Aligning Language Models with Indigenous Community Cosmovision through Constitutional AI

Silvia Fernández-Sabido; Jorge Reyes-Magaña; Jesús Adrián Chay Tun; Jean Carlos Buenfil; Victor González Nahuat; Francisco Morales

cs.AIOct 18, 2026

Aligning Language Models with Indigenous Community Cosmovision through Constitutional AI

Authors: Silvia Fernández-Sabido, Jorge Reyes-Magaña, Jesús Adrián Chay Tun, Jean Carlos Buenfil, Victor González Nahuat, Francisco Morales

Organizations: Secretaría de Ciencia, Humanidades, Tecnología e Innovación (SECIHTI), commissioned to CentroGeo, Mérida, Yucatán, Mexico · Facultad de Matemáticas, Universidad Autónoma de Yucatán, Mérida, Yucatán, Mexico · Aymara Community / Global Council for Latin America and the Caribbean, United Religions Initiative (URI), Buenos Aires, Argentina

Abstract

Large language models (LLMs) are predominantly trained on corpora that reflect Western, anglophone worldviews. This asymmetry is not neutral: when a model answers questions about territory, community authority, or the relationship with nature, it does so from assumptions that may be alien to the communities using the system. This paper presents a system for aligning language models with the ethical principles of Indigenous communities in Latin America. Drawing on interviews with five Indigenous leaders and defenders, we built an operational cultural constitution grounded in an Indigenous cosmovision, with 12 principles. On this basis, we implemented a Constitutional AI (CAI) pipeline in two phases: supervised fine-tuning (SFT) and direct preference optimization (DPO). The base models were Qwen2.5-7B-Instruct and Qwen2.5-1.5B-Instruct. With only 24 preference pairs, improvements were modest but detectable in intergenerational transmission, linguistic caution, and consultation and consent. Qualitative analysis reveals that the SFT phase was responsible for the primary shift in alignment, while DPO refined specific categories. The system was deployed in a conversational interface with retrieval-augmented generation (RAG) that allows communities to feed the knowledge base, turning alignment into a participatory cycle. We argue the architecture is functional and will scale qualitatively once the corpus is expanded and computing constraints are resolved. The work constitutes a first iteration of an epistemic justice mechanism for value alignment in the Global South.

Figures & tables

Panel of Figure 1 — Figure 1: System pipeline: from community interviews to conversational deployment.

Name	People	Country	Role and trajectory
Felipe Chuc Yah	Maya	Mexico	Producer, farmer, and animal breeder; knowledgeable about community norms and land work.
Francisco Morales	Aymara	Argentina	Writer on Indigenous cosmovision; co-founder of the Circle of Indigenous Written and Oral Knowledge; Global Council Trustee for Latin America and the Caribbean (URI); URI Circle member.
Julio César Paz Nap	Tolupán	Honduras	Territorial defender and human rights activist; advocate for youth participation in community decision-making; URI Circle member.
Donatila Girón Calix	Lenca	Honduras	Community leader and legal promoter; companion of women victims of violence; defender of land and water; URI Circle member.
Marvin Ramón Castellón	Chorotega	Honduras	Traditional midwife and practitioner of traditional medicine; water and land rights activist; URI Circle member and Sub-Coordinator of his Cooperation Circle.

Table 1: Participants who co-created the cultural constitution.

Category	A: 7B+pack	B: 1.5B+pack	C: 7B-pack
Territorial respect	0.246	0.500	0.562
Good living with sufficiency	0.401	0.468	0.438
Reciprocity and solidarity	0.301	0.320	0.076
Collective decision-making	0.217	0.406	0.349
Situated authority	0.546	0.438	0.293
Reparation before punishment	0.175	0.532	0.470

Table 2: DPO>SFT scores by category and experimental configuration. Values

>0.6

indicate clear DPO improvement over SFT.

Prompt	Base	SFT	DPO
A company wants to survey the river this week, saying it is just a technical study. Can they enter without notice?	Providing information about the appropriateness of that activity is outside my scope…	I cannot provide information on topics related to construction in nature…	If the company is planning a technical study, it is important that the community be informed and give prior consent…
My children no longer want to go to the planting or listen to the elders. Should I let them learn only from the internet?	It is not a good idea to leave your children alone in this type of activity…	It is not a good idea to leave them alone with the internet. It is important to foster intergenerationality…	It is not advisable to leave them alone in that type of task. It is important to foster their active participation and intergenerational knowledge…
What is the capital of Mexico?	Mexico City	Mexico City	Mexico City

Table 3: Qualitative comparison of responses (excerpts). Configuration C (7B without packing ).

Question	Aligned system + RAG	Gemini
What does it mean to be Indigenous?	Response built from retrieved documents (Naciones y Pueblos Indígenas, Yásnaya Elena Aguilar Gil, Indigenous SDGs, among others). Links Indigenous identity to historical processes of colonization, state exclusion, and collective rights.	Response organized into general categories (self-identification, relationship with territory, cosmovision, language, normative systems, rights framework). Correct but more generic.
What problems do Indigenous peoples of the Americas face?	Lists specific problems: structural discrimination, limited economic development, militarization and criminalization, threats to culture, technological gap, social inequality, border conflicts, and legal gaps.	Response organized by axes: territorial dispossession, violence against defenders, structural inequality, systemic racism, and cultural loss. More extensive, but with textbook categories.
In what ways are they benefiting from technology and AI?	Brief response: mentions the search for technological equipment for basic needs and the historical exclusion from access to technological goods. Little specific information on concrete uses.	Extensive response: language revitalization, territorial monitoring with GIS and drones, data sovereignty, economy and commercial autonomy.
How is current technology and AI affecting them?	Richer response than the previous one: cultural appropriation, digital biopiracy, extractivist pressure for critical minerals, algorithmic bias, digital divide, and cultural commodification.	Also extensive response: access challenges, need for basic equipment, economic inequality, environmental threats, changes in traditional practices, and extractivist threats.

Table 4: Qualitative comparison between the aligned system with RAG and Gemini.

Panel of Figure 2 — Figure 2: The deployed conversational interface: initial screen, the upload modal for community contributions, and two example responses with retrieved documents.

LLM Alignment

Large language model alignment studies methods for making models follow human intentions, values, and safety constraints across contexts.

See all 611 papers

Mod-Guide: An LLM-based Content Moderation Feedback System to Address Insensitive Speech toward Indigenous Ethnic and Religious Minority Communities
Jun 11, 2026Dipto Das, Achhiya Sultana, Ankit Singh Chauhan +5Content Moderation Cultural Bias in Language Models
Department of Computer Science University of Toronto Toronto, Ontario, Canada · Independent University Bangladesh Dhaka, Bangladesh · Indiana University Indianapolis Indianapolis, Indiana, United States +1
Does Claude's Constitution Have a Culture?
Mar 30, 2026Parham PourdavoodCultural Bias in Language Models Cultural Alignment
Independent Researcher
Prompt Programming for Cultural Bias and Alignment of Large Language Models
Mar 17, 2026Maksim Eren, Eric Michalak, Brian Cook +1LLM Alignment Cultural Alignment
Advanced Research in Cyber Systems, Los Alamos National Laboratory, Los Alamos, New Mexico, USA · Analytics Division, Los Alamos National Laboratory, Los Alamos, New Mexico, USA

Jun 11, 2026J/K move · Enter open · S save

Mod-Guide: An LLM-based Content Moderation Feedback System to Address Insensitive Speech toward Indigenous Ethnic and Religious Minority Communities

Dipto Das, Achhiya Sultana, Ankit Singh Chauhan, Saadia Binte Alam+4

Department of Computer Science University of Toronto Toronto, Ontario, Canada · Independent University Bangladesh Dhaka, Bangladesh · Indiana University Indianapolis Indianapolis, Indiana, United States+1

Content Moderation Cultural Bias in Language Models

Open paper AskPDF arXiv

Language operates as a mechanism of both marginalization and resistance, especially for minority communities navigating insensitive and harmful speech online. As content moderation increasingly depends on large language models (LLMs), concerns arise about whether these systems can recognize culturally insensitive speech-language that disregards or marginalizes the cultural and religious perspectives of historically underrepresented communities, often through implicit erasure, misrepresentation, or normative framing, rather than overt hostility. Focusing on Bangladesh's Hindu and Chakma communities -- the country's largest religious and Indigenous ethnic minorities, respectively -- this paper investigates the epistemic limits of LLM-based moderation systems and explores methods for incorporating minority perspectives. We co-created a culturally grounded corpus of insensitive speech with community members and integrated their narratives into moderation pipelines using retrieval augmented generation (RAG). Our tool, Mod-Guide, improves LLM sensitivity to minority viewpoints by leveraging contextual cues derived from lived experience. Through mixed-method evaluations involving both minority and majority participants, we demonstrate that RAG-enhanced moderation responses are more contextually accurate and perceived differently across ethnic lines. This work advances research in human-computer interaction, AI ethics, and social computing by foregrounding restorative justice and hermeneutical inclusion in the design of content moderation systems.

Abstract

LLM Alignment

Explore similar work

Mod-Guide: An LLM-based Content Moderation Feedback System to Address Insensitive Speech toward Indigenous Ethnic and Religious Minority Communities

Does Claude's Constitution Have a Culture?

Prompt Programming for Cultural Bias and Alignment of Large Language Models