Last week, Meta released SeamlessM4T, a model that translates speech and text across nearly one hundred languages. A few weeks earlier, a different tool was announced. Each month brings a new translation tool presented as the end of language barriers.
We follow these developments closely because of our work in diving. Our training extends to the Master Instructor level, we have supported dive operators with their digital and social media presence, and since 2015 we have been partners in Ceningan Divers. Diving is one of the most multilingual small industries we know.
On an ordinary morning, a dive boat off Nusa Penida may carry divers from half a dozen countries. The briefing is usually delivered in English. The training materials, waivers, medical questionnaire and pre-arrival emails, however, must all work in many languages. For a small operator, this has always been a challenge. The tools are now good enough that the central question has become what they should be allowed to translate.
What is the current state of machine translation?
Machine translation is not new. Rule-based systems date back to the 1950s. The major advance came when machine translation shifted to neural networks in the middle of the last decade. Google Translate moved to a neural system in late 2016, and DeepL launched in 2017 with translations that sounded noticeably more natural in European languages.
Large language models followed. GPT-4 can translate, but it can also adjust tone, explain a choice, maintain a formal or informal register and adapt text for a local audience rather than converting it word by word. For marketing copy, this is a significant advance. A request for Quebec French rather than European French is generally respected.
For a dive business, the drafts are of good quality, good enough that many people are beginning to trust them more than they should.
Where should AI translation be used, and where should it not?
For drafts, yes; for safety information on its own, no. This is the boundary we recommend, and it should be applied firmly.
Marketing pages, social media posts, replies to booking enquiries, a "what to bring" list or a blog post about manta season can all be drafted by AI, ideally followed by a review from a speaker of the language. If a word is slightly inaccurate, no one is harmed.
Safety information is different: dive briefings, emergency procedures, the medical questionnaire, and anything concerning decompression, depth limits, equipment checks or the procedure to follow when separated from a buddy. A single incorrect word in this material is not merely embarrassing; it is dangerous.
The errors machine translation produces are precisely of the dangerous kind. The output is fluent and reads well, so a sentence that reverses a meaning, omits a "not" or turns "must" into "should" easily escapes anyone who reads quickly. A diver reading in their own language has no reason to doubt it.
For this material, a qualified person with knowledge of both the language and diving must review every line. In many cases, the training agency already publishes official translations of its safety materials, and those should be used rather than in-house translations.
Why maintain a glossary of fixed terms?
Because it ensures consistency across every language. The most useful measure a multilingual business can take with AI translation is also the least remarkable: building a glossary.
Every industry has terms that must be translated the same way every time. In diving, these include equipment names, skill names, course names, certification levels and many terms with an official version in each language. In any business, they include product names, the brand, legal terms and agreed-upon wording. A B&B in Quebec, for example, might decide once and for all whether its English pages use the word "gîte," and apply that decision consistently.
Most professional translation tools allow a glossary to be loaded so the machine follows it; DeepL offers this feature. Even with a basic chatbot, the list can be pasted at the beginning of the request. In either case, we recommend writing the glossary once, sharing it with everyone who writes for the business and updating it whenever something changes.
Although it resembles administrative work, it is what prevents a French page from referring to the same product by three different names.
What should be automated in a translation workflow?
The surrounding steps, not the judgement. When people speak of automating translation, they often imagine a button that produces six languages at once. That is the part we would automate least.
What consumes time in a multilingual business is everything surrounding the translation itself:
- copying text out of the website and pasting it back in
- tracking which language versions are outdated after the English text changes
- sending the correct draft to the correct reviewer
- formatting, file naming and publishing
These tasks are well suited to a script, a plugin or a workflow tool. A change to the English page flags the other languages for review. A draft is generated with the glossary applied. It reaches a reviewer's inbox. The reviewer approves or corrects it, and it is published.
The person remains exactly where judgement matters: reading, deciding and approving. Everything else proceeds automatically. This is the type of automation we consider trustworthy for a small business.
Why does better translation make verification more difficult?
These tools will continue to improve; next year's models are expected to translate better than this year's, and for everyday business text they are already frequently very good.
This does not change the boundary we recommend. The better the output reads, the easier it becomes to overlook the one incorrect sentence. With older, awkward translations, errors were apparent. They no longer are, and in safety material, a single sentence is enough to cause harm.
Diving operates in dozens of languages, and safety wording cannot be approximate in any of them. The same applies to many areas beyond diving: medication instructions, contracts and the rules posted at a lakeside campground.