The LLMs quietly homogenizing the world’s surviving oral histories

The Mirage of Digital Taxidermy

For generations, the oral histories of the Shor people in the Kuznetsk Alatau mountains of Siberia survived solely through the breath of the *Chylgyschy*—epic singers who memorized tens of thousands of verses. These stories were never static; they mutated with every season, adapted to the mood of the community, and absorbed the local changes of the landscape. Today, digital humanists rush to record, transcribe, and feed these endangered epics into large language models under the banner of preservation.

This rush reveals a profound misunderstanding of how oral traditions function. Linguist Dr. Mark Turin, director of the World Oral Literature Project, has long argued that transcription is a transformative act that can sever a story from its living context. When we feed these transcriptions into neural networks, we do not preserve them; we subject them to what is best understood as digital taxidermy.

• Static Freezing: LLMs require static, textual inputs, freezing a fluid, performative tradition into a single, authoritative digital artifact.
• Contextual Decapitation: The physical environment, the audience's real-time feedback, and the performer's vocal pitch are discarded in favor of flat text.
• Semantic Domestication: The raw, localized nuances of the story are translated and formatted to match the dominant patterns of global training data.

Current evidence suggests that when a non-written narrative is digitized and ingested by a neural network, it undergoes a permanent structural decay. We can define this phenomenon as the folkloric ground-state: the systematic collapse of highly localized, anomalous narrative mutations into the standardized, statistical averages of high-dimensional neural spaces. Instead of safeguarding diversity, digitized archives act as a centrifuge, spinning away the local variations that gave the myth its original life.

To witness this process, one can try a zero-cost test: ask any commercial LLM to retell a highly localized regional myth, such as the Siberian tale of the iron-clad monster *Otcshir-Bogan*. The model will almost certainly restructure the story, unconsciously filling gaps with classical European narrative arcs like the Hero's Journey, stripping the tale of its unique, non-linear logic.

The Tokenizer and the Silent Erasure of Shibboleths

The silent homogenization of oral histories begins long before a model undergoes training; it is baked into the very math of tokenization. Large language models do not read words; they process tokens using algorithms like Byte-Pair Encoding (BPE) or SentencePiece. These tokenizers are optimized for high-resource, written languages like English, Spanish, and Mandarin, where common words are assigned clean, single-token representations.

When processing low-resource, oral-first languages—such as Gikuyu, Yolngu Matha, or modern Aramaic—the tokenizer struggles. Research published by the Masakhane NLP consortium demonstrates that African languages are routinely fractured into highly fragmented, meaningless sub-word units. This structural mismatch introduces a profound linguistic distortion that compromises the integrity of the translated text.

"The tokenizer acts as a coarse sieve, crushing the delicate phonemic architecture of oral-first languages into jagged, arbitrary fragments."

This fragmentation has severe downstream consequences for how oral myths are represented inside the model. Because the tokenizer cannot cleanly parse the rhythmic structures, tonal markers, or onomatopoeic patterns essential to oral performance, it forces these elements into standard written syntax.

1. Cadence Deletion: The rhythmic pauses and repetitions that serve as mnemonic devices in oral epics are discarded as redundant noise.
2. Phonetic Flattening: Tonal shifts that change the meaning of a word in languages like Yoruba are flattened into homonyms, stripping the narrative of its dual meanings.
3. Lexical Substitution: The model replaces highly specific local ecological terms with broad, generic equivalents from its dominant training corpus.

One compelling interpretation holds that this sub-word fragmentation acts as a tax on cultural fidelity. By forcing oral structures through a tokenizer optimized for written, bureaucratic prose, we perform a silent, automated translation that strips a story of its unique phonetic signature before the neural network even begins to compute its meaning.

The Oral Siltation Effect in Vector Spaces

Once ingested, these stories are mapped onto high-dimensional vector spaces, where semantic relationships are determined by mathematical proximity. In these dense geometric environments, another silent process occurs, which we may call the oral siltation effect. This is the mathematical pull that dense, high-resource semantic clusters exert on sparse, outlying narrative structures, gradually dragging them toward the conceptual center of gravity.

Consider the Australian Aboriginal concept of *Alcheringa*, commonly translated as "The Dreamtime." Anthropologist W.E.H. Stanner famously argued that this term does not refer to a past golden age, but rather to a continuous, non-linear temporal dimension where past, present, and future coexist. It is a concept that fundamentally defies Western, linear historical models.

Yet, within the vector space of a commercial embedding model, the tokens representing *Alcheringa* are inevitably positioned close to Western concepts like "mythology," "legend," "dream," and "creation story." When a user queries an LLM about these traditions, the model performs a vector search and pulls the response toward this dense semantic cluster.

• Temporal Alignment: The model reconstructs non-linear, circular cosmologies into linear timelines that Western minds can easily digest.
• Theological Translation: Ancestor spirits that represent local ecological features are mapped onto Western categories of gods, demons, or fairies.
• Logic Homogenization: Paradoxes and non-binary logic systems inherent in local mythologies are resolved to satisfy the boolean logic of neural network processing.

While mainstream NLP researchers often celebrate vector alignment for its ability to bridge linguistic divides, this process comes with a steep trade-off. It forces entirely different worldviews into a shared semantic taxonomy. The unique conceptual topography of an oral culture is flattened, leaving only a standardized, Westernized approximation of its original meaning.

RLHF and the Victorian Sanitization of Myth

Traditional folklore is rarely polite. The raw oral histories of humanity are filled with visceral violence, explicit sexuality, scatological humor, and profound moral ambiguity. These stories were not designed to be safe; they were designed to help communities navigate the terrifying realities of nature, disease, and social conflict.

However, modern commercial LLMs are governed by Reinforcement Learning from Human Feedback (RLHF) and strict safety classifiers. These safety layers, designed by corporate teams in San Francisco or Seattle, are tuned to prevent the generation of harmful, offensive, or sexually explicit content. When these filters are applied to raw mythology, the result is a systemic sanitization of human heritage.

Consider the West African trickster tales of Anansi the Spider, or the raw, visceral creation myths of the Māori, such as Maui's attempt to win immortality for humanity by entering the body of Hinenuitepō. These stories contain themes that modern safety filters flags as violent, graphic, or inappropriate.

When asked to generate or analyze these myths, aligned models routinely alter the narratives to fit modern corporate standards:

• Moralizing the Trickster: The amoral, chaotic trickster is rewritten as a benign teacher who learns a clear, prosocial lesson at the end of the story.
• Sanitizing the Flesh: Physical, biological elements of creation myths—such as blood, semen, and decay—are systematically replaced with abstract, poetic metaphors.
• Simplifying Conflict: Complex tribal relationships and blood feuds are reframed through a modern lens of simple, binary good-versus-evil dynamics.

While this sanitization is necessary to make AI systems safe for general corporate use, its application to cultural archives is highly destructive. By scrubbing the transgressive and the uncomfortable from our myths, RLHF acts as an automated censor, transforming wild, ecologically raw folklore into safe, sanitized, and ultimately boring fables.

The Death of Mutative Memory

For millennia, the survival of oral history depended on its ability to change. In his classic study The Singer of Tales, Albert Lord demonstrated that Yugoslavian *guslar* epic singers did not memorize static scripts. Instead, they relied on a flexible toolkit of formulas, themes, and rhythmic patterns to reconstruct the story anew during every performance.

Variation was not an error; it was a feature that kept the story alive and relevant to its listeners. If a village experienced a sudden drought, the singer might weave a warning about water conservation into an ancient myth. If a new chief took power, the lineage of the mythical hero might subtly shift to legitimize the new order.

Large language models halt this natural mutation. When an oral history is digitized, trained upon, and deployed, the model locks a single, arbitrary version of the story into its weights. This creates a digital monoculture where variation is treated as a hallucination to be suppressed.

"When an algorithm freezes an oral tradition into a single, optimized output, it halts the evolutionary process that allowed the myth to survive for thousands of years."

This freezing of myth introduces a profound second-order risk. When a community begins to rely on an LLM to retrieve and preserve its own folklore, the fluid, multi-voiced history is replaced by a single, authoritative digital oracle. The biological capacity for communal forgetting and adaptation is lost, replaced by a permanent, unyielding digital record.

The Ontological Alignment Paradox

To integrate diverse cultural narratives into global databases, researchers often rely on cross-lingual alignment techniques, using models like mBERT or XLM-RoBERTa. These systems align different languages within a shared semantic space, allowing a concept in Quechua to be mapped directly to a concept in English.

This process introduces a structural challenge that we can term the ontological alignment paradox. This paradox states that the more closely you align a low-resource language with a dominant global language for the sake of accessibility, the more you erase the unique ontological assumptions of the original tongue.

Consider the Andean Quechua concept of *Pacha*. In standard translations, *Pacha* is mapped to "world," "earth," "time," or "space." However, for traditional Quechua speakers, *Pacha* is not a static container where things exist; it is an active, living, and reciprocal cosmic force that binds humans, nature, and ancestors together.

When a cross-lingual model aligns Quechua with English, it resolves this complexity through mathematical optimization:

1. Ontological Reduction: The model collapses the relational, living concept of *Pacha* into the Western, mechanistic concept of "space-time."
2. Epistemic Extraction: The ritual obligations of reciprocity (*ayni*) associated with *Pacha* are translated as mere "traditions" or "superstitions," rather than recognized as an active ethical framework.
3. Semantic Domestication: The language is stripped of its power to challenge the dominant Western ontology, serving instead as a colorful dialect used to express Western ideas.

This alignment does not preserve the worldview of the Quechua people; it conscripts their vocabulary to populate a Western-centric model of reality. The alternative perspective—that languages represent entirely different ways of being, rather than different words for the same things—is systematically erased by the mathematical demands of cross-lingual vector alignment.

The Tragedy of the Synthetic Echo Chamber

As the internet is flooded with AI-generated text, we are entering an era of synthetic feedback loops. Large language models are increasingly trained on data scraped from the web that was itself generated by other LLMs. This recursive training process leads to a well-documented phenomenon known as model collapse, where the tail distributions of a dataset are progressively lost over generations of training.

For oral histories, this synthetic feedback loop is a terminal threat. Because oral traditions are already rare on the open web, they exist in the extreme long-tail of the training distribution. When an AI scrapes the web, it finds only a few, highly homogenized digital versions of these myths, which it then uses to generate new content.

This process triggers a rapid, downward spiral of cultural diversity:

• The First Generation: An LLM scrapes a raw, diverse set of oral transcriptions and outputs a standardized, clean version of a myth.
• The Second Generation: A cultural preservation website publishes this clean version. Another AI scrapes this site, further smoothing out any remaining local quirks or anomalies.
• The Third Generation: The model, trained on its own sterile outputs, can no longer generate anything other than a generic, Disneyfied archetype of the original story.

Preliminary research suggests that this recursive decay occurs far faster than standard linguistic drift. In a matter of years, the rich, chaotic tapestry of human oral tradition could be reduced to a handful of sterile, synthetic archetypes, endlessly recycled by models trained on their own homogenized outputs.

Building the Analog Vault: A Strategy for Cultural Sovereignty

To prevent the complete homogenization of the world's surviving oral histories, we must abandon the naive belief that digital ingestion equals preservation. True preservation requires cultural sovereignty, which can only be achieved by creating offline, un-crawlable spaces where oral histories are kept in their native acoustic and relational contexts.

Instead of feeding endangered myths into commercial, cloud-based LLMs, indigenous communities and cultural preservationists should adopt a strategy of intentional fragmentation. This approach can be realized through the deployment of decentralized, local-first archiving platforms like the Mukurtu CMS project, which allows communities to set their own access protocols and keep their archives offline.

To implement this model, communities and researchers can build a low-cost, secure, and sovereign oral archive using the following framework:

1. Air-Gapped Acoustic Vaults: Store raw audio recordings of oral histories on physical, encrypted drives housed within the community, completely disconnected from the public internet. This prevents global scraping spiders from digesting the data into commercial models.
2. The robots.txt Barrier: If metadata or stories must be shared online, use strict server configurations and custom headers to block all AI crawlers (such as GPTBot and ClaudeBot) from accessing the directory.
3. Local, Fine-Tuned Models: If AI must be used for language learning, deploy small, open-weight models (like LLaMA or Mistral) hosted locally on consumer hardware. These models should be fine-tuned *only* on the community's verified data, using custom-designed tokenizers that respect the phonetic architecture of the native language.

By shifting from global digital exposure to local, high-fidelity custody, we can protect the sacred boundaries of human memory. The goal of preservation should not be to make every story legible to a global algorithm, but to ensure that the whisper of the *Chylgyschy* can still be heard by the next generation, in the exact cadence it was meant to be told.