One question people foreign to the translation industry tend to ask me is this: ’Most of the people I know use AI or apps to translate, is what you do really sustainable?’
I have multiple answers to this thorny question.
- Artificial Intelligence (or Large Language Models) doesn’t inherently understand or think; it predicts based on giant datasets. Humans may be fallible in some aspects, but they recognise when they don’t know and do the work (research, asking questions, …). LLMs will hallucinate before they admit fault.
- Some file formats (scanned/faxed documents, desktop publishing, …) can’t be translated by AI or apps as such. They need human manipulation or handling.
- Some subjects are more sensitive than others, some data can’t be shared to cannibalistic apps that’ll mine the data you give them to improve their algorithms.
- The intellectual property issue around AI and AI translations isn’t yet clear; a person well-versed in the tech needs to intervene to protect IP and personal data.
- AI is very resource-intensive. One ChatGPT query could use up to a glassful of water, and it was calculated in 2023 that it consumes as much energy as a 60W incandescent lightbulb uses up in 10 minutes. While human translation also consumes resources, it’s nowhere near the same level.
Over the next five weeks, I’ll take a deep dive into each of these.
1. AI Doesn’t Understand; It Predicts.
Large language models aren’t sentient beings that comprehend meaning. They are sophisticated pattern-matching algorithms. The reason they generate text that seems coherent, and sometimes is, stems from extensive training and the resulting statistical probabilities.
AI models mine vast quantities of data, text, and code from the Internet, books, and other sources. This allows them to extrapolate statistical relationships between words, phrases, syntax, and apply them to the same content in another language. While LLMs are exposed to vast corpora in multiple languages, the less human-generated content there is available online for a language, the poorer the quality of the AI output.
Where a human translator understands a given text and then rewrites it in another language, AI predicts the most probable sequence of words in the target language based on the patterns it has been taught, through constant calculations of the next most probable token. This creates surprisingly similar outputs to human translation, but not quite. That’s where professional translators need to step in and check, adapt, and rewrite until the target text is fit for use in a professional context.
AI has also been known to “hallucinate” details when its datasets aren’t extensive enough to yield results. The more LLMs are trained on AI-generated content available online, the poorer the output. Linguistically, that translates to loss of colloquialisms, of sentiment, of rich syntax inherent to human creation.
So it’s clear: AI may be a useful tool, but it can’t replace HI (human intelligence!) and won’t do so for some time to come…
2. AI Can’t Access Text in All File Formats
While AI, particularly LLMs, has revolutionised translation for plain text, complex or image-based formats present significant hurdles.
AI cannot directly ’read’ an image. It relies on Optical Character Recognition (OCR) technology to convert the image into machine-readable text. Low-resolution, blurry, skewed, or handwritten scanned/faxed documents severely degrade recognition accuracy.
If the layout is complex and the file format is not editable, OCR will ’read’ from left to right and then from top to bottom. Text in columns, text boxes, and line breaks in the middle of words are all hurdles that OCR and, therefore, AI can get very wrong. Where the human brain makes logical leaps, Large Language Models can’t.
If OCR fails to identify characters correctly, structure, or words, the LLM receives incorrect input, leading to wrong or even nonsensical translations.
Scanned or photographed documents often contain stamps, signatures, diagrams, or annotations that are critical to the document’s meaning but are non-textual. AI struggles to interpret and incorporate these visual elements into the translation, requiring human review to ensure all relevant information is captured.
Translators are very much aware of how translated text expands or shrinks depending on the language pairs. When document layouts are slightly less straightforward, this leads to anchors and graphics needing to be adapted to the new text, sometimes painstakingly (translators, I know you know). Machines can’t do this, and also can’t rephrase or shorten text according to certain character limits without losing meaning. This type of ’extra mile’ is one technical translators like myself are particularly aware of: oh, how often do we innerly scream at software developers when our language is one that ’extends’!
Translators can also evaluate ’on the fly’ if certain content is to be translated or not, whereas AI will try to translate everything, leading to proper names, loanwords, and such being translated where they shouldn’t be. It also doesn’t research any topics to create a culturally sensitive, historically accurate translation. It will work one word after another and never have a doubt. Human translators will double- and triple-check sensitive topics to make sure their translation meets the highest standards.
Essentially, AI excels at processing and generating text based on statistical patterns. When the source material is not readily available as clean, structured text or when the translation requires complex visual readaptation or interpretation of non-textual elements, human intervention becomes indispensable. The most effective approach often involves a ’human-in-the-loop’, where AI provides an initial draft or extracts text, and human experts refine, format, and ensure contextual and cultural accuracy.
Come back soon to read part 2.