Transformative Translation
The AI ability not being talked about
In the last year or two, AI has done things that were supposed to take decades. Most of the awe is pointed in the wrong direction.
The thing everyone's overlooking
Large language models trained on multilingual content can take input in any language and produce output in any other — perfectly, with full fidelity of tone and intent. Not word-for-word substitution. Actual transformation. The input is processed across a shared internal representation and reconstituted in the target language as though it originated there.
This is qualitatively different from what Google Translate and DeepL were doing before. Those tools improved dramatically through deep learning, and in their best moments could match register and intent, not just surface meaning. But LLMs blew them out of the water. Ask ChatGPT or Gemini something in any language, request output in another, and you get back something that doesn't read like a translation. It reads like original writing.
That distinction matters. But we're mostly not talking about it.
Where I'm coming from
I grew up the child of two people from two different countries, raised in a third. Three native languages. This gives you a particular angle on the language barrier that's hard to acquire any other way — you exist inside multiple linguistic worlds simultaneously, and you can feel the walls between them.
When I was young and the European Union was taking shape, I was enthusiastic. But talking to people across those countries over the years, I kept noticing the same thing: the project didn't land for them the way it landed for me. Not politically, not emotionally. They didn't feel European the way I did. And the more I thought about it, the more I came to understand why. Empathizing with another country's concerns — genuinely internalizing their perspective — is hard enough between people who share a language. Across a language barrier, it's close to impossible. The EU was asking citizens of different nations to care about each other's problems. But it couldn't help them speak each other's languages.
The language barrier runs from individual conversations all the way up to geopolitical relations. At every level, it stifles communication, collaboration, empathy.
Ambassadors as protocol bridges
We've long had a word for the people who try to bridge this gap at the highest level: ambassadors. But translation undersells what they actually do. A skilled ambassador isn't converting words from one language into another. They're functioning as a transformative conduit between two systems running on fundamentally different protocols — different assumptions, different emotional registers, different cultural contexts.
Am I saying it's time to send Gemini to negotiate the next peace treaty? Obviously not.
But there's a genuine and underestimated transformative power in what LLMs can already do. I've seen it directly. I had an AI agent take old projects from my early programming days — written in languages nobody uses anymore — and transform them into modern equivalents in different languages. Flawlessly, with some minor intervention on my part here and there. Of course, you might say a machine has an inherent edge when it comes to translating machine languages — fair enough.
But the same agent was also able to take a poetic lyric I'd written in Spanish and produce an English version that respected the structure, found new rhymes, and kept the meaning almost verbatim — not just translated, but culturally adapted to some degree. That ability, to transform content from one language to another while preserving every nuance of intent, is in some ways more remarkable than the ability to generate content from scratch. Generation is impressive. Transformation that holds the voice intact is something else.
It's not just about countries
The language barrier isn't only about French versus German versus Japanese. Within any single life, there are countless moments when a person can't express what they actually mean — wrong vocabulary, wrong emotional state, wrong moment. I'd bet there is no one who hasn't said something in exactly the wrong way and only recognized it afterward.
I certainly have.
For years, I've preferred communicating through chat rather than speech. Not out of antisocial tendency — I eventually figured out the actual reason, and I'm certainly not alone in it: chat gives you a pause. Before you hit send, you have a moment to read back what you've written and consider its effect. Speech doesn't. Speech prompts you to respond immediately, which means you're constantly operating one step behind your better judgment.
The cooler-head problem, solved differently
Text communication gave me something I valued: the ability to wait. If I received a message that landed badly — something that provoked an emotional response — I could sit with it. Come back to it later. Write the reply I'd want to have written, rather than the one I felt like writing in the moment.
LLMs change this calculus entirely. Now I can write the reactive version — unfiltered, emotional, structurally a mess — and hand it off to an appropriately instructed model to transform it into the measured response I'd have written with a cooler head a few hours later, or the next day. The waiting was never about waiting for its own sake. It was about the quality of the output. That quality is now available immediately.
And it goes beyond emotional state. If I'm responding within my domain of expertise and get a detail wrong, my AI assistant can catch the error and correct it before the message goes out. Fact-check me in real time, expand on what I said, sharpen the accuracy. This raises obvious questions about dependency and authorship and agency. But the potential to improve the quality of communication — at every level, across every kind of barrier — is undeniable.
What it means in practice
I've been publishing my articles in multiple languages for some time now. Before LLMs, that meant a translation process that, even with the best available tools — and DeepL, a product out of Cologne, was genuinely good — required extensive manual editing if I wanted my voice to survive the crossing. Now, when I proofread an AI translation of my English writing into German — my mother tongue — it reads as though I had written it myself. Not "this is a passable translation." More like: this is an adapted version that fully contains everything the original had. The voice made the crossing intact.
Babel, inverted
There's an old story about humans once speaking a single language. That unity let them collaborate on something ambitious enough that it required divine intervention to stop them. Then the languages fragmented, and the collaboration ended.
That's likely not history. But it's a useful frame. The language barrier has consistently constrained what humans can build together — not just between nations, but between individuals, within relationships, inside a single person's ability to say what they mean.
AI is generating a lot of justified awe right now. Most of the coverage focuses on what it can do directly: research outputs, tasks previously impossible for machines, capabilities arriving faster than anyone expected. That's all real. But running underneath it, maybe less visible, is something potentially more consequential: AI as a bridge for human-to-human communication. Not AI doing things for us, but AI enabling us to do things with each other that the language barrier has always prevented.
That's where the greatest potential might actually lie.