Tha obair a' moladh a' bhan-cheàrd

When ChatGPT Translation Anticipates Instead of Translates

With my students turning sixteen this school year, we have started discussing something that feels distant to them but is coming faster than they realise—the transition to adulthood. As part of their Transition Planning, we talk about what comes next: work, further education, independence, and the choices they will have to make. For many, this is the first time they’ve been asked to think about their futures in a concrete way, beyond vague notions of “getting a job” or “moving out one day.” It is an unsettling conversation, because until now, much of their world has been defined by external expectations—by what teachers, parents, and school systems tell them to do. But adulthood will demand something different. It will ask them: Who are you? What do you bring to the world? How do you want to be seen?

To get them thinking about these questions, I asked them to consider their reputation—not the fleeting reputation of popularity, social media presence, or how their peers see them today, but the reputation that will follow them into adulthood. What will their work say about them? How will others judge them when they are no longer in a structured environment where second chances are built in? It is an uncomfortable but necessary realisation: what you do now lays the foundation for how others will see you in the future.

This is something I learned early from my grandmother. She would remind me of an old Scots Gaelic saying: Tha obair a' moladh a' bhan-cheàrd—The work praises (or recommends) the craftswoman. Her message was clear: my work should be my calling card. People should look at what I have done and want to hire me or work with me based on that alone. It was a lesson that shaped my entire approach to life. What I create, what I leave behind, is what speaks for me. I don’t need to declare my worth or argue for my abilities—my work will do that for me. And if my work is careless, rushed, or half-finished, that too will be a reflection of me.

This is the lesson I wanted my students to sit with. In this moment, at sixteen, many of them are still figuring out their place in the world, and it’s easy to dismiss schoolwork, jobs, and responsibilities as just another set of hoops to jump through. But I asked them to think differently. What if the work they do now is already shaping the way people see them? What if their teachers, employers, and even their younger selves will one day look back and see a pattern—not just of what they did, but how they approached it?

It’s not an easy shift in perspective. Many of them have spent years just trying to get through school, navigating systems that were not built for them. For some, the idea of putting in effort when past effort has gone unnoticed or unrewarded feels pointless. But I asked them to consider this: if their work is how they will be judged, what do they want it to say? Not just to others, but to themselves.

Because one day, they will look back at what they built in these years, and whether they realise it now or not, it will already be telling a story about them.

A Case Study: Translating Scots Gaelic to English

As part of our discussion on reputation and the way our work speaks for us, I decided to bring in a phrase from my own background—one that had shaped my thinking from a young age. I wanted to present my grandmother’s words in their original form, in Scots Gaelic, before translating them into English. Since many of my students speak more than one language, this was also an opportunity to model something I do frequently in my teaching: breaking down meaning across languages in a way that preserves intent, not just structure.

The phrase I wanted to share was:

“Tha obair a’ moladh a’ bhan-cheàrd.”

Before presenting it, I wanted to refine how I introduced it. My students often benefit from phonetic breakdowns of new words, so I sought help with IPA (International Phonetic Alphabet) transcription and an English phonetic approximation. This is something I do often when working with my multilingual students—mostly Spanish/English speakers—since pronunciation can be just as important as meaning when making a language accessible.

In my mind, the expected English output was clear:

“The work praises the craftswoman.” (or possibly, “The work recommends the craftswoman,” depending on how one wanted to nuance “moladh.”)

But what I received instead was:

“The work praises the female smith.”

At first glance, it was a minor difference—ban-cheàrd had been rendered as female smith rather than craftswoman. But that slight shift in meaning was significant. It wasn’t a direct translation. Instead, it was anticipatory—a choice made not based purely on the word’s meaning, but on what the AI predicted I meant based on historical patterns. In some contexts, ceàrd can refer to a smith, especially given how traditional Highland craftspeople (particularly Travellers) were often metalworkers. But my intent was “craftswoman,” not “smith.” The AI had guessed, rather than translated, and in doing so, it had imposed meaning that wasn’t there in my original request.

This moment resonated with me because I’ve been navigating similar challenges in my work with Spanish. Recently, I’ve noticed that some of my Spanish translations have been off—not incorrect, but failing to reflect the nuance of dialect and region. Spanish, like all languages, is not monolithic. Word choice, connotation, and even syntax can shift dramatically not only between countries—Mexico, El Salvador, Guatemala—but also between regions and dialects within those countries. The Spanish spoken in Mexico City differs from that of Oaxaca; the Spanish of San Salvador is not the same as that of a small town in western El Salvador. Even within the same country, pronunciation, vocabulary, and phrasing can carry distinct regional markers, shaped by history, migration, and Indigenous language contact. We readily recognise these differences in the United States—Boston English is not Dallas English, which is not Miami English—but beyond our borders, we often fail to extend the same awareness. There’s a tendency to treat languages like Spanish as monolithic, flattening out the richness of regional and dialectical variation, even as we acknowledge it in our own. It’s something I’ve had to be more mindful of, refining my prompts and approach as I learn—not just to ensure accuracy, but to show my students that their language, in all its variations, is seen and valued.

Thankfully, my students have been endlessly patient with me, showing a kind of grace that reminds me why this process matters. As a Gestalt Language Processor (GLP), I experience language differently myself, often absorbing meaning in chunks rather than discrete pieces, which makes learning new languages a challenge in ways I can’t always articulate. But the effort matters, and my students recognise that.

This is why I wanted to be precise in working with this important Scots Gaelic phrase. If I am asking my students to reflect on the message their work sends, then I owe them the same care when presenting mine. And yet, even with a direct request, I encountered an example of how AI translation is not neutral—it predicts, it anticipates, and in doing so, it sometimes distorts.

This is something I want my students to think about as well. Language is never just words—it is meaning, it is culture, it is history. And when we allow predictive systems to shape how we understand one another, we must be aware of the assumptions they bring with them.

Why Did ChatGPT Get It Wrong?

So why did ChatGPT get it wrong? Why did it give me a translation that anticipated rather than directly reflected my prompt? The issue wasn’t a simple mistranslation but something deeper—an interaction between historical bias, predictive modelling, and the power dynamics between marginalised and dominant languages.

First, there’s historical and contextual bias at play. The word ceàrd has historically meant craftsperson, but in more recent usage, particularly in dominant Scottish cultural narratives, it has become closely associated with metalwork and smithing. That shift in meaning didn’t happen in a vacuum—it evolved alongside broader cultural attitudes, including biases against Travellers, who have often been associated with itinerant metalwork and trade. ChatGPT, unaware of my personal heritage or the linguistic nuances of West Highland Scots—and how our interactions with Gaelic differ in many significant ways from dominant Scottish culture—made an assumption based on the majority’s usage rather than my actual words. But I was deliberately avoiding that association. There is already enough hostility towards the Traveller community across the UK and Ireland, and I don’t want to contribute to it in any way. I have no ill feelings towards Travellers, regardless of where they live, and I certainly don’t want my choice of words to reinforce a bias that has been imposed on them rather than one they chose for themselves.

Then there’s the issue of predictive modelling vs. literal translation. ChatGPT didn’t just translate—it predicted what it thought I meant, based on statistical patterns. This is not a quirk or a bug; it is fundamental to how large language models work. GPT is not a dictionary or a simple lookup tool—it is a predictive system designed to generate the most statistically likely response based on patterns from vast amounts of text. That means it doesn’t just retrieve translations; it anticipates meaning, pulling from broader historical usage rather than adhering strictly to the words as given. In many cases, this makes AI-generated text seem remarkably fluent, but in translation—especially between a marginalised language and a dominant one—it introduces significant distortion.

Which leads to the last, and perhaps most critical, issue: the imbalance between marginalised and dominant languages. When translating into English, a language with vast historical documentation and cultural dominance, AI models pull from majority-held narratives and historical biases rather than respecting the intent of the speaker. This is especially pronounced in cases where the meaning of a word has evolved differently across dialects and subcultures. ChatGPT’s training data is shaped by what is most commonly written and recorded, which means it is heavily skewed toward the dominant group’s version of the language. The minority experience, the personal interpretation, and the nuances that exist outside the mainstream often get flattened or erased.

This is not just a problem for Scots Gaelic—it happens with all marginalised languages. When a language is minoritised, its speakers often carry cultural and historical knowledge that AI cannot access, because that knowledge is embedded in lived experience rather than in the large-scale text corpora that models are trained on. When ChatGPT translated ban-cheàrd as female smith, it wasn’t simply substituting one word for another—it was reinforcing a historical and cultural perspective that was never mine to begin with.

This is why AI translation requires critical awareness. It does not simply mediate between languages—it mediates between histories, power structures, and entrenched biases. And for those of us working across languages, especially those with histories of suppression, that awareness is essential.

The Risks of Predictive Translation for Marginalised Languages

The risks of predictive translation for marginalised languages go far beyond simple mistranslations. AI models don’t just translate words—they shape meaning, and in doing so, they can erase nuance, reinforce colonial biases, and undermine linguistic autonomy in ways that are often invisible unless you know what to look for.

One of the biggest dangers is erasure of nuance. Predictive translation relies on statistical likelihood rather than speaker intent, meaning it often defaults to outdated, stereotyped, or even harmful interpretations (which is why it always defaults to a medical model view of autism). Instead of reflecting the richness of a language, it flattens it, reducing complex ideas to what is most commonly assumed to be correct, rather than what actually is. For marginalised languages—many of which have already endured centuries of suppression—this is yet another way their depth and flexibility can be lost. A language is more than its dictionary; it is shaped by its people, their histories, their struggles, and their ways of understanding the world. When an LLM (Large Language Model) ignores that nuance in favour of the most statistically probable answer, it is not just mistranslating—it is misrepresenting an entire culture’s way of expressing itself.

Then there is the issue of colonial and linguistic biases. LLMs are trained within existing power structures, and dominant cultures and languages have a far larger presence in their datasets than minoritised ones. This means that LLM translations will always skew toward dominant-language interpretations, reinforcing the colonial framing of minoritised languages rather than preserving how those languages actually function. But this isn’t just a passive issue of linguistic drift—it’s an active, systemic problem rooted in how these technologies are being developed. The major LLM companies are capitalists, working within capitalism, and fighting like the Jack Welch acolytes that they are in their race to monopolise the ChatGPT space. Their interest is not in linguistic preservation, but in market control, and in that race, anything that is not profitable—like the protection of minoritised languages—will always be an afterthought. AI translation is not an act of neutrality; it is an extension of the capitalist, colonial structures that already dictate whose voices are amplified and whose are ignored.

The most insidious risk, however, is the loss of linguistic autonomy. If AI translation continues to be used uncritically, marginalised language speakers may find their words misrepresented in ways that shape external perceptions—not just in real-time translation, but in how their languages are understood at all. And this isn’t just Orwell’s Newspeak—it’s more subtle, more pervasive. This is a form of Neuro-Linguistic Programming (NLP) at a societal level, and the most dangerous part is that most people won’t even realise it’s happening. Unless they know how to detect it, unless they actively safeguard against it, their language—and by extension, their culture—can be reshaped without their consent and without their awareness.

This is precisely why this discussion matters. It’s not just about one translation gone wrong. It’s about the broader implications of what happens when we let AI models mediate language, meaning, and history without understanding how they do it. And if we don’t challenge these systems now, we may find that by the time we do, the damage has already been done.

How to Approach LLM Translation with Caution

So, how do we approach LLM translation with caution? If we know that these systems predict rather than simply translate, that they reinforce dominant narratives rather than preserve intent, and that they privilege statistical probability over linguistic autonomy, then how do we engage with them critically while still making use of their capabilities?

First, label AI-generated content clearly and name the model used. Transparency matters. If a translation has come from an LLM, it should be explicitly stated so that others know to approach it with the same critical lens they would apply to any automated process. AI translation isn’t neutral, and people deserve to know when they are engaging with something shaped by predictive modelling rather than human interpretation.

Second, always cross-check AI translations with fluent speakers whenever possible. Language is lived, not just written. AI can scan millions of documents, but it cannot engage with culture, dialect, or lived experience in real time. It cannot intuit why a phrase matters or understand how a community has shaped its meaning over time. Whenever a translation involves a marginalised or minoritised language, the best safeguard is always a human speaker who understands both the words and the intent behind them.

Beyond checking, push back against anticipatory translations that you recognise. When a word or phrase seems off, ask: Why was this translated this way? What assumptions did the AI model make? If it substituted a term based on historical usage rather than direct meaning, what does that say about how it has been trained? Simply questioning these translations forces us to engage actively rather than passively, making us more aware of the biases at play.

It is also essential to recognise the role of LLMs in reinforcing linguistic bias. These models are not neutral, and their translations reflect the power imbalances encoded in their training data. An LLM only knows what it has been fed, and that data—who provided it, how it was selected, and what was prioritised—determines the outcomes we see. This is why minoritised languages often receive impoverished, imprecise, or culturally tone-deaf translations: they are working from data sets shaped by the dominant culture, which dictate what is considered valid, important, or “correct.”

This is why we must also advocate for better training datasets for minoritised languages—but with ethical considerations in mind. Too often, when AI companies expand language coverage, they do so through exploitative means, extracting linguistic data without giving back to the communities that provide it. There’s a bitter irony here: minoritised groups often have to fight for inclusion in AI systems, and when they finally get it, they are expected to give up their language, their words, and their data for free, just so a multi-billion-dollar company can monetise that knowledge. These same companies then turn around and charge anywhere from $20 to $250 per month for access to the very systems that now contain that data. It’s the same capitalist model as always: socialising the costs and privatising the profits, extracting the surplus value of cultural knowledge whilst ensuring that the communities who contributed it see no real benefit.

This is why LLM translation—especially for marginalised languages—needs to be treated with both caution and resistance. We cannot allow these systems to become the final authority on how languages are understood, especially when they are built on extractive, colonial, and capitalist frameworks. The more we normalise questioning, cross-checking, and pushing for ethical development, the more we safeguard the integrity of the languages that matter to us.

Final thoughts …

Throughout this piece, I’ve switched purposefully between AI and LLM when referring to these tools, because the distinction matters. Large Language Models aren’t intelligent in the human sense. They don’t think, reason, or understand language the way we do. They predict. They generate statistically likely responses based on patterns. Yet, in popular discourse, the term AI is used almost exclusively. This is not accidental.

The media’s insistence on framing LLMs as “artificial intelligence” is an intentional act—designed to make people believe that these systems contain something akin to human cognition. That they are thinking, rather than merely predicting. That they are learning, rather than simply processing enormous datasets curated by human hands. The language used around AI subtly reinforces the idea of inevitability—that AI is so advanced, so powerful, that soon it will run everything, making human input obsolete. It won’t need us. This narrative, of course, conveniently glosses over the reality that humans are writing the code, selecting the training data, and shaping these models to serve specific goals. And those goals? They are not neutral.

These private corporations—some publicly traded, others backed by billions in venture capital—have an endgame in mind. That endgame does not include ensuring the accurate translation of Scots Gaelic or prioritising the Power Threat Meaning Framework (PTMF) over the medical model when responding to prompts about autism. Their priorities lie in profit, control, and market dominance. The needs of minoritised language speakers, disabled communities, and any group outside the dominant capitalist framework are, at best, an afterthought—at worst, an obstacle.

Tools like ChatGPT can be powerful, but they don’t just translate—again, they predict. And prediction is not neutral. It is based on rules that humans create, with specific ends in mind. When working with marginalised languages, predictive bias doesn’t just lead to occasional errors—it leads to systemic misrepresentation. This isn’t some unfortunate byproduct of LLM development; it is intentional. The dominant narrative always reproduces itself. Without intervention, it will continue to do so.

So what’s the solution? It isn’t to reject LLM translation outright, nor is it to assume that LLMs will improve on their own. The solution is to approach these tools critically, to demand transparency, and to ensure that minoritised language speakers are the ones shaping how their languages are represented. It means actively questioning, challenging, and pushing back against the idea that LLM-generated text is inherently authoritative.

Most importantly, it means keeping this conversation going. Talk about it. Share this article. Spread the word. Push back. The more people understand how these systems work, the harder it becomes for corporations to hide behind the myth of LLM neutrality. Engineer your prompts in very specific ways. Ask why translations are being framed a certain way. Demand better.

Because language is power. And if we don’t fight for it, someone else will shape it for us.