Which translation tools are LLM free? Will they remain LLM free?
Looking at the submission rules for Clarkesworld Magazine, I found the following:
Statement on the Use of “AI” writing tools such as ChatGPT
We will not consider any submissions translated, written, developed, or assisted by these tools. Attempting to submit these works may result in being banned from submitting works in the future.
EDIT: I assume that Clarkesworld means a popular, non-technical understanding of AI meaning post-chatGPT LLMs specifically and not a broader definition of AI that is more academic or pertinent the computer science field.
I imagine that other magazines and website have similar rules. As someone who does not write directly in English, that is concerning. I have never translated without assistance in my life. In the past I used both Google Translate and Google Translator Toolkit (which no longer exist).
Of course, no machine translation is perfect, that was only a first pass that I would change, adapt and fix extensively and intensely. In the past I have used the built-in translation feature from Google Docs. However, now that Gemini is integrated in Google Docs, I suspected that it uses AI instead for translation. So I asked Gemini, and it said that it does. I am not sure if Gemini is correct, but, if it doesn't use AI now it probably will in the future.
That poses a problem for me, since, in the event that I wish to submit a story to English speaking magazines or websites, I will have to find a tool that is guaranteed to be dumb. I am sure they exist, but for how long? Will I be forced to translate my stories like a cave men? Is anyone concerned with keeping non-AI translation tools available, relevant, and updated? How can I even be sure that a translation tool does not use AI?
My cousin - who I this very month visited - runs a language school in Italy. He said, very frankly, that he has to do fewer tone corrections of ai translations than those of the humans he used to employ in the same role.
That industry is dead.
There remains a market for English teaching, but it's remarkable to see one of the first definitive terminal career casualties of the ai era....
It's certainly not dead. Good translation (with localization) cannot be automated because languages are constantly changing. It's subjective, not deterministic, especially when it comes to creative works. Even a teacher's opinion is just that, an opinion. There's also a consideration of the author's opinion. Hardly anyone will want to read a translation if they know the author doesn't agree with how it was translated.
This is an important point. I'm not a professional translator, but if I were, I would ask the author for clarification on unwritten subtleties in the source material to ensure that the translation is accurate to their intention. An LLM does not do this (ask follow-up questions).
This would be good practice in the ideal circumstances, but unfortunately it very often is not possible (or at least is not facilitated in any sensible way) for professional translation work, unfortunately.
There‘s a difference between language learners translating something (I guess it‘s something way shorter than a novel and more straightforward than a poem) and an accomplished translator translating a novel, conveying not only vocabulary and tone, but a shitload of context.
The question is, will the demand for „little“ jobs, which would allow a translator to become accomplished go away? Like the junior problem in software development.
Maybe eventually, but so far only humans can interpret signed languages. Part of the reason is that it’s visual, but a big part is not understanding the interaction with facial and body modifiers.
I’ve seen a few papers claiming fairly good results interpreting signing from video, although I’m not a language specialist at all so I’d be the wrong person to judge the accuracy.
What I will say is that automated identification of body pose and facial expression in general is almost absurdly good nowadays, so at least in theory I’d expect processing those as inputs in the same way we do with written or audible words now shouldn’t be too much of a stretch - assuming enough data and resources. It would be a bit more compute intensive thanks to the nature of video vs text or audio, although not prohibitively so given how efficient the pose detection step is: the actual video can be discarded in near-realtime with everything afterwards then using the much smaller detected keypoints.
Honestly the somewhat sad (or relieving, depending on your stance on AI/ML!) truth is that signing probably isn’t currently considered a big enough market to get the research spend, rather than it being a uniquely hard problem per se.
Signs in signed languages tend to have a lot more detail and be a lot more complex, so existing ability to read larger body poses doesn't necessarily indicate that it'll be able to handle the intricacies necessary to translate from a signed language. I've heard of some work in machine transcription of signed language, but it's very early days.
I think the intuition that it's theoretically possible with current types of models is correct, but there is more to this than just throwing the videos directly at a big multimodal model. We don't know that it would be able to translate a sign language to a spoken language even with loads of sign language data to train on -- and even if it could, the magnitude of the data problem is absolutely monumental. The current trend is models that need a lot of training data, and we absolutely don't have even close to that for any signed language. Any solution that would work for signed languages needs to be structured in a way that requires LOADS less training data than current trends in language models.
Then, of course, there's the additional hurdle that translating into signed languages is the more commonly needed component, and generating a signed translation from spoken language has even larger issues that would be incredibly difficult to solve. Plus the fact that any solution ideally would need to be adaptable to multiple different signed languages with considerable differences between them.
The lack of money behind something like this is definitely the biggest factor in why more hasn't happened in this space. But there are some very unique challenges with signed languages as well that do make it a lot more difficult.
This is super interesting context, I really appreciate it. You can tell how much background I'm missing just on the basis that my first thought was sign-to-text rather than the text-to-sign that is, now that you point it out, clearly much more useful!
You've piqued my interest now - especially because video synthesis is a lot closer to what I specialise in - and it looks like this is roughly where we are at the moment: https://signllm.github.io/
To my eye, one of the big weaknesses there is the actual presentation, they're doing it in a way that's fairly low resolution/frame rate and one that I'd consider suboptimal. I think if you envisage the quality we're seeing from Google's Veo or OpenAI's Sora and then map that onto the animated pose skeletons in the link, that's more representative of what should be possible here and now.
Out of interest, when you mentioned detail at the start were you meaning linguistically or in terms of the physical subtleties of the signs? I'm wondering whether the skeleton-based approach in the link above would be better replaced by a mesh approach on the hands specifically, something like https://seanchenxy.github.io/HandAvatarWeb/ (skip to 3:30 in the demo video for examples).
I agree on training data being the likely bottleneck, but I'd actually be fairly hopeful on this being doable. The advantage is that it doesn't need to learn video processing, or language processing, or pose estimation from scratch: those can be done as generic tasks, unrelated to signing in particular, and then a few hundred to a few thousand hours of content (I'm thinking existing video that has sign language interpretation alongside known good audio and text conveying the same information?) should be sufficient to train a model that can do the sign-language-specific parts and then pass the resulting tensors directly into lightly fine tuned existing models for further processing.
[Edit] A nicer example of what's possible in terms of motion synthesis: https://actors-hq.com/ - this is probably more appropriate for a detail-oriented task like this than a full scale Sora-style diffusion transformer would be, now that I think about it. It's also a fraction of the compute, and a couple of years old now, so a much more tractable approach.
[Edit 2] Also, sorry @lou, this is wildly derailing your topic but it was too interesting not to follow up!
I think the issue is that higher-quality general-purpose video generation models may not be able to cleanly present the linguistic content in a consistent way. A successful sign language interpretation model would need to handle the subtleties of the linguistic content and the details of the physical movement without introducing inconsistencies. You wouldn't easily be able to do this by "slotting in" a more general purpose model, afaik, so you need something more purpose-built to at least some extent. I'm actually very intrigued by SignLLM and am gonna read further into it!
Not sure if I got that first edit in before your reply, but yeah I agree, they probably wouldn't be the be best actual tech for the job - I was more going for an example of the fluidity and resolution that's currently on the table, because I think the reduced quality in the SignLLM outputs is enough to harm their usability (but again, untrained eye on the language side!). The radiance field approach in HumanRF would likely do a better job of creating smooth motion with quality that's near-indistinguishable from reality as well as tighter control over the details.
I'm well over the line of excessively technical at this point but there are actually some great examples of existing, frozen model weights being used in conjunction with newly trained layers to adapt a model to a new modality: ControlNet was the original example, way back, and is still used in SignLLM to move from pose skeletons to rendered output - although I actually think AnimateDiff captures the "slotting in" technique best. Their breakthrough was to take a pretrained image model, freeze it, and then add just a minimal number of new layers to turn the batch of generated images into a coherent animation - only the new layers are trained, in conjunction with the existing frozen layers being used as normal at training time to give those new layers contextual examples of the latent space tensor input they should expect. It's a really cool piece of research!
It's definitely not usually a matter of being able to pipe stuff from one model to another in a human-readable form and still get good results, but breaking off the input and output heads of a pretrained model and adding different ones to make use of the broader structures encoded into the remaining 90% of the weights really can get some very impressive results in terms of both quality and training speed. When you actually get down to the granular details a huge number of models now even in cutting edge research papers are bootstrapped from whatever open source weights exists for a similar-ish architecture rather than from a randomly initialised state - and things like autoencoders are shared pretty freely between different models too.
I think you might be pleasantly surprised at how fast the training could go with a sensibly patched-together architecture that minimises the number layers that start with totally unknown state!
This is a really interesting conversation to me because I know a lot less about the image side of things and a lot more on the sign language side of things (though I'm not an expert in signed languages by any means and I don't know any myself).
At least when it comes to the examples on SignLLM's Github, I think they're almost certainly legible enough at the given quality. I don't understand any of the signed languages here, so I can't speak to the quality of the translations, but I could almost certainly transcribe all of them into SignWriting or HamNoSys, even without a ton of experience on my part. Based on that, I don't think the quality of the video is likely to be a major issue here. The quality is definitely not so low that major details are lost. The parts where the hands leave the borders of the vertical video are the hardest part for me, but they may be clearer for more experienced signers or transcribers.
I would be worried about models like AnimDiff distorting details in ways that are linguistically impactful. While the examples on their Github are impressive, their examples with hair and the ocean demonstrate inconsistencies in the physical form of waves, and this doesn't inspire me with a ton of confidence that the fine-grained details in signs wouldn't be lost or distorted by the model -- something that could ruin the quality of the result even if the language model component were absolutely perfect at generating accurate translations into a given sign language.
I do wonder how well such systems would account for, for example the dialect differences in even ASL vs Black ASL (or Black Sign Variation) or even just the interpretation done for concerts vs a news conference.
Not learning ASL younger is one of my big regrets, up there with not learning Spanish
There’s a technique called LoRA (low rank adaptation) that tends to work very well for that kind of granular tuning for style and tone, and doesn’t need too much training data, but that’d rely on a decent base model existing first and then someone with a good understanding of the nuances helping to create the adaptation layers.
Totally technically achievable, unlikely to be considered worthwhile in a commercial setting, so depends on the right person or people with the right overlapping fields of knowledge being able to work together to open source it, I’d say.
I suspect coverage of dialectical variation would depend heavily on the training data's coverage thereof. SignLLM has the very lofty goal of covering multiple sign languages entirely, so it would hopefully handle dialectical variation fairly well, but I'm dubious about how good its translations themselves are based on my very limited understanding and their examples.
Fair, I generally am unimpressed with how all sorts of LLM/AI models handle race - whether parroting racist information, consistently making assumptions that lead to racist or exclusionary imagery, or simply the people building it being absolutely unaware of things they don't know (like Black ASL existing), while thinking they've obviously accounted for everything.
I definitely agree that LLMs have a tendency to flub anything related to race, so I totally get the worry. I wouldn't be surprised if the existing ASL datasets don't have enough coverage of black ASL. That said, I don't think sign language machine translation is good enough yet in general, so I think that's the bigger hurdle atm.
Makes perfect sense!
Likewise, it's really cool to get insight on a new potential use case for the kind of tech I'm familiar with from someone who actually understands the needs!
Totally makes sense on the detail concerns with AnimateDiff - I don't think it'd be the optimal base model to use, I was throwing that in there more as an example of the kind of results that are possible from reusing existing models with limited additional training (in this case existing still image model turned into an animation model, as an analogy for turning a spoken or written language model into a sign language one), but I also realise I have a habit of doing this a bit when it comes to visual model stuff so I definitely could've been clearer there!
Either way, the fact that SignLLM is already getting to the point of legibility is great to hear. Part of my reason for asking was wondering whether expanding on their work with some improved visual output would be a potentially worthwhile use of time, and I've made a note to dig a bit more into their paper just to see if there are any quick wins I could potentially contribute back their way, but honestly if their fairly basic visual synthesis approach is doing the job fairly well that already sounds like good news to me.
SignLLM's legibility definitely seems sufficient to me, but I'm also kinda dubious about its linguistic competency. The actual translation between languages is definitely the most difficult component (especially with how small the sign language data is), and I wager it's where the biggest improvements are needed. I don't know much sign myself, but even I can tell that the man in the generated DGS example signed "Belgien" with the wrong handshape (he used an open palm, whereas the actual sign uses the B handshape with the thumb crossing the palm). It's possible that this is some under-documented regional variant, but I kinda doubt it, and it makes me worried about how SignLLM handled other much more nuanced aspects of translations. I wish I knew someone on Tildes who was fluent in ASL, because I struggled to look up the signs used there in either English-to-ASL dictionaries or reverse dictionaries.
If enough publications ban AI translation, I suppose there might be enough users for at least one small dumb translation tool that everyone can trust.
Arguably, text translation (quickly and easily verified by a bi-lingual) is amongst the best use cases for ai, practically and ethically.
They are language models after all.
Most likely all translation tools fall under the category of artificial intelligence. Google Translate has used some form of machine learning for at least a decade, and has received updates in its techniques as new technologies become available. Other tools like Grammarly also make use of machine learning to power their models. They may not use transformers currently, but still use RNNs and other AI approaches in their models.
It's likely that even basic grammar and spell checkers in document editors are now using some limited form of AI that might be considered an "assist", per Clarkesworld's rules. Unfortunately, AI is a very widely-scoped subfield of computer science, and includes essentially any form of computer decision making which covers a lot of useful tools.
I assume that Clarkesworld means a popular understanding of post-GPT LLMs specifically and not a broader definition that is more academic or specific to the computer science field.
You’re probably right about them having a specific meaning in mind, but this is one of those cases where I think their lack of clarity is actually a genuine problem for anyone trying to follow the rules in good faith.
You’re seeing that issue yourself in the comment just above - maybe their intent was “this needs to be translated by a human”, but as written the rules exclude you, the bilingual author and absolute best-placed person to write the translation, from doing so in your preferred way. I think that’s highly unlikely to be something they consider a good outcome, but the ambiguity is pushing in that direction.
In general I’ll happily be the technical pedant pointing out the loopholes when a poorly written rule or law needs to be undermined - but this isn’t that. This is a situation where I think following the spirit of the rule probably makes sense, but it’s actually hard to parse out what the spirit is.
I guess what I’m trying to say is that “AI” is a poorly defined term, but even by common understanding it would apply to pretty much every machine translation software that actually works, even ones that long predate ChatGPT and that broader use of AI as a term. And I’m not saying that in a “well, actually, if we’re being technical about it…” way, I’m saying it in a “they’re doing roughly the same thing under the surface by any meaningful layman’s definition, even if the people writing the rules didn’t realise that” way.
I think the only real option here is to contact them for clarity:
Is their intent to avoid AI slop? If so, a human translator in the loop - i.e. you - is the perfect outcome and your using modern tools to assist shouldn’t be a blocker.
Is their intent to enforce an ideological objection to AI? If so, do they understand and intend that as written their rule would exclude situations like you, the author, using Google Translate as a starting point? Do they also understand and intend that it would apply just as much to someone who used Google Translate in the same way in 2016, long before ChatGPT even existed? And that it would exclude any paid, professional, human translators hired to do the job from using similar tools in their workflow?
Maybe the intent really is option two and they really did think through all of those implications, I don’t know, but I’d speculate that the technical nature of the question and the general fuzziness of what “AI” actually means is leading to some unintended consequences that they’d be happy enough to rethink and clarify.
Unfortunately on the internet people writing the rules often have to resort to overly broad rules like this. See also my comment here where I mention it a bit. You are entirely right that it becomes a hurdle for people acting in good faith. At the same time the people not acting in good faith, specifically rule laywers tend to take up an extraordinary amount of work and time in these situations.
I'd also say that in any case where an author has used a general use LLM to do a small aspect of some translation, double-checked the translation, had it proofread, etc it simply doesn't matter anymore. Because we have no reliable way to detect LLM writing other than the fact that for large pieces of text the quality simply is low. Which, as I said in my other comment, is already something they also look for.
So, in my mind, this very likely is written up to be overly broad and have people err on the side of caution.
Yeah, that makes a lot of sense. I don't envy anyone who has to come up with a reasonable policy on this kind of thing and expect it to survive contact with the realities of a rapidly evolving technical landscape and the realities of the general public at large!
Even if that is the case here, I would be shocked if any machine translation is not leveraging an LLM somewhere in its pipeline these days. That especially goes for Google, but I'm fairly certain other vendors in the space are also using similar models. It's almost certainly impossible for you to verify that they aren't.
DeepL has started using LLMs, specifically models they trained themselves. Which makes sense as LLMs, as far as I know, are closely related to the machine learning used in the past decade (NMT). The biggest caveat of general purpose models like chatGPT is that the amount of training material used for different languages varies wildly.
I'd assume that when they train LLMs specifically for translation this is accounted for, making sure that for all languages there is sufficient training material used.
Yeah I'm not surprised DeepL uses LLMs, and their machine learning has been so much higher quality than the competition for most languages that I'm glad they're training them themselves -- they clearly know what they're doing. Training their own models also means they can avoid wasting time on aspects of a more general-purpose model that they don't need for a machine translation product (such as the conversational chatbot style for something like ChatGPT, for instance).
I believe that they intend human translation to be the solution, as the stories I've seen have a human translator. (And I do mean like a paid translator)
That is certainly a barrier to entry but it feels realistically like the correct option for literature.
Where We Stand on AI in Publishing
The reason for their stance
It Continues
And if you look at any of their "translation" tags, you'll see at least one human translator credited
Yeah. Literature feels like a different case compared to more technical writing because it can have a lot of nuance that can be lost in translation, whether it be cultural or things like wordplay. Interpretation is VITAL for literature since it can change the meaning and tone, and I'm not sure AI would be able to catch some of that.
My main exposure to this aspect of translation is manga, where many fan translators will often explain some context or give insight into their choices. One of the most impressive translations I saw was for Assassination Classroom, because that series is rife with puns relating to assassination. Even the teacher's name, Koro-sensei, is a pun on the Japanese word for "unkillable" (korosenai) and sensei. I was endlessly impressed by the fan translators finding suitable translations for all the puns that came up.
Then when it comes to sentences with multiple potential translations... Well, My Hero Academia has possibly my favorite example. During a scene where everyone was giving the protagonist a "come to Jesus" talk, one character (who is very loathed by the fandom for being a pervert towards women) told the main character he admired him.
The problem: He specifically used the word "horeta" (惚れた), a verb that can mean deep respect and admiration. But the first (and often only) definition to come up when looking for English translations? "To fall in love with."
So guess which one the official translation used?
Yeah, the fandom was not happy that the first "official" bisexual character was the pervert. At least we got some hilarious memes and edits though.
Context seriously matters when interpreting sentences in fiction. And I'm not sure AI will be able to always detect and account for that context when translating.
It works if you are the author and fully verify, adapt, and fix the translation. So in my case it is only supposed to be a first pass to avoid some grunt work.
How do you “fully verify, adapt, and fix the translation” for a language you don't speak?
Edit: apparently I didn't fully pick up on the context. I wasn't referring to the OP or any particular author, nor was I singling out English or any other particular language. I was just saying that “the author can just vet the translation” is not an option in most cases.
I feel like lou's ability to speak, at least write, English is fairly good. Given that they are doing so in the context of this thread and everywhere I encounter them on Tildes ;)
If you speak a language, just not natively, it is very easy to come across things where you just aren't entirely sure what the best way is to translate it from your native language. Using a translation tool here is often more akin to using a thesaurus as a native speaker. To find a word, or sentence construction, that might be a slightly better fit than what you yourself would have used or thought of.
Verifying the translation can come in various forms. If it is a single word I might google it as well, look it up in dictionaries like, Merriam-Webster or the Cambridge Dictionary. With sentence constructions you can typically ask for the reasoning behind them, then verify it against online resources.
Then of course there are sometimes native speakers you might be in contact with. If you are wondering why someone might not ask that group directly, having a more targeted question can be useful.
Your assumption seems to be that someone might use such tools to translate an entire text wholesale. Which is just one end of a spectrum where people might employ translation tools.
This is why a professional translator is preferred though, they should have the knowledge/experience to do it without a "first pass" by an LLM. And that is the standard that Clarkesworld has set. Other publishers will have other standards, but I don't understand what feels like a vibe of "go ahead and do it, they can't tell." This isn't a huge corporation, they're just trying to keep their own heads above water while still publishing authors from around the globe.
There is a lot more nuance to this than simply saying "you should just hire a translator". There is an entire spectrum of (partial) translating things where using a professional translator is complete overkill.
I know the original context is about translating entire works. If we were talking about throwing an article book/article through an LLM and using that without any adjustments I would completely agree with you. But that isn't what I am talking about, which I had hoped would have been abundantly clear already.
I was responding to this question.
Challenging the assumption that people only use these tools for aid in translation when they don't speak the language. Which is far from the truth and also what makes this much more complex.
To take myself as an example, English is not my native language. There have been instances where I wanted to write a very similar comment or even article in both my native tongue and English. Sometimes I will write the comment first in English, and sometimes I will write it in Dutch first. Then when translating it I might run into a situation where I can't quite recall the right words, or sentence structure. A LLM can be incredibly useful in those instance, specifically where I know there is a specific word or saying and traditional direct translations come up blank as they are too literally. Here I have no qualms over asking an LLM to give me some suggestions based on a description and context and after doing my due diligence in verifying it use the result in anything I write.
Again, because that is not what I am advocating for. What I am describing is a situation where someone writes, say a 10,000 word piece. And maybe for a few tricky sentences they use an LLM to get an initial phrasing or translation. They then check it, revise it, and make sure it fits. At that point, the involvement of an LLM is minimal such that, to me, it is hardly relevant anymore. As I said before, to me, this puts this at the level of an assistive tool similar to someone maybe using a thesaurus.
Now, I am also fully aware that in the context of Clarkesworld magazin submissions this can lead to a slippery slope kind of situation. Where the question becomes "how much LLM use is too much". So, as I said in other comments, I think I fully understand why they have set the rule as strict as they did.
But that is also not really relevant for the question I set out to answer. ¯\_(ツ)_/¯
I understand the specific context you were answering in - the vibe I was referring to is less about your specific response and more throughout the entire thread.
I also definitely don't think it's easy or cheap to "just" hire a professional and I don't think I used the word "just" for that reason, but I think that authors will find different standards in amateur vs semi-pro vs pro magazines and zines and will have to be able to adapt.
Otherwise we broadly agree.
Right, frankly, this confused the heck out of me. I went back several times to see if I didn't somehow imply it. Certainly given the comment you directly replied to.
Overall I can see where you are getting the vibe from. Though, personally, I get a slightly more nuanced vibe. I don't see anyone here advocating for just using LLMs. The comments that do lean towards the "they can't tell" aspect do that with the clear distinction and expectation that is being done in a context where the author puts care and work into making sure it is correct.
Anyway, vibes are personal. I mostly just got confused about it being applied to this specific comment thread.
No worries, it was just the comment I was making when I had the overarching feeling so apologies for the confusion and lack of clarity.
And I know the comments generally were intended as "with care by author" but perhaps I'm just too anti-LLM and too inherently rule-following to feel comfortable with the balance they were striking in the context of responding to a prohibition of the use of those tools.
I absolutely don't like "they can't tell" as a justification, but I think it's very reasonable firstly to question whether Clarkesworld really did mean "no Google Translate" or if the broader intent was "no autogenerated slop".
And if they did mean "no Google Translate", I think it's fair to further question that decision in cases like this, where it's an assistive tool that allows a fluent but non-native speaker to be published rather than excluded by the cost of having to pay for a native speaker's time.
I think that's a reasonable question, and one that could probably be solved by messaging them directly on Bluesky or by email, because it won't be solved by us all guessing and Neil/whoever else uses their Bluesky has been pretty responsive.
But I acknowledged it is a barrier of entry, but so is not allowing LLMs at other points in the process, and the magazine paying a real live editor and using real live readers to filter the stories. But I don't think it's an unreasonable barrier.
Replying to myself for visibility
From Clarkesworld on Bluesky:
The question with answer in replies
Clarkesworld reply:
It also looks like they did a Kickstarter in partnership with an org in China that allows for them to cover (subsidize?) translation costs for Chinese stories.
That sounds like a very reasonable take - strongly discouraged for reasons of quality makes sense to me, but discouraged isn’t outright banned, which leaves enough space for someone like @lou who is confident in their own English to use it as a tool.
I'm not sure they wouldn't ban it if they found out, but it'd be worth asking if someone tries to submit rather than risking getting banned from submission entirely.
They're very ethically opposed to its use as well and didn't address that so I can't tell where their line would be ultimately
Just for clarification, my comment didn't say “lou” nor did it say “English”.
I do speak English very well.
I'm sorry for any misunderstanding, but I was referring neither to you nor to English specifically.
At least for now, and particularly for Clarkesworld given their experiences, it seems like the only correct option. I don't know if, for example, English language authors experience the same thing trying to be published in other countries (in this magazine/short fiction format), or if it's too infrequent to know
Yes, Google Translate is AI according to most definitions.
Background: natural language translation is one of the original problems that AI research attempted to solve. Google Translate launched a lot earlier than AI chatbots, but chatbots use algorithms that are direct descendants of the ones used for translation.
I think having your work proofread by someone who knows English well is probably your best bet.
AI chatbots started in the 1960's. "AI" has being continuously co-opted for marketing purposes for decades now.
I should have said, modern AI chatbots starting with ChatGPT that can chat about a wide variety of topics. There have been lots of chatbots before that, but they were often quite narrow and implemented using traditional programming. (Eliza-like bots are mostly a lot of if statements.)
@cfabbro can you please change AI to LLM in the title? Otherwise people will debate on what is AI and that's not really why I created this post.
Thanks ;)
Done.
I replied in more detail above, but I can't resist a good opportunity for an analogy: when the photography magazine banned digital modifications, did they mean "no Photoshop", or did they mean "no digital cameras, only film that you've developed yourself"? Did the sculpture magazine mean "no 3D printing and CNC", or did they mean "no power sanders or band saws, only things you've made with a hammer and chisel"? On both, I'd gently lean towards assuming the former but could totally believe the latter - especially when the rules are covering something as varied and subjective as creative work.
You're quite right to want to avoid Tildes-level mega nerds like me debating the deep technical minutiae on this one, but I do think the question as a whole hinges on what the person who wrote the rule considers to be AI - and, more broadly, whether they realise that the definition most people would give actually applies to a whole giant swathe of tools, including quite a few that are a decade or more older than modern LLMs.
Basically, I think we have to parse their intent to get to a reasonable answer, even if that answer doesn't mesh with whatever technical definition I might give.
Frankly, that strikes me as a backwards and idealistically naive policy. And also one that can’t really be enforced.
I assume they’ll reject submissions that are mechanical in tone or read like stereotypical ChatGPT output. But I very much doubt they’d flag anything that has genuine care put into it. If you’re writing content yourself, and it’s good, and then use AI to translate it with more than the laziest barebones effort, you’ll probably be fine. More specifically— there’s a spectrum between literal word-for-word translation and loose paraphrase. I would suggest prompting for something closer to the former, with exceptions for untranslatable idioms. If you give an LLM too much creative freedom it’ll start to lean into those AI stereotypes. And don’t just say “translate this into English” because it will do the same. You want it to preserve as much of your original meaning and intent as possible. If you’re feeling doubtful about it, you might ask a native speaker to read through the resulting copy afterwards to verify that your human personality does in fact shine through it.
Of course I won’t be held liable for your intentional disregard for Clarkesworld’s policies… 😉
Deliberately breaking the rules that a magazine sets for submission isn't a great way to continue to get paid by them, or to not end up with a poor reputation if "caught" and struggle to be printed.
Clarkesworld has had to deal with a ridiculous number of straight-LLM-written submissions and shut down temporarily (unclear at the time if it was going to be permanent). The rule is totally reasonable given their history.
I really hate breaking rules :P
Seems like an overly broad policy of clarksworld to me. Specifically when translation is concerned as in that area machine learning has been in wide use for well over a decade now (NMT). In fact, if my understanding is correct, LLMs are closely related to NMT and in some way an evolution of concepts used in NMT.
The caveat of general use LLMs is that they highly depend on being trained on both languages being in the training data to a meaningful degree. Where specialized translation models (even if they use LLM technology) will be trained specifically with that goal in mind.
There is also no realistic way they are going to be able to detect that you did use a general purpose LLM to translate certain things. I'd say that for entire texts any translation tool will still fall short and any submission that was entirely machine translated would simply be rejected because of its extreme poor quality. Even more so given this statement:
My gut feeling here is then that they simply don't want works that are majority generated by generic LLMs or wholesale translated. I highly doubt they are going to care if you used chatGPT or any other LLM to translate a specific word or sentence fragments. But, putting that sort of nuance in rules on the internet is just going to invite rule lawyering. So it is easier to ban these things entirely, so people err on the side of caution making it overall easier for the folks reviewing submissions. See also @RheingoldRiver's comment here.
Many commentors here mentioned that the LLM today are related to machine translation. I recently learned that the paper that introduce Transformers (the "T" in "GPT") to the world wrotes:
I read this first on Hacker News where people mentions that LLM could be seen as a machine translation tool that translate English to English or question to answers.
Theoretical discussions about AI are stimulating. In addition to that, I would also appreciate some suggestions for a dumb
non-AInon-LLM translation tool I could use.Given that some tools are likely no longer updated, at least one must still exist in that state, even if only temporarily. I would appreciate it if anyone had a practical suggestion in that regard.
The tool doesn't even need to be particularly good, it is only supposed to do a first pass to avoid some basic, mindless grunt work. Given that I am the author, I have the liberty to (and will) fully verify, change, adapt, and add to whatever it outputs.
I have no intention whatsoever to break or bend any rules, please keep that in mind.
Thanks!
EDIT
Please keep that in mind:
Secondary reply, based on your emphasised edit: if you really don't consider that assumption to be a case of rule-bending, you could try out https://libretranslate.com/, which is based on https://opennmt.net/
It's still a machine learning model, it still bears all the technical hallmarks of a proto-LLM, and I personally would consider using it to be "using AI" not just in an academic or computer science sense but as a literal example of the same tech that just happens to predate the popularisation of the term "AI" in public (fuzzy and imprecise as common usage of the term is). I consider that far more rule-bending than relying on the defined difference between "discouraged" and "banned" and using Google Translate.
I'm not saying any of this for the sake of argument, I genuinely think there's some miscommunication happening here about what distinguishes an LLM from a translation model and how that interacts with the letter and spirit of their rules. But I'm not the one making that call, you are, so I went looking for potentially useful open source tech either way!
Great suggestion, thanks!
To the best of my knowledge there aren’t, and weren’t ever any worthwhile machine translation tools that didn’t use machine learning.
To translate a language in a meaningful way, the system has to model the language. This isn’t me playing semantics; a machine translation tool is literally a language model, even though the general public didn’t know them by that name until ChatGPT came along.
I get that a lot of the threads here probably seem a bit frustratingly tangential to your original question, but I really don’t think it’s possible to give you an answer naming any specific tool that doesn’t inherently raise the “is or isn’t this AI?” question and step quite firmly into rule bending territory if AI tools are outright banned.
Honestly I think the most useful and direct response in the thread is this one, linking to a thread where the magazine editor clarifies that machine translation is strongly discouraged for reasons of quality but not actually banned. I personally wouldn’t consider it rule bending at all for you to make your own judgment call and continue to use your preferred tools despite them being discouraged.
Nitpick, transformer model. But models that needed to be trained regardless.
Counter nitpick, Google Translate was using an LSTM at least a year before the original transformer model paper was published, and some newer tools are experimenting with moving to state space models rather than transformers ;)
I meant more that they were not called language models. But fair enough, the technology is very closely related, to the point that arguing about the differences is not all the relevant.
This is a difficult ask, given how broad that definition is. As I said elsewhere, google translate and tools like deepl have used machine learning in the form of NMT for almost a decade. Before that a technique called SMT was used which, is effectively also a form of machine learning. I am pretty sure companies liked to call both AI back than as well. The truth is that translating text, including proper grammar and context is a difficult task so there aren't really "dumb" tools out there. Unless you just take a dictionary to the task and go by literal translations.
What I am getting at is that your question is way too broad to give a reasonable answer.
If you don't want any "AI" then changes are that you will find there is more work in fixing all the translation oddities compared to doing it yourself.
From your start post
Isn't this an opportunity as well? Your grasp of the English language is clearly good. It might feel daunting to not use any translation tool as a first pass. But it also might present an opportunity for you to further refine your own skill set by translating larger bodies of text directly.
With that out of the way, if you are just looking for "not LLM" then your best bet is likely anything based on OpenNMT, specifically anything making use of Ctranslate2. Here is a desktop application I found that wraps around Ctranslate2. The caveat is that you need the proper transformer models for the languages you are using them in.
For SMT based translation there are also still tools around like the Apertium project they also do have a demo online but it is fairly limited.
Honestly, I still think you are approaching this in a way that isn't productive for yourself. Considering the highly specific bit you are replying to and the fixation on needing/wanting to use a tool. I might be wrong here, but I feel like that is more likely to lead to the sort of text they don't want.
I might be wrong entirely, but if I am not I'd really would like to challenge you to have a go at not using automated translation as a first pass.
Edit:
Response to your edited in bit.
The way a machine translation translates entire sentences can subconsciously lock you into a certain pattern. Even if you go over it by hand to correct individual things. At least that is my experience when I have used both NMT based translation and LLM based translation for entire texts. They tend to stick to the pattern of the original text, even if the target language generally would structure sentences and sometimes entire paragraphs differently.
So, the times I used these tools do translate an entire text instead of just a sentence I often found myself rewriting that text to such a degree that I might as well could have done it myself.
Anyway, just my 2ct.