You've moved to different country with a language that is vastly different from yours. Let's say you're an American moving into the Czech Republic. You need to sign an important document that has legal and business ramifications. Would you trust an AI translation on the document or ask for a professional to be in the loop?
Commercial translator services lately are the worst they have ever been. You cant validate that they aren't directly sending your excel with the translation lines into a LLM with no tweaking/checking.
For a indie videogame i work on, we tried a couple translation agencies, and they gave terrible output. At the end, we built our own LLM based agentic translation, with lots of customization for our specific project like building a prompt based on where the menu/string is at, shared glossary, and other features. Testing this against the agencies, it was better because we could customize it for the needs of our specific game.
Even then, at the end of the day, we went with freelancers for some of the languages as we couldn't really validate the AI output on those languages.The freelancers took a month to do the translation vs the 2-3 days we ourselves took for the languages we knew and we could monitor the AI output. But they did a nice job, much better than the agencies.
I feel that what AI really completely kills is those translation services. Its not hard at all to build or customize your own AI system, so if the agency is going to charge you considerable money for AI output, just do it yourself and get a better result. Meanwhile those freelancers are still in demand as they can actually check the project and understand it for a nice translation, unlike the mechanical agencies where you send them the excel and they send it to who knows what or an AI without you being able to check.
I will likely be opensourcing this customizable AI translation system for my project soon.
> While using AI tools for everyday tasks like finding directions is “low-risk,” human translators will likely need to be involved for the foreseeable future in diplomatic, legal, financial and medical contexts where the risks are “humungous,” according to Benzo
Now it's a classic, you need an expert in order to check the work of the machine, because the "customer" is by definition not able to do it.
Aside from highly technical domain, in purely literary works, I think that the translator is a co-author - maybe IP laws acknowledges that already? I remember the translation of E.A. Poe by C. Baudelaire for instance; I think you could feel Baudelaire's style because it is a lot "warmer" than Poe's. I've also read a translation of a Japanese novel and I was quite disappointed with it. I don't know Japanese but I have read/watched quite a few mangas/animes, so I could sense the speech patterns behind the translations and sometimes thought they could have made better choices.
In any case, one will still need a translator who is good at "prompt engineering" to get a quality translation. I don't know. Maybe translators can add this skill to their CV, so they can propose quick-and-dirty/cheap translations, or no-AI high quality translations.
Some suggest "no-AI" labels on cultural products already - I think if it becomes a reality it will probably act as "quality signaling", because it is becoming more difficult every year to tell the difference between AI and human productions. It won't matter if what you read was written by an AI or a human (if it quacks and looks like a duck...), but what the customer will probably want is to avoid poorly-prompted machine translation.
The same is happening to a lot of website owners. You lose half or more of your traffic to AI summaries trained on your own content. The cost of producing original content is the same as before.
It makes research harder too, since more and more public information is infected by AI content. Both published posts and internet discussions are tainted.
And then the AI companies threaten to crash the whole economy if we don't let them do it.
Would like to point out that professional translation has been under pressure for much longer than AI.
I have friends that made a descent buck 20-30 years ago translating technical documents like car manuals. Over the years, prices fell from quarters per words to fractions of a cent.
And even though machine translation was barely existent, tools were used to argue higher productivity and therefore lower prices.
AI translations are getting good. I've been working on our company website in the last month. It's a static website. So, I use Agentic coding tools to implement both content and technical changes. A simple prompt "align the translations with the English version for all pages" and some guard rails in the form of skills and AGENTS.md instructions are doing most of the heavy lifting. To be clear.
These translations are not perfect, yet. But good enough for my needs. Any professional translator services would in any case be beyond our budget. The advantage of using agentic coding tools here (enabled by using a site generator rather than a CMS) is that I can get systematic about dealing with jargon, SEO, and frequently used phrases. I simply document all that and instruct the tool to refer to that. The funny thing is that most of the models are pretty good at fixing their own mistakes if you just ask them too. I asked it to look for examples of "denglish" (German English) in its own German translations and then to fix it. It found a few examples and the suggested fixes were fine.
A lot of people are focusing on the negative here. I like to look at the positive. We're approaching the moment where any person on this planet will be able to communicate directly with any other person on this planet without the need for translators. The tools already exist for this. But they need a lot of work on quality.
A second point here is that the role of English as the most popular intermediary language is disappearing as well. I'm not a native speaker. When I talk to foreigners from wherever, it's mostly in (bad) English. By definition that limits me to talking to people that have had enough education and exposure to English. This is very limiting. A lot of the people we need to talk to here in Germany aren't all that comfortable speaking English.
Recently I've noticed that the subtitles on TV shows have become garbage. Often the translations makes no sense in the context of what is happening, but are literal translations of what is being said on screen. I assume it's some kind of AI translation, but who knows. Maybe the translator is just stupid.
To be clear, we had bad subtitles before as well, but that was due to the translators lack of cultural understanding. Most of it made sense at least. Now it's just straight up bad and often makes no sense. Sad!
I will take a more positive view, all the translators being put out of buisness will take up literary translation and we will enter a golden age when it comes to translated literature. AI still has a very long ways to go here and there is a great deal of untranslated literature out there and literature which only has a terrible translation. Probably won't happen but maybe the struggling publishing industry will see the opportunity provided by a glut of out of work translators and jump on it.
I don't doubt translators face massive job losses. I hate this "personal interest" style anecdotal news stories though, because it gives me no actual sense/data on how fast/slow this transition is, how widespread etc. It's basically junk food.
AI translation is so much usable these days I think.
I had a report/article (using Claude) on methods to fall asleep. First method worked amazingly and I am doing that every night. I translated it to Urdu using Gemini (because they are doing translations for a long time, they must be better). First translation used very stiff/dry language, something you might read in a research paper etc. I asked it to simplify it for my mom the language and to my astonishment, it did that very very well.
Urdu surely must not have as much data as other languages. I could not find a single mistake. Yeah there were some weird choices of words in original version, but the simpler was so good it can be published like an article.
My mom retired as an independent translator 5 years ago.
She worked freelance 40 years from age 25 to age 65.
In the 15 years that preceded her retirement, she would get less and less work.
Partly she's an introvert who relies on her network to provide work, and her network gradually retired.
But machine translation was the big killer.
Before LLMs, the early versions of Google Translate killed paid translation.
As the market adapted to machine translation, and as the internet became a globalised platform for knowledge work, it also opened up to lower grade translation, and there were suddenly many more translators willing to work at a lower wage.
Prior to Google Translate there were semi-automated systems that would fuzzy match from large databases.
But it'd still rely on a human-in-the-loop for adapting the sentences.
With Google Translate you'd get a super sketchy translation out, very crude and not at all correct or idiomatic in the target language. Any distance between the source and target language (e.g. English -> Chinese) and it'd be one big joke. With plain Google Translate it still is. But the market spoke: Probably you don't need a very good translation most of the time. Especially not if the shitty one is free.
In her later years she moved to transcription of board meetings. She'd type up everything that was said.
I work for a company now that automates transcription via the whisper model and generates summaries that can be adapted by the customer. You pay per minute of transcription, and you can regenerate summaries as much as you want after that until your prompts give the right results.
All of this manual labor that provided for my childhood is gone now.
I couldn't imagine being a professional translator today and not use AI extensively.
But unless I have a legal reason to consult with a professional translator, I probably don't even need one, since LLM-based translation is as good as it gets with just plain LLM usage, and near perfect with automated translation tools that will help you pick both the mood, formality and alternative formulations for your translation.
High-grade translation is massively parallelisable, and a human-in-the-loop is entirely for final proof-reading.
The premise is absurd. Imagine if coders decided to stop working "on principle" because AI agents "learn from your work"
People would rightly call it out. Somehow gets a pass because its about translation instead of coding?
14 comments
[ 0.23 ms ] story [ 29.9 ms ] threadFor a indie videogame i work on, we tried a couple translation agencies, and they gave terrible output. At the end, we built our own LLM based agentic translation, with lots of customization for our specific project like building a prompt based on where the menu/string is at, shared glossary, and other features. Testing this against the agencies, it was better because we could customize it for the needs of our specific game.
Even then, at the end of the day, we went with freelancers for some of the languages as we couldn't really validate the AI output on those languages.The freelancers took a month to do the translation vs the 2-3 days we ourselves took for the languages we knew and we could monitor the AI output. But they did a nice job, much better than the agencies.
I feel that what AI really completely kills is those translation services. Its not hard at all to build or customize your own AI system, so if the agency is going to charge you considerable money for AI output, just do it yourself and get a better result. Meanwhile those freelancers are still in demand as they can actually check the project and understand it for a nice translation, unlike the mechanical agencies where you send them the excel and they send it to who knows what or an AI without you being able to check.
I will likely be opensourcing this customizable AI translation system for my project soon.
Now it's a classic, you need an expert in order to check the work of the machine, because the "customer" is by definition not able to do it.
Aside from highly technical domain, in purely literary works, I think that the translator is a co-author - maybe IP laws acknowledges that already? I remember the translation of E.A. Poe by C. Baudelaire for instance; I think you could feel Baudelaire's style because it is a lot "warmer" than Poe's. I've also read a translation of a Japanese novel and I was quite disappointed with it. I don't know Japanese but I have read/watched quite a few mangas/animes, so I could sense the speech patterns behind the translations and sometimes thought they could have made better choices.
In any case, one will still need a translator who is good at "prompt engineering" to get a quality translation. I don't know. Maybe translators can add this skill to their CV, so they can propose quick-and-dirty/cheap translations, or no-AI high quality translations.
Some suggest "no-AI" labels on cultural products already - I think if it becomes a reality it will probably act as "quality signaling", because it is becoming more difficult every year to tell the difference between AI and human productions. It won't matter if what you read was written by an AI or a human (if it quacks and looks like a duck...), but what the customer will probably want is to avoid poorly-prompted machine translation.
It makes research harder too, since more and more public information is infected by AI content. Both published posts and internet discussions are tainted.
And then the AI companies threaten to crash the whole economy if we don't let them do it.
I have friends that made a descent buck 20-30 years ago translating technical documents like car manuals. Over the years, prices fell from quarters per words to fractions of a cent.
And even though machine translation was barely existent, tools were used to argue higher productivity and therefore lower prices.
These translations are not perfect, yet. But good enough for my needs. Any professional translator services would in any case be beyond our budget. The advantage of using agentic coding tools here (enabled by using a site generator rather than a CMS) is that I can get systematic about dealing with jargon, SEO, and frequently used phrases. I simply document all that and instruct the tool to refer to that. The funny thing is that most of the models are pretty good at fixing their own mistakes if you just ask them too. I asked it to look for examples of "denglish" (German English) in its own German translations and then to fix it. It found a few examples and the suggested fixes were fine.
A lot of people are focusing on the negative here. I like to look at the positive. We're approaching the moment where any person on this planet will be able to communicate directly with any other person on this planet without the need for translators. The tools already exist for this. But they need a lot of work on quality.
A second point here is that the role of English as the most popular intermediary language is disappearing as well. I'm not a native speaker. When I talk to foreigners from wherever, it's mostly in (bad) English. By definition that limits me to talking to people that have had enough education and exposure to English. This is very limiting. A lot of the people we need to talk to here in Germany aren't all that comfortable speaking English.
To be clear, we had bad subtitles before as well, but that was due to the translators lack of cultural understanding. Most of it made sense at least. Now it's just straight up bad and often makes no sense. Sad!
One can dream.
I had a report/article (using Claude) on methods to fall asleep. First method worked amazingly and I am doing that every night. I translated it to Urdu using Gemini (because they are doing translations for a long time, they must be better). First translation used very stiff/dry language, something you might read in a research paper etc. I asked it to simplify it for my mom the language and to my astonishment, it did that very very well.
Urdu surely must not have as much data as other languages. I could not find a single mistake. Yeah there were some weird choices of words in original version, but the simpler was so good it can be published like an article.
She worked freelance 40 years from age 25 to age 65.
In the 15 years that preceded her retirement, she would get less and less work.
Partly she's an introvert who relies on her network to provide work, and her network gradually retired.
But machine translation was the big killer.
Before LLMs, the early versions of Google Translate killed paid translation.
As the market adapted to machine translation, and as the internet became a globalised platform for knowledge work, it also opened up to lower grade translation, and there were suddenly many more translators willing to work at a lower wage.
Prior to Google Translate there were semi-automated systems that would fuzzy match from large databases.
But it'd still rely on a human-in-the-loop for adapting the sentences.
With Google Translate you'd get a super sketchy translation out, very crude and not at all correct or idiomatic in the target language. Any distance between the source and target language (e.g. English -> Chinese) and it'd be one big joke. With plain Google Translate it still is. But the market spoke: Probably you don't need a very good translation most of the time. Especially not if the shitty one is free.
In her later years she moved to transcription of board meetings. She'd type up everything that was said.
I work for a company now that automates transcription via the whisper model and generates summaries that can be adapted by the customer. You pay per minute of transcription, and you can regenerate summaries as much as you want after that until your prompts give the right results.
All of this manual labor that provided for my childhood is gone now.
I couldn't imagine being a professional translator today and not use AI extensively.
But unless I have a legal reason to consult with a professional translator, I probably don't even need one, since LLM-based translation is as good as it gets with just plain LLM usage, and near perfect with automated translation tools that will help you pick both the mood, formality and alternative formulations for your translation.
High-grade translation is massively parallelisable, and a human-in-the-loop is entirely for final proof-reading.