AI translation: an honest take from people who do this for a living.

Every six months, someone declares translation is dead. Every six months, our work continues. Both observations are true at the same time, and the gap between them is where the actual story lives.

## What AI translation does well now

Let's be honest. Modern neural machine translation is good. Not "impressive for a computer" good — actually good. For internal docs, technical content, draft translations of common language pairs, the output is often usable with light editing. Pretending otherwise insults our clients' intelligence.

Common language pairs (EN, ES, FR, DE, etc.) at fluent quality

Repetitive content where consistency matters

First-pass drafts that humans then revise

Real-time gist translation for low-stakes communication

## What AI translation still doesn't do

And yet. Eight years into the neural MT revolution, the work that matters most still requires human judgment. Not because machines can't generate fluent text — they can — but because fluent text isn't the same thing as right text.

Voice preservation — keeping an author's distinct style across languages

Cultural mediation — knowing what to localize, what to footnote, what to leave alone

Sensitive content — where mistranslation has real human consequences

Low-resource languages — where training data is thin and bias is high

Inclusive language — where defaults baked into training data are exactly what you don't want

The new question: It's no longer "can the machine do this?" It's "is this the kind of work where the machine being mostly right is good enough?"

## How we actually use AI

We use it. A lot. As a productivity tool for our linguists, not a replacement for them. CAT tools have included MT integration for years. The good translators have always been faster than the slow translators — AI just changes the productivity ceiling, not the floor of skill required.

What we don't do: send sensitive client content to public AI tools, deliver raw MT output as finished work, pretend our human work doesn't increasingly involve AI assistance, or undercut translator pay because "the AI does most of it now."

## Where this goes

Translation as a profession is changing. Not dying — changing. The volume of low-stakes content that gets machine-translated will keep growing. The work that remains for humans will be more concentrated in high-stakes, high-craft, high-judgment territory.

If you're a client wondering when to trust the machine: ask whether the cost of getting it slightly wrong is acceptable. For internal Slack messages, sure. For your annual report on gender-based violence in conflict zones — please, no. We'd rather you went to a different agency than skip the human work for that one.