These days, AI hallucinations are common and expected. We love to spot mistakes and criticize celebrities, politicians, and brands trying to get away with AI edited (or created) images.
However, when it comes to using large language models for translation (especially when we don’t speak the target language), it’s easy to forget these same risks apply.
AI Mistakes in Translation
Large language models are the technology behind AI translation tools like Google Translate, DeepL, and Amazon Translate.
Hallucinations in AI translation are far less funny than a celebrity gaffe. AI hallucinations are statements that are presented as fact, but are actually completely made up, misrepresented, or untrue. Publishing this content unedited can create significant risk for organizations.
AI tools prioritize delivering an answer (even if wrong) over clarifying uncertainty. As such, mistakes resulting from hallucinations in AI translation are common, such as:
- Content present in the translation, but that doesn’t appear in the source file
- Citing nonexistent publications or research
- Inaccurate representations about real people, such as age, birthday, or accomplishments
A minor hallucination may result in an incomplete translation with some content left untranslated, or incorrect word choice due to ignoring context clues. However, a major hallucination can cause real harm, such as reversed do vs. do not guidance due to AI misreading sentence syntax.
Imagine if a health center was using AI alone to support a patient’s discharge after a procedure. What would happen if the tool translated a prescription label to instruct the patient to take medication on an empty stomach, when the instructions actually said it should be taken only with food? Repercussions could range from a person feeling uncomfortable side effects, all the way to a malpractice lawsuit.
Mistakes resulting from hallucinations are often compounded when using multiple levels of AI, such as AI voice matching interpretation or AI translated captions.
The quick and seemingly nuanced output of these tools makes them tempting to use, but be aware of the high risk of flawed output given the opportunity for errors at every step.
Human in the Loop: Refining Machine Translation Since the 1940s
While machine translation tools have been popularized in the last few years, studies into the technology have actually been going on since the early 1940s. For nearly a century, research has concluded the same thing – a human in the loop is critical to have confidence in translation quality and accuracy.
Given the hallucinations of today’s technology, HITL is more important now than ever.
HITL Translations
A human in the loop translation process allows organizations to take advantage of the speed afforded by AI and repetitive content matching available in translation memory while also allowing for critical checks.
Human in the loop translation includes checks and reviews for:
- Completeness, word choice, accuracy, as well as spelling and grammar in a translation
- Cultural appropriateness
- Flow and consistency
- Tone and formality
- Ambiguity or determination of idioms make sense given the audience and context
Alboum’s human in the loop process results in natural-sounding translations that are accurate and hallucination free.
Please contact us to request a quote and discuss how we can offer you an affordable option to minimize the risk of costly errors resulting from AI hallucinations.
