AI translation produces fluent text that looks right and often is not. In regulated content, the failures are not random: they are specific and they repeat. This is the map of the most common errors in machine and AI translation, and of what human review does to catch them before they reach the reader.

If you work in life sciences, medical devices, pharma, technical or IT content, the risk is not theoretical. A manual, an instruction for use, a clinical document or a safety data sheet that fails in the target language can compromise compliance, user safety and customer trust. AI speeds up the work, but it does not take responsibility for the result. That is still human.

Where AI translation goes wrong

Omissions

AI can drop content without warning: a sentence, a clause, a negation or an item from a list. Because the text still reads smoothly, the omission goes unnoticed unless it is compared against the source. In an instruction for use, a missing sentence can be a missing safety instruction.

Terminology

AI picks plausible terms, but not always the correct or approved ones. It ignores glossaries, translation memories and the regulatory terminology of the sector. A term the reader understands is not the same as the exact term required in a medical, technical or legal context.

Locale and culture

Beyond the words, there are conventions: date, number and unit formats, forms of address, cultural references and the right variant, European or Brazilian Portuguese for example. AI misses these details, and the text ends up reading as foreign or ill-fitting for the target market.

Factual inconsistencies

AI can alter numbers, doses, units or names, and it can add information that is not in the source. The result is a confident but false statement, the most dangerous kind of error, because it raises no suspicion in the reader.

Polarity and directional reversals

AI can reverse meaning: turn a negative into a positive, or flip the sense of an instruction, with "do not" becoming "do." In a safety warning or a clinical indication, that reversal changes everything.

What the linguist actually checks

Human review is not a cosmetic touch-up. The linguist checks the text against the source, against the glossary and against regulatory requirements, verifies consistency across the whole document and weighs the risk of each choice. This is the work AI cannot perform on its own output.

Admissibility

In legal and regulatory contexts, AI output may not be admissible before authorities, notaries or certification bodies. A certified translation requires the responsibility and signature of a professional, something a machine cannot assume.

Reproducibility

The same sentence can produce different translations on each run. This variation undermines consistency, complicates auditing and version control, and makes it hard to guarantee that a set of documents speaks with one voice.

Confidentiality and data protection

Sending content to public AI tools can expose confidential information and personal data, and breach non-disclosure agreements and the GDPR. The data can be retained or used to train the models, outside your control.

The cost that shows up later

The upfront saving is misleading. The rework to fix errors, the problems with file and format integrity and the time of your internal teams all add up. Once you count the total cost, the cheapest option at the outset tends to end up more expensive.

What human supervision adds

At L10N, AI and machine translation have their place, but always with human supervision. Our linguists translate only into their native language, have at least five years of experience in their field, and every project runs through our ISO 17100 certified process: translation, revision and a final quality check by three different professionals. When we use machine translation, we always add human post-editing, so the content is reliable and ready to publish. A dedicated translation memory and glossary per client keep your terminology consistent across everything you send us.

Need volume and speed without giving up reliability? That is exactly what machine translation with post-editing offers: the speed of the machine with the judgment of a professional. Learn more on our Machine Translation page.

Frequently asked questions

Is AI translation safe for medical or regulated content?

Raw, no. AI produces fluent text but makes specific errors, such as omissions, meaning reversals and imprecise terms, that in regulated content can have safety or compliance consequences. With human post-editing, it can be a suitable option for certain content.

So is machine translation useless?

No, in the right context. For high-volume, lower-sensitivity content with short deadlines, machine translation with post-editing is effective. What you should not do is publish AI output without human review in critical content.

Can I use public AI tools with confidential documents?

Not without care. Many public tools may retain the data or use it for training, which raises confidentiality and GDPR risks. A professional workflow ensures information is handled securely.

What sets post-editing apart from a simple review?

Post-editing starts from the machine output and corrects it against the source, the terminology and the requirements of the sector. It is specialist work, which calls for knowing both the typical errors of AI and the subject matter itself.

Working with content that cannot fail? Tell us about your project and request a quote. We will reply quickly with a clear proposal.

Speaking to the world with confidence is easy!

1. Request a quote 
2. Validate the terms of the proposal 
3. Receive accurate translations 
You are using an outdated browser. Please upgrade your browser to improve your experience.