Just send it to the LLM and ta-da! Translations are done! Hassle free, worry free and perfect.

With all the wonders and hype around AI, many execs treat language as a "solved problem". Is that really the case?

Caught in the whirlwind of the AI hype, it’s natural to question, why do we need to worry about translations at all. And if translations are reduced to a simple input output mechanism, we would need to worry about managing them either?

1. On the Nature of Language

LLMs are not that different from highly sophisticated Neural Machine Translation in terms of translation quality.¹ At times a bit more nuanced, at times less predictable, and most often computationally more expensive, the big difference with LLMs is in terms of public perception. They awaken fear, awe, and other big human emotions contrary to Machine Translation that was just attempting to solve an isolated problem.

And when we take a closer look at how an LLM behaves in terms of translation, its power and limitations become clear even in the simplest of interactions. Take a look at the example below.

I asked the GPT-4o to translate a sentence into Portuguese.

Before in going into the text itself, the model didn’t ask me whether I was looking for Brazilian Portuguese, or Portuguese from Portugal. It didn't bother to ask me for more context around the text itself, or my audience. It just executed the command.

At first glance, this looks like an accurate translation but a quick read shows it treated Bureau Works using the feminine article A instead of the masculine article O. It decided for me that Bureau Works in this context was the company Bureau Works and not the software Bureau Works.

One article, two identities. Machines don’t always know the difference.

This isn’t a technical mistake. This is a stylistic decision that underpins the entire way we position ourselves as a brand. Whether we decide to position ourselves as a company or a system has deep ramifications for how our users understand us. It may be disguised as a grammatical ambiguity but it quickly evolves into a matter of identity.

And that is the nature of language. It’s not merely descriptive. It’s relational. It’s about how we identify ourselves with our own selves and our world and each and every choice is chiseling at an overall sense of self, sense of belonging, sense of direction.

So language is much more than a matter of accurate or inaccurate representation. Language is a form of art that is sculpting the world with tiny decisions and if these decisions are no longer managed by humans deeply embedded in a micro-cultural context, then that means that there are no longer drivers at the wheel of language. It means culture is no longer a human thing.

2. Post-human Culture

Ok, so we’ve established that language is not just about good enough representation. And we’ve established that decisions left only to machines make culture no longer a human thing. But it seems clear as well that as a global society we are long past the point of thinking about culture as only a human thing. Machines are an undeniable part of our social fabric, in the form of search, algorithms, apps, models. Thinking that we can go back to a human-only society is a denial of the past 10,000 years of human history. It’s naive and counterproductive.

What we can do is navigate into a post-human Culture that still has humans at the wheel.

3. Managing Human-Machine Interactions

Models while incredibly powerful, are not really specialized at any specific task other than guessing the next token. When you dive deeper into the idea of translation, the limitations of working directly with a model become pretty clear.

A. It’s hard to teach a model

If I told the model in the example above that I wanted to treat the company as a system, it would understand my instruction and produce the correct result in that specific case.

But if I continue with more and more interactions in this chat, the likelihood of it departing from this instruction becomes greater and greater because this command will become buried in a bigger and bigger sea of vectors and embeddings and it will become less of an imperative.

And if I begin to work with bigger chunks of text it’s more likely that I will arrive at exceptions. There will be cases where I do want to treat Bureau Works as a company and those cases are likely to be mishandled either because it followed the instruction to literally or because it’s output was erratic. Either way the key thing that gets lost is control. As the reader, or verifier, I can fix these issues, but it becomes really hard for me to predict how it will behave in cases to follow.

B. The Positive Legacy from Translation Management Systems

If there’s one thing translation management systems have done well over the years, it was consolidating a framework that favored predictability. Translation memories (relational databases that stores and retrieves previous translation sentences) and Glossaries (relational databases that stores and retrieves terminology) have worked hand in hand to ensure that the large surge of volume from the latest wave of Globalization starting in the 90’s was taken care of.

These relational frameworks were really good at following strict commands. If something was translated in a given way and stored, you could be 99.999% certain that it would be translated in that given way again and again. The problem with these frameworks is that they are purely syntactical and require a whole lot of human output management to become useful.

For instance, if I had previously translated the sentence,

“Bureau Works is an advanced translation platform that helps organizations make use of AI efficiently and securely.” that would fall into a 70-79% match range and let’s say I had certain terms in my glossary such as “platform” being translated as “system”, it would be up to the translator to fill in the blanks, to comb through the linguistic corpus and make decisions that they could live with.

This framework laid out the bones for something important: at scale language needs to facilitate consistency. With many people working on the same assets for a single brand, over a long period of time, this framework made consistency at scale possible and economically feasible.

C. What LLMs can Do and Can’t Do

An LLM can easily do this heavy lifting, reducing human edits by at 39% and cutting down on edit time by 50%.² But they need the right tools and framework to do so. If you try submitting a document with a few pages to an LLM, a database with past translations and another one with terms to be followed you can sit back and watch them fail horribly at that.

I am trying to translate a 10 page document with some additional context provided in the form of these relational databases.

It hesitates, but I tell it to push on.

It tells me it does the job but here’s the document it generated:

It tells me it’s translated but everything is still in English with “PT:” appended to the start of every paragraph. It may be possible for me to prompt-engineer this model into oblivion and get the job done, but I am just a regular user who needs a document translated. I don’t have time for this crap.

And ultimately, even assuming that I would have infinite time and resources to get the job done, this would mean I would have to vibe-code my own Translation Management System.

4) Enter Bureau Works

If I were to do the same thing in Bureau Works, I can upload the file

I can translate it respecting all of my external context.

I can predictably download the translated asset.

But most importantly I can edit the document in an environment that will truly take note of my edits and leverage them in the future, not just for me but for as many people that I want to include as a part of my team.

Bureau Works is a tool that is specifically designed to facilitate the collaboration between humans and machines when it comes to translation. It’s a very specific thing. And in this context, it’s designed to work with small bits of text or thousand page documents. It’s designed to work with code or with scanned PDFs. It’s versatile and predictable within a very limited scope.

5) The Temptation of a Single Tool to Rule them all and Hidden Dangers of Work-Arounds

LLMs contrary to Bureau Works can do pretty much anything. Or at least they feel this way. They can translate, analyze a legal document, provide input on a critical decision, or comb through data. As such they evolve into a go-to tool for all kinds of things.

And our brains, adaptive as they are, love that level of convenience. So even when they don’t produce the results we need, we tend to work within its set of limitations. A good example is someone translating slides, may choose to copy and paste back and forth every single text box into an LLM so as not to preserve formatting.

They’ll get the job done with the illusion of simplicity but all of those minutes copying and pasting text will never come back. And all of those edits done after the translation is saved won’t be leverageable by anyone in the future. It’s pure wasted effort that becomes disguised as getting things done with a model.

6) File types and flexibility

What about when you need to translate an HTML web page, or a JSON piece of code, or an elearning course. By now it should be pretty clear that an LLM’s strength doesn’t lie in parsing files and making them compatible with translation processes. But that’s a mechanical issue. And it can be arguably worked around with pre and post processing mechanisms that can ostensibly bypass the need of a translation management system even though now you have the overhead of managing some kind of middle ware to enable your LLM interactions.

7) Economic Considerations

This also has economic implications as well. How much is it worth to me being able to decide whether or not we position ourselves as a system or as a company to our users. Some will argue that this is worthless. Others that this is priceless. And while I am not here to make the case of how valuable this is. You can decide that you need to hire a brand agency to make a decision worth tens of thousands of dollars or you can leave it to GPT to decide for you for a few tokens.

It’s a highly contextualized thing, in my opinion. Some situations will warrant deep study and reflection. Others may require only a few tokens. My only opinion here is that you do need humans to make these value calls. My second consideration is that these humans need the right software to be able to make these calls, follow through with them and ensure that they are stored in a way that’s meaningful and retrievable in the future. Otherwise, everything is simply left up to chance and we’re back to not having a driver at the wheel of that micro-culture.

A side note on the driver at the wheel metaphor. With the rise of autonomous driving, it may not seem that bad to not have a driver at the wheel. But contrary to autonomous driving where decisions to turn left or right don’t have many considerations other than the safety of all parties involved, deciding whether to turn left or right with text, is one chisel at a time sculpting an entirely different cultural landscape.

8) The Heart of the Issue

Let’s go back to the beginning of our discussion. Language is a matter of identity. It’s personal and it has ramifications that extend far beyond simple grammatical choices. Language needs owners, and owners need to work in environments where they can leverage the good stuff that LLMs bring to the table and also leverage the great stuff that they bring into the table. The little stylistic decisions here and there, the terminological choices, the silent acquiescence for certain things that should remain unchanged. Editing is not an act of correction. Editing is an act of ownership.

Bureau Works is fundamentally designed to enable that level of ownership. And it’s not just ownership in a single slice of time, it’s past present and future ownership over a body of language that may be specific to a brand, a product, or a movement.

9) Where Translation Becomes Content Creation

In this spirit of ownership, we understand that translation shifts from the end goal to the starting point. And in that spirit, our Free Flow editor enables users to release themselves from the shackles of the source and explore text in whichever way they feel they most need.

I can begin with a simple to paragraph source in English for instance and translate it into German:

But then end up with a short Linkedin Post as one of my deliverables

And all of this can get done in an environment where my edits matter. Where they continuously inform my databases whether vector or relational, where these databases will then inform future decisions.

It’s not about correctness. It’s not about humans or machines.

It’s about exploring the best machines have to offer but still emerging as owners on the other side of this rabbit hole.

And you need the right software, the right people and right processes to do that.

1. ChatGPT Vs. Machine Translation: https://www.bureauworks.com/docs/chatgpt-vs-mt

2. Findings on Context Sensitivity: https://www.bureauworks.com/docs/context-sensitivity-efficiencies