Google's translation service processes one trillion words monthly. François Yvon, a research director at the French National Centre for Scientific Research (CNRS), discusses the challenges AI faces with idiomatic expressions, such as the French phrase "Il pleut des cordes," which translates literally to "It's raining ropes." He notes that while AI can recognize common idioms in some languages, it may struggle with others, like Indonesian.
In April, Google announced that its translation service is handling one trillion words each month, not accounting for other services like DeepL, Microsoft, Claude, ChatGPT, and Gemini. The direct sales of machine translation services were valued at approximately $1.4 billion in the previous year.
John Zimmerman, founder of the Open Knowledge Association (OKA), states that AI translation is suitable for many applications, such as internal documents or product listings, where the cost of human translation may not be justified. He emphasizes that for high-stakes content, like encyclopedic entries, accuracy is crucial. OKA, which has translated over 10,000 articles primarily into Spanish, Portuguese, and English, relies on AI to draft translations that are then reviewed by human translators.
Zimmerman notes that the quality of AI-generated translations has improved significantly, reducing the need for extensive human editing. However, Ariella Baltié, a legal translator, highlights the potential for serious errors in AI translations of legal documents, citing examples where AI misinterpreted terms, leading to awkward or incorrect translations.
In the medical field, accuracy is critical, as Carol Bereuter points out. She recalls an instance where an AI translation confused the Greek letter μ, leading to a potentially dangerous dosage error. Bereuter also mentions that as AI translations become more fluent, spotting errors has become increasingly difficult.
François Yvon warns that context is essential for accurate translations, as AI may misinterpret phrases without understanding the surrounding context. A study by OKA found that 27% of AI-generated text required corrections before publication, with a significant portion of those corrections altering the meaning of the original text.
Yvon believes that while AI translation systems are improving, achieving 100% reliability remains a challenge due to the probabilistic nature of AI. He anticipates advancements in technology that will enable translation systems to self-evaluate and correct errors. The future breakthrough may involve AI systems that can express uncertainty in their translations.