In 2018, some of us observed that the problem with AI isn’t ethics; it’s epistemology. A major Phase 1 finding is a corollary: ‘epistemology is architecture’. The Niitsitapi Pod developed a five-category Blackfoot knowledge-domain framework for evaluating what knowledge can be shared into AI systems; the Haudenosaunee Pod created the Digital Wampum Treaty, a living, relational framework where data is knowledge bound within ongoing responsibilities to land, community, and future generations, establishing a treaty-based model for accountable computational systems; the T’Karonto Pod drew upon the Dish with One Spoon Treaty to design a system of reciprocal sharing and layered disclosure protocols to enforce sovereign control in AI systems; the Wíhaŋble S’a Pod uses wíhaŋble (dreaming) as a sovereign mode of Lakȟóta knowledge-making that any computational system must accommodate; Hiringa te Mahara produced a Kaupapa Māori-based protocol to guide any work with AI; and the Hawai‘i Pod looked to Kapu and Kanawai to incorporate AI into their cultural context. These efforts converge along several dimensions: 1) Sovereignty not only over data, but over the conditions that make computation possible—how models are built, how they learn, and how they operate; 2) Governance as a fundamental component of any AI system; and 3) Projects that integrate land, material practice, community knowledge, and computation into new configurations.

Our community collaborators are increasingly concerned about the environmental impact of AI technologies. Spurred by reporting on the enormous resource demands required to both train LLMs and use them, they have become increasingly skeptical as to whether engaging with AI at all is ethically acceptable for communities that define themselves through the territories which they steward.

This led to numerous spirited discussions within the Pods, across the Partnership, and in public fora involving Indigenous participants. We learned that understanding, addressing, and ameliorating the environmental impact of AI systems is as fundamental to successful Indigenous AI as solving how to build sovereignty into every aspect.

The Abundant Soils (Crow and Lewis) project at the Hawai‘i Pod leads the way here. This project explores how to integrate pedology (understanding of the physical, chemical, and biological soil properties, patterns, and their genesis), quantitative models of soils, and traditional Hawaiian soil nomenclature and knowledge practices, with large-scale landscape models that draw on machine learning techniques to create holistic models for understanding and increasing soil health.

The health and survival of their languages, and the storytelling-as-knowledge-practice built on top of them, are top priorities across all the collaborating communities. Thus Natural Language Processing (NLP) is the network’s most concentrated technical cluster, and it illustrates how Indigenous design constraints produce technically novel research. Building language models for languages that are low density (Te Reo Māori) and/or polysynthetic (Kanien’kéha, Niitsi’powahsin, ʻŌlelo Hawaiʻi) poses challenges that commercial architectures are not well-designed to address: tokenisation, morphological parsing, training-data scarcity, and the need for community validation rather than benchmark performance as the measure of model quality. Across the network, small language models are emerging as a shared technical direction, emphasizing community-appropriate, locally controllable systems per the governance architectures. The network has assembled substantial language corpora (e.g., over 2,800 hours of Blackfoot waveform speech data) and produced early evidence that commercial language models cannot adequately hold the cultural and logical structures of Indigenous languages, a finding with implications for low-resource NLP broadly. An example here is the Lauleo Project (Mahelona) which explores innovative methods for collecting community-verified data to train AI tools for ōlelo Hawai‘i speech recognition and language preservation.

This area explores how AI systems can move beyond isolated optimization toward relational, context-aware intelligence, drawing on both neuroscience and Indigenous knowledge systems. It reflects an innovative shift toward socially grounded AI, where interaction, reciprocity, and meaning-making are central computational principles. Mācistan: Reciprocity in Multi-Agent Reinforcement Learning as a Credit Assignment Problem (Richards and Master student Malenfant) examines how cooperation can emerge among AI agents without explicit social contracts, inspired by Métis resource-sharing practices of the North American plains. In Yuíyeskapi (Translating) (Kite and Jerbi), researchers developed a system in which dream-derived inputs are translated into symbolic representations embedded in a shared latent space, enabling context-sensitive semantic mapping and generative AI systems for real-time performance outputs.

The Partnership is advancing embodied, multimodal AI systems that move beyond text-centric approaches by treating non-verbal, environmental, and collective signals as primary inputs. Projects include ML on thermal and 3D land-based data (Two Bears), community-authenticated access systems (T’Karonto collaborator McConnell), and EEG neurofeedback with semantic mapping (Kite and Jerbi), reflecting a shift toward edge computing and embodied AI. This approach supports localized, community-owned infrastructures, reduces reliance on centralized models, and enables integration of cultural, environmental, and experiential data.