In modern machine learning, data is often treated as a natural resource: an unowned, raw material waiting to be cleared, processed, and fed into models. For dominant global languages, this pipeline operates largely unnoticed behind web scrapers and massive compute clusters. However, when applied to Indigenous and endangered languages, this paradigm breaks down.
A growing body of research warns that treating these languages merely as "data-scarce" optimization problems perpetuates a form of digital colonialism. By viewing language communities as passive data repositories rather than sovereign intellectual authorities, standard Natural Language Processing (NLP) practices risk replicating historical patterns of extraction and exploitation.
The Extraction Paradigm and the AI Gold Rush
To understand why standard NLP practices fail Indigenous communities, we must look at how data is acquired. In the race to build massive Large Language Models (LLMs), developers routinely scrape the open web for any available text.
As Marie-Odile Junker highlights in Data-mining and Extraction: the gold rush of AI on Indigenous Languages, this has triggered a digital "gold rush." Developers aggressively pursue linguistic data from vulnerable communities without adequate consideration for the creators or the cultural context of the data.
This approach treats language as a commodity. In Decolonising Speech and Language Technology, Steven Bird argues that framing Indigenous languages as "zero-resource" or "gold data" strips them of their social realities. This technical framing views languages as raw materials for extraction, ignoring the community's own knowledge systems, aspirations, and sovereignty.
When text is scraped indiscriminately, it violates basic principles of data ownership. For instance, Linda Wiechetek et al. demonstrate in The Ethical Question – Use of Indigenous Corpora for Large Language Models that web-scraping for LLM training routinely disregards community needs and data ownership, moving past simple legal compliance into deep ethical violations.
Extractive Paradigm (Digital Colonialism):
[Community Knowledge] ──(Scraping/Extraction)──> [Data Commodity] ──> [Model Benchmarking]
Sovereign Paradigm (Decolonized NLP):
[Community Sovereignty] ──(Participatory Design)──> [Contextual Technology] ──> [Language Revitalization]
The Deception of "Low-Resource" Terminology
The problem is baked into the very vocabulary computational linguists use. Terms like "low-resource" or "data scarcity" sound like neutral, objective descriptions of statistical distributions. However, they obscure why the data is scarce.
In Primum Non Nocere: Before working with Indigenous data, the ACL must confront ongoing colonialism, Lane Schwartz argues that categorizing Indigenous languages solely as low-resource data sets hides the historical and political factors—such as active, colonial suppression—that led to their current status. Treating data scarcity as a mere technical hurdle erases a history of systemic harm.
Furthermore, this terminology misdiagnoses the problem. In ”It’s how you do things that matters”: Attending to Process to Better Serve Indigenous Communities with Language Technologies, Ned Cooper et al. critique "low-resource" as a Eurocentric framing. They propose the term under-served instead. This shift in language clarifies that the issue is not a natural lack of data, but rather a systemic failure of the technology community to adequately support and service these language communities.
When researchers fail to interrogate these underlying assumptions, they risk doing active harm. Craig John Carpenter et al. warn in Seeding Alignment between Language Technology and Indigenous Methodologies: A Decolonizing Framework for Endangered Language Revitalization that standard computational linguistics terms carry inherent colonial biases that promote extractive research practices, potentially exacerbating language endangerment rather than alleviating it.
Structural Power Imbalances in Tech Development
Digital colonialism is not isolated to North American Indigenous contexts; it is a global structural issue. For example, in Decolonizing African NLP: A Survey on Power Dynamics and Data Colonialism in Tech Development, Nan Yan et al. document how historical colonial legacies continue to influence modern NLP development in Africa. This manifests as a systemic marginalization of African languages and local needs, resulting in biased, ineffective AI systems designed from the outside.
The core conflict lies in a misalignment of values. As Claudio S. Pinhanez et al. point out in Balancing Social Impact, Opportunities, and Ethical Constraints of Using AI in the Documentation and Vitalization of Indigenous Languages, current AI development practices rely on data extractivism and colonial thinking that are fundamentally at odds with the principles and values held by Indigenous communities.
When tech companies or academic researchers build tools solely for benchmarking or high-level metrics, they create utilities that do not serve the actual linguistic or educational priorities of the people speaking the language.
Shifting to Community-Centric Frameworks
To move past digital colonialism, the NLP community must shift from a model-centric focus to a process-centric and community-led focus.
| Dimension | Extractive NLP Paradigm | Decolonized NLP Paradigm |
|---|---|---|
| Data View | Raw commodity / Open-source benchmark | Sovereign intellectual property |
| Primary Goal | Maximizing model performance metrics | Supporting community language revitalization |
| Research Design | Top-down, researcher-led | Bottom-up, participatory, community-led |
| Language Framing | "Low-resource" or "Zero-resource" | Systemically "Under-served" |
To guide this transition, researchers have proposed several concrete frameworks:
- Ethical Pillars: Lane Schwartz outlines a four-pillar ethical framework—Cognizance, Beneficence, Accountability, and Non-maleficence—to fundamentally restructure how engineering communities engage with sovereign nations.
- Data Sovereignty: In Ethical Considerations for Machine Translation of Indigenous Languages: Giving a Voice to the Speakers, Manuel Mager et al. emphasize that tools like Machine Translation must prioritize the right of speakers to exercise data sovereignty. This includes the right of communities to decide what—and what not—to translate or digitize.
- Revitalization Roadmaps: Rather than treating languages as benchmarking commodities, researchers must design tools around specific educational and revitalization goals. For instance, Shiyue Zhang et al. present a collaborative roadmap for the Cherokee language in How can NLP Help Revitalize Endangered Languages? A Case Study and Roadmap for the Cherokee Language, demonstrating how NLP can ethically support revitalization when guided directly by community priorities.
Ultimately, decolonizing NLP requires recognizing that language data cannot be separated from the people who speak it. True progress in language technology is measured not by the perplexity scores of a model on a scraped dataset, but by how effectively the technology honors the sovereignty and serves the vitality of the community.
Key Foundational Work
Decolonising Speech and Language Technology (2020) — contributed 3 concept elements to this direction.
Primum Non Nocere: Before working with Indigenous data, the ACL must confront ongoing colonialism (2022) — contributed 1 concept elements to this direction.
How can NLP Help Revitalize Endangered Languages? A Case Study and Roadmap for the Cherokee Language (2022) — contributed 1 concept elements to this direction.