialized document translation when:

• It's a contract, medical record, or technical manual.
• You need formatting preserved perfectly (tables, headers).
• You need a "paper trail" of data handling.
• The document has legal or financial consequences.

Takeaways

It comes down to this: LLMs are creative engines. Translation is often an exercise in restraint.

You don't want a "creative" translation of your patent application. You want a boring, accurate one.

If you're relying on ChatGPT to negotiate a deal, you're not just betting on the translation. You're betting that a probabilistic token predictor didn't get creative with your liability clause.

Seems like a bad bet.

Further reading

• Which translation services should I use for sensitive documents?
• Why not Google or DeepL?
• How your data is handled
• We don't want your data