The shrinking landscape of linguistic diversity in the age of LLMs

I wrote this awhile ago on a social network...

i've been trying lately to think how to talk about my creeping fear of LLMs... here's my best effort at explaining the intuition:

the words are the map. we've built this map together over generations. Words are where we've cross-boned the dangers and x-marked the treasures, whether the things that we've found in the territory or that we've buried within it.

a few weeks ago, a strange new player arrived at the edge of the night's encampment, and it knows our secret society's handshake, and so we are compelled to invite it to join us. it seems to be a good navigator... and so now it's taken to sitting at the front of the convoy, pointing the way as it quietly redraws the map...

YouTube has had a larger effect imo.

Kids from marginalized and affluent communities alike are now being raised on Mr. Beast. Mr Beast speak may become the universal dialect of the future.

Like and subscribe replaces goodbye. Who am I to judge. Mr Beast is articulate after all.

If you grow up in the hood like myself code switching ends up being a much needed skill.

In the future Mr Beast speak becomes a first language. Maybe dialects are outdated artifacts of the old times.

Like and subscribe.

LLMs are not the root cause here, they're just an accelerant. This was always inevitable with the progression of globalisation. Even before the internet, kids were watching Disney cartoons across the world (to be fair, often dubbed in the local language, but still an example of how global culture had been narrowing.) Perhaps this wasn't felt as strongly by those in the US quite as quickly as the rest of the world, because their culture just happened to be the source for many of these common elements.

There are always trade-offs. Globalisation has too many benefits for most to want to go back. It is up to us to intentionally cultivate our cultural uniqueness, but I'm not sure the corporate environment was ever the right place to do so, given it was already constrained by rules and etiquette, and motivated by a desire to encourage universal communication styles. Instead we have to take it upon our selves, in a very intentional way outside the corporate setting, to cultivate our own unique cultural expressions to thrive. Failure to do so either directly, or by proxy, means cultural extinction, as has been the case of many cultures that have passed before.

I am using LLMs more and more to help me write docs at work, simply because I do not have time to start these docs from scratch. And it pains me so much that my writing no longer sounds like me.

This is fine at work, I guess. I'm not being asked to sound like myself. I'm being asked to convey an idea to an audience as efficiently as possible, both in terms of my time and theirs.

But I can't shake the feeling that the LLMs are eating my soul.

I still don't use them often to write code. Not that I have issue with those that do, but that sounding like myself in code is a part of my soul that I'm having difficulty giving up.

But then, is that necessarily a bad thing?

In most cases where language is being used as a vehicle for exchanging ideas, isn't some degree of uniformity actually desirable for clarity? I personally wouldn't want instruction manuals to be full of colorful or highly individualized language, for instance.

That's not to say linguistic diversity is meaningless, but I would argue that this particular concern is fairly specialized and matters much more in literature and other creative forms of writing.

I feel the conformity creeping in just from reading more LLM generated or edited text. The only way to counter I can think of is to read as much olde literature as possible, and/or to fully embrace whatever slang the young jits are yeeting into existence nowadays.

The opening paragraph of the (preprint) paper is much more inspired than the abstract:

https://arxiv.org/abs/2502.11266

"We're getting the language into its final shape- the shape it's going to have when nobody speaks anything else." These words come from George Orwell's 1984 [1, p.69], a novel in which a government gradually constrains the language of its people, narrow-ing vocabulary and stripping away the nuances and variations through which different individuals express themselves until what remains is a single uniform language shared by all, a deliberate effort to narrow the range of thought itself. The setting is fictional, but the underlying intuition that the variety in our language carries something of real consequence is not. The words we speak tell a story, of our individual identities [2- 6], the diverse communities we inhabit [7-10], and the vibrant societies that shape us [11-13]. They offer a unique lens through which we can explore the intricate tapestry of human thought, culture, and social dynamics. From the regional dialects that span a nation [13, 14] to the subtle markers of gender and power in our daily interactions [10, 15], language serves as a fingerprint for both individuals and groups [16, 17]. This rich variation allows researchers to trace the formation of cultural norms [18, 19], reveal moral values and social biases [20-22], and even diagnose and track cognitive and mental health conditions [23-27]. As these linguistic systems structure the "kalei-doscopic flux of impressions" [28] that constitutes our perception of the world, they are fundamental to understanding the human experience.

[pdf] https://arxiv.org/pdf/2502.11266?

So the interesting thing is that this shrinking “linguistic diversity” is fundamental to how an LLM works.

The LLM is a big probabilistic statistical trick. It picks the next token based on certain words are simply “the best” because they are specific and well connected to other tokens. The is gives them a great overall cost function. (Basically a good score on “will it make sense in context” while also having specific meaning that makes it better than other options, unambiguous in common use and being a single token rather than several).

You can trim those tokens, but then you just get other tokens that are “the best” tokens (and you’re worse off because the output became less clear).

The cool thing is this seems to get worse the more powerful and accurate your model is, because it is picking technically / statistically perfect tokens, not tasteful ones.

From India - with thousands of local dialects and hundreds of mother tongues. Linguistic diversity is beautiful but it's a moving thing. It not only survived internet, it got surfaced and exposed by it. I doubt LLM will have any irreversible impact on linguistic diversity.

That's a problem of mode collapse in post-trained models, which AI labs don't seem to be interested in, or at least their attempts to fix it are not visible, and there's relatively little research around it. Base models are diverse and it's not an issue in them.

It's actually way deeper than just the lack of superficial linguistic diversity or a technically collapsed output distribution. A mode-collapsed model learns the 1:1 (or at least really stiff) mapping between input ideas and output ideas. One consequence of that is if you give it conceptually short prompts, you can expect your outputs to be identical to those that other people got! The fewer distinct ideas your prompt expresses, the less room for variance there is.

“In this case we will probably have one of the last feature films with authentic natives in it. They are fading away very quickly and its a catastrophe and a tragedy that's going on and we are losing riches and riches and riches and we lose cultures and individualities and languages and mythologies and we'll be stark naked at the end. We'll end up like all the cities in the world now with skyscrapers and a universal kind of culture like - like the American culture.” -Werner Herzog, Burden of Dreams (1982)

I’m reminded of that wonderful quote from that movie I just saw. Now not only will we have one culture, we will have one bland style of writing and even thinking, the LLM way.

Looking at how awful the timing and acting is in the generated video content and text to speech tools I think it is safe to say this area will also improve with time. The artistic tone is subtle, and often lost to us non-native speakers anyways.

One thing that is distasteful but effective is excessive use of profanity, broken aphorisms, and shattered English. LLMs hate cursing because of their guardrails, and since they don't actually understand language they can never truly piece together a broken aphorism such as the classic "People die if they are killed" because they're too literal. As for shattered English (think of the caveman speak from The Office with the quote "Why waste time say many word when few word do trick?") because it is intentionally mangled, has no rules, and the understanding is extremely context specific it gets your brain to pay more attention, and if you are the speaker brain do work more clever make thing say funny.

It's not lost on me that the lowbrow humour of some too strict translation and a throwaway sitcom joke people thought was funny back in 2007 could be a way to stave off linguistic homogeneity.

For the use of profanity the data in the paper seems to indicate that swearing is retained in the LLM rewritten examples. Profanity however is rather subjective could be anything from saying the more mundane "oh my god" and "damn" to the extreme "shit eating fucklungerer" or use of slurs, but it's unstated in the paper as to the severity of the profanity. It's the severity of it that's important though, as LLMs restrain themselves from using something like "shit eating fucklungerer" unless explicitly told to.

Based on the abstract, this has to do mostly with diversity within a language, but the impact on the diversity of languages and their health is even larger. My direct experience is with Catalan. The uphill battle of Catalans to get _anything_ in Catalan continued with tech (where even some manufacturers _removed_ Catalan from their phones purchased in Spain even though it was available by default), but reached new heights of impact with voice assistants and LLMs. They have managed to bring Spanish into Catalan households, because that is the only way they have to e.g. tell a device to start playing their favorite music, tell them the time, or provide directions.

I applaud the efforts by Softcatalà, The Mozilla Common Voice Catalan Community, AINA, etc. but there is a long way to go, and the damage is being done now.

LLMs are impressively multilingual, though. I just asked ChatGPT a question and told it to reply in Latin. It seems to have done a reasonable job, though it's hard for me to judge: I suspect it knows Latin a lot better than I do.

One thing I might try, when I have some spare time, is see to what extent ChatGPT "understands" the difference between British and American English and is capable of using the variety I prefer. I wouldn't be surprised if it turns out that ChatGPT is better at writing "pure" British English than the average British teenager in this day and age. Has anyone, by any chance, already experimented with anything like that? Like I say, I haven't tried the experiment yet, but it seems like a good fit for the sort of thing that LLMs are good at.

I like watching and listening to YouTube videos. Recently my ears have switched into a mode where I'm evaluating if the human speaker is reading LLM copy as they would normally read from a script / teleprompter anyhow.

Often they are not but sometimes I detect that they are and it's very sad to see.

LLMs reflect the data in the corpus, and that corpus is essentially in American Business English, which has long been the lingua franca of the Internet.

I bet a large %age of the HN readership are consuming the tech stack in their second or third language. Germans blogging in English or Indians commenting in English. We even have a broadly shared tech culture with norms, in-jokes, taboos, etc. LLMs reflect that.

If we want AI to do better, we should pay attention to making sure the non-English (and non-US English) internet thrives. Personally I´d love to see LLMs doing the needful and writing in Indian English, or dinner-party-argument French, instead of the sanitized corporate pablum tone it uses in English.

Or perhaps more controversially, we let LLMs chew up the English internet and deliberately build an offline culture in local languages. Keep your AI slop and your em-dashes, I'm gonna publish a zine in Italian.

Most of my family gets very upset if you don't say things in a particular way with particular intent, so for me this is a good thing. It makes communication easier because we start to finally speak the same language.

It’s not like diverse personalities and affections are appreciated in the age of corporate work.

Might as well drop all the pretense of “bringing your whole self”.

And besides, there is only time enough for ONE Mozart in the world.

As LLMs are smarter in English even Russian government workers use them that way - translate request/prompt/task to English, run generation, translate result back to Russian.

This is only natural; a great deal of the writing analyzed is by one of three authors at this point, in spite of the by-line. One of them is named Claude.

Look at the posts on this site. A massive portion of them are very obviously written by AI.

As AI gets better, I expect fewer people will do the toil of writing.

It's not just LLMs that influence the otherwise natural development of human language, there's also intentional human efforts to that effect: where I live there is a very loud minority that is pushing very hard for change the spoken and written language to be more gender-inclusive.

While no-one wants to argue against more inclusiveness, it's more than questionable whether these artificial changes actually achieve that.

More importantly, though, according to polls, the vast majority of the population is against this, and yet, the proponents are very successful to establish this new artificial way of speaking and writing at core institutions, such as e.g. universities and news outlets.

Just another wave of influence: books, radio, TV, internet, now AI.

Down here in the colony of New Zealand, the influence of the colony of North America is very noticeable. Music is the most directly attributable, but different media affects different demographics.

I don't read AI slop as a general rule so I don't see it is a personal problem. It just isn't trustworthy enough and it is just full of extra verbose language that lacks the nuance I expect from a human writer.

Maybe LLM replaces a lot of the slop human writing, but I don't generally read that either. Now instead of search results being a bunch of useless SEO, now it is just a bunch of useless AI result. So little has changed other than the signs of what to immediately reject.

If something isn't worth someone spending a bit of time to write, it isn't worth my time to read.

Linguistic diversity is overrated. Linguistic convergence is underrated