Previously, Gizmodo en Español had a small but dedicated team who wrote original content tailored specifically for Spanish-speaking readers, as well as producing translations of Gizmodo’s English articles. The site represented Gizmodo’s first foray into international markets when it launched in 2012 after being acquired from Guanabee.
Newly published articles on the site now contain a link to the English version of the article and a disclaimer stating (via our translation from Google Translate), “This content has been automatically translated from the source material. Due to the nuances of machine translation, there may be slight differences. For the original version, click here.”
As both a translator and a tech writer, this article touches upon a lot of aspects of my professional life. As a translator with a master’s degree in translation and over 13 years of experience, I can confidently say these AI-translated articles won’t be anywhere near the quality of a professional translation, let alone that of original content written in Spanish. Computers are actually not that great at language, and every time I play around with machine translation tools – they tend to be integrated into the various translation software suites I use – it’s barely passable as coherent text.
There are things you can do to increase the success rate of machine translation. It’s crucial to write the source text in a very formulaic manner, using short sentences with basic sentence structure any primary schooler can easily follow. Avoid complicated clauses, literary devices, sayings and wordplay, and words that can carry multiple meanings. To further increase the success rate, make sure your writers reuse the same formulaic sentences in different articles, so the machine translation software can learn from earlier corrections.
By the time you instilled all this and more into your writing staff, not only will they quit because writing in such a way is not engaging at all, it will also tank your SEO – something the kind of people who would fire translators to rely exclusively on machine translation would care about – into the ground. It wouldn’t feel natural, and nobody will enjoy reading it but computers.
…it’s going to end up as AIs writing for other AIs.
Well, this seems short sighted. Certainly it is not going to be as successful as true Spanish language content.
Then again, it may actually be a choice between completely shutting down Spanish language publication and relying on translation. From that angle, transition would be the better, less drastic choice. I guess only they know how much value this content adds relative to resources expended.
A genuine question for translators, how have LLMs like GPT impacted this analysis. The author says that computers are not good at language. Certainly that has been my experience with translation. However, while AI like ChatGPT can spout nonsense, the one thing I have been impressed with is how “correct” the English that GPT generates generally is. It can be verbose sure but it is otherwise high quality. While are all getting better at spotting it, it was generally good enough to fool native speakers at first.
If we allow an LLM not just to “translate” sentences completely word for word but rather “rewrite” the content to some extent, how does that impact this “computers are bad at language” equation. My own sense is that they are not really that bad at it anymore. If you give an LLM a few sentences in English and tell it to summarize, you get back a few different sentences that are more or less equivalent in terms of meaning ( depending how careful you are with your prompts ). I would think that moving between languages would not impair this all that greatly. Does it though?
Any translators familiar with how LLMs have changed the game? How do they do these days?
My experience has been that LLMs probably won’t help enough, because A.I. translation still often struggles to make sense of things unless there’s a human operator fluent in the language on the receiving end who can identify and massage the translation past those sticking points.
(While I haven’t published it yet, I’ve written a draft of a guide to how, as an English-speaking enthusiast with no money to pay professional translators, you can get decent results out of OCRing and Google Translate-ing of Japanese fan-works from Pixiv.)
LLMs may be very effective plausible gibberish generators, but “garbage in, garbage out” still applies.
Context is very hard. Even human translators have issues with it. I have KDE in my local language, and right clicking on the Trash brings a menu with «Empty» translated as a noun, instead of a verb.
Parodper,
I agree, sometimes the context is just not there. I’ve seen how software translations are done in practice: send out spreadsheets with english columns and having translators fill in other columns. There are no screen shots or other context clues to help the translator and generally the service completes their task of translating text without an opportunity to ask questions. For companies that don’t have translators on their software teams who are also knowledgeable about local context, they just assume the translations are correct.
Thom Holwerda,
I’m skeptical of the “won’t be” claim. If you say they’re not there today, then I believe you completely, you are an authority on the subject after all. I’ve seen silly translations too. Reminds me of this…
https://www.osnews.com/story/136228/windows-11s-latest-endearing-mess-rigorously-and-wrongly-enforces-britishisms/
But AI is going to keep improving and I believe the view that AI will not approach or exceed human abilities is naive. I predict the day is coming when AI will regularly beat professionals at their own professions at least in the digital domain where physical presence isn’t required. IMHO it’s less of an ‘if’ and more of a ‘when’.
Alfman,
With the greatest of respect to your viewpoint (I have read a lot of your contributions on this site and they are well thought out), I would have to at least differ in opinion on your last paragraph, whilst not necessarily entirely disagreeing with it.
Generally, technology improves for sure. That said, based purely on scientific principles, there isn’t a lot of evidence to say the current crop of LLM-type solutions will actually continue to improve – indeed, there is just as much speculation that they may plateau at some point. Will something else replace them? Maybe, even probably, but without knowing what it is, I wouldn’t necessarily think that AI as we currently imagine it will keep improving in the short to medium term, given that we don’t know how to build what comes next.
Let me ask you, are you thinking that only the “current crop of LLM-type solutions” will plateau or AI in general will plateau? I believe there’s still plenty of room for new AI innovations to keep the field moving.
I still think there are plenty of ideas to try. One of the more radical/frankenstein-esque approaches that I saw in a documentary was training AI out of actual “living” neuron cells. (Are biological cells “living” when grown in artificial environment? Whatever! The show must go on 🙂 ) The density, efficiency, and cost advantages of this technology over power hungry compute servers would be enormous. It’s still early days for this research, but it’s a proven technology “in the field” so if we master it in the labs, the potential for artificial applications is huge. Anyway, let me get off the crazy sci fi rails and pivot back to our discussion, haha.
Still, even if we dismiss AI’s potential to become more powerful, there’s another aspect we should be considering. We are being critical of how general purpose AI handles specialized jobs. Just because AI is better than an average person doesn’t make it better than an expert, which is a fair point. However I don’t believe it is this general purpose AI that’s going to replace most of the jobs. It’s going to be specialized AI trained on private data and records that ChatGPT and the like never had access to via public sources. Once businesses (and governments) start training AI using their internal customer & case records, etc. I expect it will create lots of redundancy even for those with expert level domain knowledge. Rank and file employees who are processing those cases may find that AI can do their jobs faster, with more consistency, and at scales that humans cannot come close to matching. AI, while not cheap to develop, could save millions annually if it can replace dozens/hundreds/thousands of case workers. Even allowing for some humans to remain for exceptional cases, I still find it plausible that AI could take over the bulk of the work.
I was referring to large language models in particular yes. Plenty of room for further innovation elsewhere in the field for sure, totally agree, but it certainly isn’t moving at the break neck speed we are often lead to believe.
The neuron cell stuff is definitely crazy interesting, but the ethical concerns are potentially there too. When do a collection of biological neurons become alive/able to suffer, as opposed to a bundle of cellular machinery – no one knows right, that’s the scary part.
In regard to training specialised LLMs (that is what we are effectively talking about here in the short to medium term), I don’t fully agree with your conclusions – on the one hand, they will open up some interesting potential in terms of untapped business value. On the other however, their reasoning and logic skills will continue to remain incredibly limited in my opinion, and this will limit the jobs they can replace (at least in some areas anyway). This is not a solved problem (in terms of the technology), nor is there any indication it will be solved soon.
The assumption that larger model sizes would help with reasoning skills was (as far as I can tell from available evidence) based on not much more than assumptions, and certainly hasn’t worked out. As for smaller, more targeted models – these I find interesting due to the democratising effect they could have on the technology. That said, the ‘last 10%’ issue is likely to be with us for a while yet anyway.
I don’t necessarily think this stuff is going away, but I think people will be surprised by just how difficult it ultimately ends up being for that last 10% (or 1%, whatever we want to label that last mile) of the human element to be replaced. For sure the tech bros would replace all of us overnight if they could, but that doesn’t mean it will pan out that way. Is that naive? Maybe, I guess we’ll see 🙂
PhilPotter,
It’s an interesting debate. To me logic is logic, regardless of how the machine gets implemented. Maybe it’s a case of too close to our own biological processes for comfort?
How so? IMHO AI capabilities have grown exponentially and I see no signs of it stopping. Granted I have to concede your point that all my speculation about the future of AI is built on assumptions, like costs coming down, more sophisticated models, training, and more specialization. However isn’t your opinion that AI is plateauing also built on assumptions that these won’t continue to improve?
Even accounting for exceptional cases, wouldn’t you agree that most office work could be automated already? I think many jobs are already at high risk, they just don’t realize it yet because nobody came up to their employers offering AI wrapped up in a nice bow. AI still requires a very large upfront technology and training costs. But we’re still in the early adapter phase, as the technology improves and becomes easier and more accessible, I expect it could snowball into hugely disruptive technology.
Of course we can’t prove the future and it’s all just speculation. But I do feel like people are too quick to assume their jobs cannot be automated. I think a lot of the workforce today are going to be affected before they’re able to retire. Employees keep getting more expensive meanwhile AI is projected to get less expensive. Employers will be under immense financial pressure to give AI a go at least for parts of their business. And once the foot is in the door, AI will likely take on more and more work that previously required employees.
Yes, I know that I could be completely wrong about all of this, but I really don’t think it’s safe to assume it won’t happen.
And so it begins, like site owners care about quality when $ are on the line. It’s not and never was about removing all the human translators, it’s about having 1 do the job of 10!
I suspect the same is going to happen with much of the mundain coading as well, and AI will bash out the bulk and some human will tweak and tune the edges.
Some say we’re at the end of the era of the written word anyway. AI cannibalizing other AIs cannibalizing human writers might be both a sign and an accelerator of this.
dsmogor,
Very interesting take. So the AI isn’t getting it wrong after all, it’s just indicative of what to expect of our languages in the future 🙂
On a serious note though, the data we’re using to train AI, especially from the internet, may not be our most high caliber work. A great deal of human writing doesn’t get proof read and is riddled with errors. So if this is the quality of data we’re using to train AI, then it would be unsurprising for errors and poor quality to be reflected in the results.
As a computer science problem, it’s not enough to feed the AI all the papers in the class so to speak. The AI also needs to know the grade and feedback those papers got as well. This way the AI can gain a better understanding of what quality is and knows what not to do. I think current models and training data are missing this type of feedback even though it’s important for AI to be able to discern quality of human work. Otherwise it gets trained on a mediocre dump of everything. Once AI can effectively rate the quality of input, it will be in a stronger position to create quality output too by weeding out/auto correcting it’s own bad work.
All the buzz coming from higher ups at various companies I interface with is about how they won’t have to pay as many of those dirty workers any more. The opportunity, if anyone is paying attention, is for other people powered companies to come take their lunch. These clowns are so delusional.
CaptainN-,
I’m not so sure. Many jobs are menial and just require a little help from AI to displace employees. Like when we call support staff who are following a script or just provide a human interface for a computer program anyway. I’m of the belief that any jobs that can be automated will be automated whether we like it or not (and no I don’t always like it). Automation creeps in with the low hanging fruit first, but it won’t stop there. Over time it will become more proficient and will be able to displace more human workers. I don’t think employers necessarily have it out to replace employees, however since this is capitalism it is a matter of costs that will ultimately decide who wins. Even a company that prefers humans may find itself pressured to use AI to stay competitive or else be leapfrogged by more profitable companies.
You’re seeing the early stages of it. It would be silly to think they won’t continue or be successful eventually. I have relied on auto translate more than a few times to understand foreign news, parse complex documents , etc. The struggle for years was between labor and capitol with peace only coming when capitol was forced to realize they needed labor. But now they don’t for many things. I don’t think they’ll keep people around from the kindness of their hearts. I don’t know what’s in store now. I think Capitol has been too successful for too long, and has such a huge lead on labor organizers. A significant number of laborers are too busy trying to figure out what laws to pass to punish those with different values than them to do anything to counter the situation.