Key facts
- AI translation services are increasingly used for internal documents and product listings.
- Human translators are still needed to check AI-generated drafts for accuracy.
- Errors in AI-translated legal documents can lead to delays in detainee releases.
- Mistranslations in medical contexts, such as dosage units, can be life-threatening.
- AI translators struggle with idioms and context-dependent phrases.
- Future AI translation systems may incorporate self-correction and express uncertainty.
While artificial intelligence has made significant strides in translation, human oversight remains essential for ensuring accuracy, particularly in fields with high stakes such as legal and medical documents. Experts highlight that although AI translators are becoming more sophisticated, they can still misinterpret idioms, context, and nuances, leading to potentially serious errors.
François Yvon, a research director at the French National Centre for Scientific Research (CNRS) specializing in automatic translation, noted that while AI can often correctly interpret common phrases like "it's raining cats and dogs," it may struggle with equivalent expressions in other languages. He also pointed out that AI can sometimes misinterpret words with multiple meanings, leading to nonsensical translations, such as mistaking "translate into Polish" for a reference to shoe cream.
John Zimmerman, founder of the Open Knowledge Association (OKA), which uses AI to help translate Wikipedia articles, stated that AI translation is cost-effective for many use cases, such as internal documents or product listings. However, for an encyclopedia where accuracy and trust are paramount, human review is still necessary. OKA's process involves AI generating a draft, followed by human translators checking every sentence. He observed that the quality of AI output has improved to the point where skipping human review entirely is a tempting option for many in the industry.
Translators like Ariella Baltié, who specializes in legal documents, and Carol Bereuter, a medical translation expert, emphasize the critical nature of accuracy in their fields. Baltié recounted instances where AI mistranslated legal terms like "the requester" instead of "the plaintiff" and misinterpreted marital status descriptions. Bereuter warned of dangerous errors in medical translations, such as confusing milligrams with micrograms, which could lead to a thousand-fold overdose. Both noted that as AI improves, its errors become subtler and harder to detect.
Yvon believes that future translation systems will likely incorporate AI tools that can evaluate their own output and express uncertainty, a key step towards greater reliability. He suggests that a new generation of AI might use multiple tools simultaneously: one for translation, a second for criticism, and a third to generate a revised version. The ultimate breakthrough, he posits, will be when AI can indicate when it is unsure of a translation.