Position Paper: Losing our Tail – Again: On (Un)Natural Selection And Multilingual Large Language Models
Eva Vanmassenhove · 2025
Read Position Paper: Losing our Tail – Again: On (Un)Natural Selection And Multilingual Large Language Models by Eva Vanmassenhove (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-03933.txt.
Introduction
Eva Vanmassenhove presents a 2025 position paper warning that multilingual AI systems may flatten the world's languages. Earlier tools helped people write; today's large language models increasingly write for them, gaining direct power over language itself. The risk, she argues, is a slow loss of rare words, grammar, and cultural flavor.
The mechanism is model collapse: machine-made text feeds back into training data, shrinking the spread of language toward its most common forms. Translation tools are a special worry, since they can sand down idioms and local constructions across many languages at once. Drawing on work in vision and language technology, Vanmassenhove urges the field to treat expressive diversity as something to protect, not average away.
Catalog record
| CREATOR | Eva Vanmassenhove |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 43 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-03933
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Position Paper: Losing our Tail – Again: On (Un)Natural Selection And Multilingual Large Language Models Eva Vanmassenhove Abstract Multilingual Large Language Models (LLMs) considerably changed how technologies can influence language. While previous technologies could mediate or assist humans, there is now a tendency to offload the task of writing itself to these technologies, enabling them to change our linguistic ecosystem more directly. While they provide us quick access to information and impressively fluent output, beneath their apparent sophistication lies a subtle, more insidious threat: the gradual decline and loss of linguistic diversity. With this opinion piece, I explore how model collapse, with a particular focus on translation technology, can lead to the loss of linguistic forms, grammatical features, and cultural nuance. Model collapse refers to the eventual consequence of self-consuming training loops, where models reinforce their own biases and lose linguistic diversity. Drawing on recent work in Computer Vision, Natural Language Processing (NLP) and Machine Translation (MT), I argue that the tails of our linguistic distributions are vanishing, and with them, the narratives and identities they carry. This is a call to resist linguistic flattening and to reimagine NLP as a field that encourages, values and protects expressive multilingual lexical and linguistic diversity and creativity. 1 A Simple Evolutionary Trade-Off A few million years ago, we lost our tails. Once useful (and presumably quite fun), they became obsolete, absorbed by evolutionary trade-offs. They disappeared through natural selection. Darwin already drew parallels between the evolution of languages and those of species stating that they follow ‘curiously parallel’ paths in The Descent of Man ( Darwin 1871 ) . He went so far as to suggest that “ the survival of certain favoured words in the struggle for existence is natural selection ” ( Darwin 1871 , p. 61) , highlighting how language itself is subject to various evolutionary forces. However, it is unlikely that even Darwin could have anticipated the current path of language evolution and the latest forces that came into play. Today, it is not just ‘natural’ 1 1 1 While the technological forces that shape language can certainly be seen as part of the ‘natural’ forces influencing our linguistic ecosystem, there is a crucial shift when we move from tools that help us record, polish, or disseminate language to models that actively generate it. In such systems, the technology becomes not just a medium but a core driver of language change. I will return to this later in the paper. forces but technological ones that exert strong selective pressures on languages and influence the struggle for linguistic survival on various levels. Given that Multilingual LLMs are essentially next-word predictors over large datasets, they amplify statistically likely languages and linguistic forms, whether these are words, subwords, syntactic constructions – pruning away our rich, long statistical tails. Will this technology-driven artificial selection merely accelerate an already inevitable trajectory or does it mark a more disruptive turn? In other words, does losing our ‘language’ tails imply the disappearance of obsolete, fun and decorative elements or does it signal a more fundamental erosion of diversity and richness? A quick note on ‘language tails’ In language, as in other complex systems, there are various long-tailed distributions: a few elements are very frequent, while most are rare. These “tails” appear across multiple linguistic levels. For example, a word like ‘the’ appears more often than ‘disperse’ or ‘eloquent’. Yet, these words are fully grammatical, meaningful and frequently used in specific domains or contexts. These tails appear across different levels of language: morphology, syntax, phonology but also on a cross-linguistic level. 2 Natural Language Selection Before turning to how technology may be reshaping language, it is helpful to briefly consider how languages evolve naturally. Evidently, this is not meant to be a comprehensive or nuanced account of language evolution since doing justice to the topic of language evolution requires several volumes. 2 2 2 For readers interested in a broad, comprehensive, nuanced history on, among others, the evolution of language, I recommend “The Language Puzzle” Mithen 2024 and “How Language Began” Everett 2017 . Rather, the goal here is to highlight a few key aspects of how humans learn and transmit language, to lay the groundwork for drawing parallels with how Multilingual LLMs learn, transmit and might affect language. More broadly speaking, natural selection in languages operates at two main, related levels when discussing linguistic diversity: across languages (Section 2.1 ), where some languages survive and thrive while others gradually disappear, and within languages (Section 2.2 ), where certain words, expressions, or grammatical structures persist over time while others fall out of use. 2.1 Across Languages Focusing first on the cross-linguistic level, history shows that shifting populations, cultural dominance, and environmental factors have repeatedly driven the extinction of many languages Bromham et al. 2022 . Today, nearly half of the world’s approximately 6000 languages 3 3 3 This is a commonly used figure that many reference books use but estimates vary in recent decades between 3000 and 10000 Crystal 2002 . are endangered, with some disappearing at an alarming rate. 4 4 4 https://www.unesco.org/en/articles/multilingual-education-bet-preserve-indigenous-languages-and-justice . This natural process of ‘language death’ is not new; throughout history, languages have risen and fallen, often due to the expansion of dominant languages and the demographic shifts that accompany political and social changes Crystal 2002 . Such shifts have led to the gradual disappearance of many so-called ‘long-tailed languages’, languages with relatively fewer speakers or less economic visibility. The disappearance of individual languages has been widely recognized by scholars for decades Krauss 1992 ; Hale 1992 ; Simons and Lewis 2013 ; Crystal 2002 and international organizations such as UNESCO 5 5 5 https://www.unesco.org/en/articles/new-unesco-report-calls-multilingual-education-unlock-learning-and-inclusion have set up initiatives to promote language revitalization. It is evident, that across languages, also current and past cultural technologies have introduced additional pressure or accelerants which we will briefly touch upon in Section 3 . While language loss is often viewed as something tragic, it can stem from a shared communicative goal or need. For instance, speakers might adopt a dominant language to improve their chances of access to education; or simply to improve mutual understanding. Evidently, even in these seemingly positive scenarios, something linguistically and culturally valuable is lost. Perhaps more subtly and at the same time more universally , we might also be witnessing a different type of loss: the loss of richness and diversity within languages themselves. Such a loss is less likely to be driven by collective needs like mutual intelligibility 6 6 6 Of course, in some specific cases or contexts, simplification can help communication across a more diverse set of speakers. , but can most likely be attributed to converging pressures and constraints from language technologies. Therefore, I want to turn the attention more inwards, and focus on the erosion of diversity within languages themselves. 2.2 Within Languages Whereas the extinction of entire languages is often driven by external forces, changes within a language are also shaped by internal cognitive bottlenecks. The acquisition of a language, for instance, happens under rather constrained conditions: humans are exposed to sparse, ambiguous, and often noisy and imperfect input. In addition to this so-called poverty of stimulus Chomsky 1980 , we face cognitive