Turns out, AI’s a bit of an arsehole

The hidden cost of AI translation

Translation used to mean a person, a skill, and time. Now, for a huge share of the world’s content, it means a prompt and a few seconds’ wait. 

Machine translation handles the overwhelming majority of translated content produced globally, and generative AI is accelerating that shift. Every product description, support article, and social caption can be “translated” instantly, for next to nothing.

Except it isn’t for nothing. It just moves the cost somewhere you can’t see it: onto quality, and onto the planet.

When speed costs you the message

Machine translation is fast because it isn’t really translating. It’s predicting the most statistically likely words in the target language, without understanding tone, intent, or the picture sitting next to the text. Most of the time that’s good enough to get the gist across. Occasionally it isn’t, and the gap between “good enough”, “actually right” and “culturally nuanced” is exactly where brands get burned.

In 2017, Facebook’s automatic translation tool turned an Arabic “good morning” into “attack them” in Hebrew. The post was a photo of a construction worker smiling next to a bulldozer, holding a coffee. There was no ambiguity to a human reader. A single mistranslated letter turned a greeting into a threat, and the man was arrested and held for hours before anyone caught the error.

That’s an extreme example but the underlying failure is ordinary: AI translation has no idea what it doesn’t know. It can’t picture the product, sense the sarcasm, or flag that a phrase lands differently with a Brazilian audience than a Portuguese one. It won’t warn you when a slogan reads as a threat, an insult, or a punchline. It will simply serve up its best statistical guess with total, unearned confidence. A human catches that gap before it becomes a headline. A machine, left unchecked, becomes the headline.

We hear examples of these failures from our clients a lot. A recent, more jovial one comes courtesy of a company we’re currently in talks with to replace a well-known AI-powered translation agency. Amongst other clangers, the agency delivered a line stating “[Company] will take you on safari” (worth noting the client in question is a digital employee experience platform). Sadly, no safaris included.

It gets better: the client was paying for that agency’s “human service” for the work in question. Another clanger.

The cost you don’t see

Every one of those “free,” instant AI translations draws real water and real electricity from the physical world.

This isn’t only a free-tool problem, either. Plenty of translation providers, like the one mentioned above, now market themselves as AI-powered or AI-first, running client work through the same large language models behind a paid, branded interface. The invoice looks more professional than a free consumer tool. The underlying computer, and the water and electricity it draws, is the same.

A single AI query is estimated to use anywhere from 0.3 mL to 10 mL of water, depending on whether you’re counting just on-site server cooling or the full picture, including the water used to generate the electricity behind it. Much of that water doesn’t return to the local supply it was drawn from, evaporation releases it into the atmosphere, where it comes down as rain somewhere else entirely, unlike household wastewater, which is typically treated and returned to the same system.

Scaled up, the numbers stop looking small. At a billion queries a day, even the conservative estimate (the tech giants will have you believe) adds up to roughly the daily water use of a thousand households. Data centres in Virginia alone used 1.85 billion gallons of water in 2023, up 65% in just four years, and U.S. data centre water use is projected to reach 74 billion gallons annually by 2028. In parts of the UK, that demand will inevitably compound an already-growing water shortfall.

None of this shows up in the two seconds it takes to get a translated sentence back. But it’s there every time, and it’s only growing as more everyday tasks, translation included, get handled by AI.

Why human still wins

A skilled translator, working from scratch, catches what a model can’t even see coming: the wrong gender assigned to a name, the wrong “you” in a formal email, the joke that doesn’t survive the border crossing. That’s not a matter of reviewing someone else’s guess and correcting it. It’s building the translation the right way the first time, with the context, intent, and audience in mind from the first word. They do it without asking a data centre in a water-stressed region to run a little hotter on your behalf.

At CultureSmith, our translators work from the source text, not from an AI’s first attempt. Translation and transcreation are human-first from the outset for that reason. Technology can help in the process, and often does, but the judgement, cultural nuance, and accountability your brand needs still comes from a person translating with your audience in mind, not a model guessing at it, and not a human relegated to cleaning up after one.

Who gets paid?

There’s another cost worth naming plainly: when translation defaults to a machine, a professional translator doesn’t get the work.

Translation and transcreation are skilled trades, built over years of bilingual fluency and an ear for how a market actually speaks. Every project routed to an AI model by default is income that doesn’t reach the translator, the reviewer, the in-market specialist who built that expertise in the first place.

It’s a shrinking pipeline. If fewer paid projects go to human translators, fewer people can afford to build the specialism at all, and the industry loses the expertise it needs to catch what AI gets wrong. Quality and cultural judgement don’t appear from nowhere. They’re built on paid work and years in the trade. Choosing a human translator isn’t just about a better outcome on this project. It’s an investment in having skilled people around for the next one.

Faster isn’t the same as better. Cheaper isn’t always cheap. Someone, somewhere, is always paying for it: a water-stressed grid, or a translator who didn’t get the brief.

Want translation that gets more than the gist across, without the hidden costs? Let’s talk