AI Translation Dominates, But Human Oversight Remains Crucial for High-Stakes Content

Machine translation has become a dominant force in the translation industry, handling trillions of words monthly across platforms like Google Translate, DeepL, Microsoft, and others. The direct sales of machine translation services reached an estimated $1.4 billion last year, reflecting its widespread adoption for use cases where cost-effectiveness outweighs the need for absolute precision.

For many applications, such as translating internal documents or product listings, AI-generated translations are deemed sufficient. John Zimmerman, founder of the Open Knowledge Association (OKA), a non-profit that translates Wikipedia articles, notes that for such purposes, "raw AI translation is now the right call: multiplying your costs by 100 to catch the last few percent of errors isn't rational." OKA itself relies heavily on AI to produce draft translations of Wikipedia articles, a process that would be prohibitively expensive if done manually.

However, the reliance on AI falters when accuracy is paramount, particularly in fields like law and medicine. Legal expert and translator Ariella Baltié recounted instances where AI mistranslated legal terms, such as rendering "Le demandeur" (the plaintiff) as "the requester," which is inappropriate in legal contexts. Another error involved misinterpreting "Madame X, épouse Y" (Ms. X, wife of Y) as "Ms. X marries Y," confusing a noun with a verb and altering the meaning entirely. In the United States, errors in machine-translated asylum applications have reportedly led to delays in the release of detainees.

Similarly, in the medical field, precision is critical. Carol Bereuter, a specialist in medical translations, described an instance where a Greek letter used in dosage units was mistranslated. The symbol μ (micro) was converted to "milli" (one thousandth), a thousandfold difference that could result in a dangerously incorrect dosage. Bereuter also observed that as AI translators improve, their errors become subtler and harder to detect, as the output often sounds fluid and natural, masking grammatical or contextual mistakes that were once more apparent.

François Yvon, a research director specializing in automatic translation at the French National Centre for Scientific Research (CNRS), highlighted the challenge AI faces with context and nuance. Idioms, such as the French "Il pleut des cordes" (It's raining ropes) for heavy rain, can perplex AI systems, which may struggle to find the appropriate equivalent in another language. He noted that even seemingly clear sentences can be misinterpreted if the AI misunderstands the context, citing an example where the phrase "translate into Polish" could be mistakenly interpreted as relating to shoe cream.

Zimmerman's OKA study of 119 articles found that 27% of AI-generated text still required corrections before publication. Of these, three-quarters were errors introduced by the AI, with 5-6% altering the original meaning. He suggests that the probabilistic nature of current AI translation systems makes achieving 100% reliability difficult.

Yvon believes that future translation systems will likely incorporate self-correction capabilities, using multiple AI tools simultaneously for translation, criticism, and revision. However, he posits that the true breakthrough will occur when AI systems can not only translate but also indicate their level of uncertainty, signaling when human review is most critical.