1. AI output needs to be reviewed since AI systems are trained with imperfect human translations, AI output is also imperfect.

2. AI systems degrade over time: The training data will eventually become outdated. Another cause of degradation is recursive training. AI-generated content has been proliferating on the internet without being marked as AI-generated. Much of this content will flow back into the training data of foundation models.

3. AI models do not want to produce good quality — actually, they don't have any needs at all.

4. Language models have no connection to reality. More precisely, the only connection to reality is the training data.

5. Requirements for a certain type of translation are different, so the same AI model or even the same automation approach is not suitable for every situation. Finding the right model, the right prompt, the right automation process is a task for a human engineer — at least for now.

6. Language combinations are not equal: there is a lot of materials in one language (like in English), and not much in another. Because of this, the quality of the AI system's work depends heavily on the language it has to work with.

According to a report by UserPilot, around 52% of users have uninstalled an app due to poor localization. But that‘s not a translation problem. That‘s a product problem.

When most people outside our industry think about translation, what comes to mind first? Google Translate. Then ChatGPT. The sophisticated ecosystem of translators, interpreters, localization engineers, and cultural consultants remains largely invisible to the very people who need these services most. <...>

While they struggle with user conversion loss, the very people building these apps often lack awareness of professional localization services. They‘re solving a product problem with the wrong tools because they don‘t know better tools exist.

This disconnect has created interesting market dynamics. We‘re seeing wellfunded, Y Combinator-backed startups
emerge <...> without any team members who have localization industry experience. This isn‘t necessarily their fault, but rather it’s a symptom of the broader disconnect between our industry and the product teams who desperately need our expertise.

We‘ve remained behind the scenes while the problem we solve has become central to product success.

The language industry follows the same pattern every time.

Translation memories triggered anxiety. So did statistical machine translation, then neural MT, and now large language models. None of these replaced translators, but each one changed how translators work and expanded the scope of what could be translated.

After a year of watching generative AI evolve, it became clear to me that this is not a disruptive innovation for translation. It's an incremental one.

Translation accounts for less than 0.001 percent of the world's content creation. What AI does is make it possible to translate more of that content. Words that would never have crossed a language barrier before are now within reach. Google Translate handles one trillion words a month. That's more than all human translators translate in a year.

Mistakes I Cannot Sea

I ran this poem thru it
I’m sure your pleased to no
Its letter perfect in it’s weigh
My checker tolled me sew.

Candidate for a Pullet Surprise

I have a spelling checker,
It came with my PC.
It plane lee marks four my revue
Miss steaks aye can knot sea.

Eye ran this poem threw it,
Your sure reel glad two no.
Its vary polished in it’s weigh.
My checker tolled me sew.

A checker is a bless sing,
It freeze yew lodes of thyme.
It helps me right awl stiles two reed,
And aides me when eye rime.

Each frays come posed up on my screen
Eye trussed too bee a joule.
The checker pours o’er every word
To cheque sum spelling rule.

Bee fore a veiling checker’s
Hour spelling mite decline,
And if we’re lacks oar have a laps,
We wood bee maid too wine.

Butt now bee cause my spelling
Is checked with such grate flare,
Their are know fault’s with in my cite,
Of nun eye am a wear.

Now spelling does knot phase me,
It does knot bring a tier.
My pay purrs awl due glad den
With wrapped word’s fare as hear.

To rite with care is quite a feet
Of witch won should bee proud,
And wee mussed dew the best wee can,
Sew flaw’s are knot aloud.

Sow ewe can sea why aye dew prays
Such soft wear four pea seas,
And why eye brake in two averse
Buy righting want too pleas.