The Good Mind Guide to Artificial Intelligence

A Haudenosaunee Playbook for Indigenous Organizations

How to Use This Playbook

This playbook is a public resource for Indigenous organizations: First Nations administrations, community programs, non-profits, and cultural institutions who are weighing whether or not and how to use artificial intelligence in their work.

It is written differently than most AI guides. Rather than starting from a technology and asking where it fits, this playbook starts from Haudenosaunee principles: the Good Mind, Seven Generations, the Two Row Wampum, and the Great Law of Peace.

This includes information on the tradeoffs of using mainstream tech platforms. These systems can offer real benefits, but they can also quietly cost us sovereignty over our own data, knowledge, and decisions. This guide was written to help weigh that tradeoff deliberately, one decision at a time, rather than following AI adoption strategies developed by organizations that may not be aligned with our values.

A note on identity: this guide was developed from an Indigenous perspective, but does not include all Indigenous perspectives. Indigenous is a broad umbrella term that includes nations from all over Canada. Although many of the teachings and ways of life are more similar than not, this guide is not meant to speak for all Indigenous people.

A note on transparency: this guide was developed in collaboration with Anthropic’s Claude, but the final draft was edited by a human and peer-reviewed by community members.

1. Words That Come Before All Else

Before any meeting, traditionally the Haudenosaunee people begin with the Ohen:ton Karihwatéhkwen (the Thanksgiving Address). This Address serves as a reminder of the interconnectedness of all life and the importance of living in harmony with nature. This playbook follows that same order, we start with gratitude. Let's take a moment to practice gratitude for our communities, our lands, and all of the natural world. I'll extend this to include gratitude for the technology available to us today that has allowed us to connect with each other frequently and instantaneously.

Artificial intelligence did not arrive from within Indigenous communities, and most of it was not built with Indigenous governance, consent, or knowledge systems in mind. That's not necessarily a reason to reject it outright, Indigenous communities have always evaluated new tools on their own terms, adopting what serves our people and setting aside what does not. It is a reason to slow down and evaluate deliberately, in the Good Mind, rather than immediately adopting whatever a vendor or tech company hands us.

This is not a technology roadmap, but a decision-making tool grounded in principles that predate the technology by centuries.

2. Foundational Principles

Four principles anchor every recommendation in this guide. They are not metaphors laid over a Western AI framework, they are the framework.

Kanikonriio — The Good Mind

The Good Mind is a state of thinking focused on peace, kindness, reason, and clear awareness of one's own thoughts and intent. Applied to technology, it means an organization does not adopt a tool simply because it is available, fast, or free. It asks first: does this serve our people well? Does it come from a clear and honest intent? Does it support peace within our community rather than erode trust and increase dependency?

Practically, this means slowing down procurement and adoption decisions enough to actually deliberate  with staff, with leadership, and with community, rather than reacting to hype or urgency manufactured by vendors who do not share our values.

Seven Generations Thinking

Decisions should be evaluated for their impact seven generations forward. For AI specifically, this means weighing costs that don't show up on a monthly invoice: the environmental footprint of the data centers powering these tools, the permanence of any data or cultural knowledge once it is uploaded to a third-party system, and the long-term effect on language, skills, and self-determination if a community becomes dependent on tools it does not control.

A short-term efficiency gain that increases long-term dependency on external, non-Indigenous infrastructure is not a gain in Seven Generations terms.

A specific note on image and video generation: these are, by a wide margin, the most energy-intensive everyday AI tasks. Independent measurements of common AI models have found that generating a single AI image can use on the order of a thousand times more energy than a typical text response, and that a short AI-generated video clip can use several hundred times more energy again than a single image. If your organization is weighing AI image or video tools — for promotional materials, social content, program graphics — that environmental cost is real and disproportionate compared to text-based drafting. Default to text-based AI use, and treat image or video generation as a deliberate, occasional choice rather than a routine one.

Guswenta — The Two Row Wampum

The Two Row Wampum belt depicts two vessels traveling the same river, side by side, neither steering the other. It is a founding principle of nation-to-nation relationship: parallel sovereignty.

Applied to AI vendors and technology consultants, Guswenta is a useful test: does this relationship keep your organization's governance, data, and decision-making in your own vessel? Or does adopting this tool quietly hand the power and control to someone else? A tech company's terms of service, a funder's reporting platform, a government department's preferred software? A tool can be useful and still need to be kept in its own lane, governed by our own protocols rather than absorbed into someone else's system.

Kaianere’kó:wa — The Great Law of Peace and Consensus Governance

Haudenosaunee governance traditionally operates by consensus. Decisions are worked through until the people affected can live with them. AI tools that summarize, draft, or accelerate decision-making are useful under this framework if they are supporting that process. A tool that produces a fast recommendation and short-circuits deliberation, or that makes it easier for a small group to bypass community input, works against this principle regardless of how efficient it is.

Where AI is used to support governance or leadership work such as drafting resolutions, summarizing consultations, or preparing materials for Council, it should be used to prepare the ground for consensus, not to replace it.

The Good Mind Checklist

Before adopting any AI tool or workflow, an organization can ask:

• Good Mind: Does this use of AI reflect good intentions and contribute to harmony, respect, and unity in our community?

• Seven Generations: What does this cost us (environmentally, culturally, in dependency) that won't show up on an invoice?

• Guswenta: Does our organization keep its own governance and data in its own control, or does adopting this tool hand control to someone else?

• Great Law: Does this tool prepare the ground for consensus and community input, or does it shortcut that process?

3. What is AI?

Artificial Intelligence (AI) is the broadest term — any system designed to do something that normally requires human thinking: recognizing a pattern, making a prediction, holding a conversation. It's the whole category, going back decades, and includes plenty of things that aren't chatbots at all (spam filters, fraud detection, GPS routing).

Generative AI is a subset of AI — specifically, AI that creates new content (text, images, video, audio) rather than just classifying, predicting, or retrieving existing information. This is the newer wave that's driven most of the recent AI conversation.

LLM (Large Language Model) is a subset of generative AI, the kind trained on huge amounts of text to generate language. Claude and ChatGPT are LLMs.

Types of AI used in our technology today:

Predictive analytics / machine learning (forecasts trends from historical data, doesn't generate new content)

• Everyday version: Netflix/Spotify recommendations, retailers predicting inventory needs, weather forecasting models

Computer vision (interprets images/video)

• Facial recognition, medical imaging analysis (reading X-rays or scans), license plate readers, quality-control cameras on a manufacturing line, OCR detection

Natural language processing (NLP), non-generative (understands language without generating new text)

• Spam filters, sentiment analysis on customer feedback, Google Translate, voice-to-text dictation

Voice assistants / speech recognition

• Siri, Alexa, Google Assistant — a mix of speech recognition plus a smaller decision-making system underneath

Recommendation and personalization systems

• Amazon/YouTube/TikTok's "you might also like" engines — this is one of the most widespread AI types in daily life, even though people rarely think of it as "AI"

Optimization and routing

• Google Maps/Waze traffic routing, supply chain and logistics optimization, scheduling software

Fraud and anomaly detection

• Your bank flagging an unusual charge, cybersecurity systems flagging suspicious login patterns

You can see how AI algorithms are already integrated into our everyday life. Generative AI is what most people are concerned about today. This guide is focused on the responsible use of Generative AI models and more advanced agentic AI systems that are built on today’s most popular LLMs.

4. Navigating the Colonial System

Indigenous organizations do not get to build AI policy in a vacuum. Most still have to file with provincial and federal funders, comply with Canadian privacy law, and report through platforms and formats they did not choose. This section is about operating inside those constraints without losing the principles above.

Whose Values Are Already Built In

Every large AI model is trained and fine-tuned according to the values of the company that built it, written into content policies, safety guidelines, and design choices made by teams that, in almost every case, did not include Indigenous people. This doesn't make these tools unusable, but it means their defaults: what counts as “neutral,” whose history is treated as settled fact, whose knowledge systems are treated as legitimate, are not neutral. Organizations should expect to correct, contextualize, and sometimes simply disregard AI output on matters of Indigenous history, law, and knowledge.

OCAP® and Data Sovereignty

OCAP (Ownership, Control, Access, and Possession) is the standard most widely used across First Nations in Canada for asserting authority over community data and research. It holds that a community owns its collective information, controls how that information is collected and used, has the right to access it, and controls where it is physically or digitally kept.

OCAP is important to keep in mind for any AI tool: Where does information actually go once submitted. Does the vendor retain it, train on it, or share it with third parties? Who can access it, and under what jurisdiction? Can your organization get it back, delete it, or move it if the vendor relationship ends? If a tool cannot give clear answers to these questions, it does not meet the standard, regardless of how useful it seems.

Compliance Realities

Canadian privacy law (PIPEDA, and provincial equivalents where applicable) still applies to any personal information an organization handles, AI-assisted or not. Practical steps:

• Maintain a clear data classification policy: what may go into a public AI tool, what may never leave internal systems (band membership data, health information, ceremonial or restricted cultural knowledge, HR files).

• Review AI vendor terms specifically for data retention and training use before adoption. Many default to using submitted data to train future models unless an enterprise or opt-out setting is enabled.

• Disclose AI use where funders or government partners require it, and keep a written record of which tools were used for which deliverables.

• Anticipate that funder and government reporting platforms may themselves use AI on submitted materials. Ask directly rather than assuming.

Strategic Options for Reducing Dependency

Organizations with the capacity to do so have several ways to reduce reliance on mainstream, US-based, externally-controlled AI systems:

• Open-source models that can be run on infrastructure your organization or a trusted Indigenous partner controls, keeping data off third-party servers entirely.

• Fine-tuning an open-source model on your own materials and values, so its defaults reflect your organization's language and governance rather than an outside tech company's.

• Indigenous-led, on-premise infrastructure hosting, for organizations that need or want full control over where data physically lives.

• Where none of the above is feasible, using mainstream tools cautiously for low-sensitivity drafting and admin work only, while keeping anything sensitive out entirely.

None of these require rejecting AI entirely. They are about choosing the level of dependency your organization is actually comfortable with, and being honest about the trade-off when a free or convenient tool is chosen instead.

5. Where AI Could Help Us

AI should be considered a powerful tool, not a decision-maker. We shouldn’t just be asking “Can AI do my job?” but “Would using AI serve our people while keeping our governance, knowledge, and data in our own hands?” 

The four areas below are practical places to explore generative AI. Start with low-risk uses with human oversight, keep people accountable for decisions, and increase the level of data governance as the sensitivity of the information increases.

5.1 Administrative Work

AI can reduce repetitive administrative work such as summarizing meetings, drafting routine correspondence, creating templates, searching approved documents, and preparing reports.

Tools to explore: ChatGPT (OpenAI), Claude (Anthropic), and Microsoft 365 Copilot (Microsoft).

Other options: Microsoft Word/Excel/SharePoint without AI, Google Workspace, traditional document-management systems, or Indigenous-controlled AI.

Example prompt:

“Summarize these meeting notes into decisions, action items, responsible people, and deadlines. Do not infer anything that isn't stated.”

More advanced use:
An organization could create a RAG AI knowledge assistant that searches approved internal policies, procedures, meeting records, and administrative documents. Staff could ask questions such as, “What is our current process for approving a new program?” and receive an answer based on the organization's own approved documents. 

Tradeoff:
The more useful the AI becomes, the more organizational information it may need access to. A system connected to internal documents creates greater risks around privacy, data retention, access permissions, and dependency on the vendor. A locally hosted or Indigenous-controlled system provides greater control, but requires more technical capacity, infrastructure, and ongoing maintenance.

Data sovereignty: Best practice: do not put HR files, Council discussions, community records, legal matters, or restricted cultural information into a public AI tool. Enterprise privacy protections reduce some risks but do not automatically give an organization ownership or control of its data.

5.2 Grants, Funding & Reporting

AI can help staff interpret funding criteria, identify gaps in proposals, develop workplans, draft reports, and organize information for recurring funder requirements.

Tools to explore: ChatGPT (OpenAI), Claude (Anthropic), Microsoft 365 Copilot (Microsoft).

Other options: Word and Excel without AI, grant-management software, shared reporting templates, human grant writers, or locally hosted AI.

Example prompt:

“Compare our approved project description against these funding criteria. Identify what is addressed, what is missing, and what needs clarification. Do not invent information.”

More advanced use:
AI could become a funding intelligence system that monitors approved funding opportunities, compares them against the Nation's programs and strategic priorities, identifies potential matches, tracks deadlines, and helps prepare the required application and reporting materials.

Tradeoff:
This could significantly reduce administrative workload and help organizations avoid missed funding opportunities. However, the system may require access to strategic plans, program information, budgets, and previous applications. The organization would need to decide whether the efficiency gained is worth giving an external AI provider access to that information. There is also a risk that AI-generated applications become overly standardized or lose the organization's own voice and priorities.

Data sovereignty: Grant applications can contain community needs assessments, health information, financial information, and strategic plans. Review what is being uploaded and where it will be stored. Do we want to trade long-term control of community knowledge for a faster grant application?

5.3 Housing & Homelessness

AI can help identify trends in housing demand, wait times, housing supply, program utilization, and community-level housing needs.

Tools to explore: Microsoft Power BI with Copilot (Microsoft), ChatGPT (OpenAI), and Gemini for Workspace (Google).

Other options: Excel, Power BI without Copilot, Google Sheets/Looker Studio, existing housing databases, or Indigenous-controlled analytics systems.

Example prompt:

“Analyze this aggregated housing dataset and identify significant trends in demand, wait times, housing types, and available units. Distinguish observed patterns from possible explanations.”

More advanced use:
AI could be used to model future housing needs by combining historical housing data with population trends, housing availability, program utilization, and other approved community-level information. The system could help explore scenarios such as: “What housing capacity might we need over the next five years under different population and construction assumptions?”

Tradeoff:
Advanced modelling could give leadership a much better understanding of future needs and help with long-term planning. However, the more detailed the data becomes, the greater the risk of identifying individuals or families. Predictive systems can also create a false sense of certainty: a forecast is not a fact, and AI should not be used to determine which individuals or families are considered "at risk." The tradeoff is therefore between better planning information and greater privacy, governance, and interpretation risks.

Data sovereignty: Housing data can reveal highly sensitive information about individuals and families. Use aggregated or appropriately de-identified information. Best practice: never put names, addresses, case notes, or family circumstances into a public AI system. 

5.4 Health & Wellbeing Metrics

AI can help organizations analyze community-level wellbeing data, identify trends, summarize anonymized survey responses, and build dashboards around indicators defined by the community.

Tools to explore: Microsoft Power BI with Copilot (Microsoft), ChatGPT (OpenAI), and Claude (Anthropic).

Other options: Excel, traditional statistical software, survey platforms, community-developed dashboards, or Indigenous-controlled analytics.

Example Prompt:

“Analyze these aggregated wellbeing metrics and identify significant changes over time. Present these as observed trends, not causal conclusions. Identify limitations and questions for our team.”

More advanced use:
An organization could develop a community-level AI analytics system that brings together approved wellbeing indicators over time and identifies emerging patterns across programs. For example, it could help identify whether participation in different community programs is changing alongside community-defined wellbeing measures and flag areas that may warrant further discussion or investigation.

Tradeoff:
This could help organizations see patterns that are difficult to identify manually and support better program planning. However, health and wellbeing information is particularly sensitive. Even supposedly aggregated information can become identifying in a small community. Advanced analytics can also create the temptation to treat correlations as causes or allow AI to make judgments about individuals. For this reason, the system should support community-defined questions and human interpretation, rather than automatically determining what a community's wellbeing needs are.

Data sovereignty: Health information requires particular care. Even aggregated data can become identifying in small communities. Best practice: do not upload identifiable health records or case notes to public AI tools. 

Before We Use Any AI Tool

Ask five questions:

1. Who owns the tool?
Know the company, its jurisdiction, and its business model. Each company has its own values it’s trained on & biases. 

2. Where does our data go?
Know where it is stored, processed, and backed up.

3. Who controls the data?
Check retention, training, access, deletion, and export provisions.

4. What alternatives exist?
Consider non-AI tools, human capacity, Indigenous-owned providers, and Indigenous-controlled infrastructure.

5. What happens if we leave?
Can we retrieve our information, delete it, and continue our work without the vendor?

The goal should not be to automatically use the most powerful AI. It should be to find the approach that gives us useful tools while retaining our ownership, control, access, and possession of our information.

As discussed above, the more advanced the AI solution is, the more data the AI requires access to, and therefore the greater the responsibility is to govern how it is used. Before adopting a more advanced AI system, we should ask whether the benefits justify the additional risks to our data, knowledge, privacy, skills, and self-determination.

Moreover, advanced systems should use a human-in-the-loop approach. AI can identify patterns, retrieve information, or prepare recommendations, but staff, leadership, knowledge holders, and community remain responsible for interpretation and decisions.

For sensitive community information, the long-term goals should be discussed with community members, and perhaps options like Indigenous-owned and/or Indigenous-controlled alternatives can be explored, rather than assuming that the largest technology companies are our only option.

The Environmental Tradeoff

More advanced AI solutions can provide greater benefits, but it can also require more computing, electricity, hardware, and infrastructure. Organizations should consider whether the value of an AI application justifies its environmental cost. When considering the next Seven Generations, efficiency should not be measured only by staff time or dollars saved. The environmental impact of the technology should also be part of the decision. New AI tools are frequently marketed as "the most advanced AI" but that doesn’t automatically mean "the best solution." A simple spreadsheet or statistical tool may sometimes accomplish the same task with far less computing and greater transparency and control.

Energy use can be difficult to understand when we are talking about AI because the electricity is consumed in data centres, rather than by a machine sitting in front of us. A person may only see a short response on a screen, while that response may require computers, storage, networking, cooling systems, and other infrastructure operating behind the scenes. Training a model is a massive, one-time energy cost. (Ex. OpenAI’s GPT-3 (175 billion parameters) consumed roughly 1,300 MWh, equal to the annual electricity use of about 130  homes.) Every single use afterward, especially image and video generation, adds its own energy cost on top.

Not all AI uses require the same amount of energy. A short text request is generally much less resource-intensive than generating an image or video. More advanced applications can also require additional computing because they may process large datasets, use multiple AI models, perform several steps in sequence, or run continuously.

One way to make these differences easier to understand is to compare AI use with familiar household electricity consumption. For example, a 1,500-watt electric kettle running for 10 minutes uses approximately 250 watt-hours (Wh), or 0.25 kilowatt-hours (kWh). If a simple AI text request uses a fraction of a watt-hour, hundreds of individual requests could have a similar electricity demand to that one household activity. However, a complex AI task—such as generating a high-resolution image or analyzing a large dataset—may use considerably more electricity than a simple text request. For example, if a complex task used 5 Wh, it would take about 50 such tasks to equal the 250 Wh used by a 1,500-watt electric kettle running for 10 minutes.

These comparisons should be treated as illustrations rather than fixed measurements. AI energy use varies depending on the model, hardware, length and complexity of the task, number of times the system is used, and the efficiency of the data centre operating it. As AI technology changes, the energy required for the same task may also change.

For an organization, the important question is therefore not simply "How much energy does AI use?" but: "How much energy does this particular AI use require, and is the benefit worth the environmental cost compared with another way of accomplishing the same task?"

For example, an organization might compare:

• AI-generated text with writing or editing the material directly.

• AI data analysis with Excel or conventional statistical software.

• An AI knowledge assistant with a well-organized document management system.

• AI-generated images with existing photographs, community artwork, or conventional graphic design.

• AI-generated video with existing footage or community-produced recordings.

A technology decision should consider not only the immediate benefit, cost, or staff time saved, but also the environmental resources required to operate the technology over time.

We don’t necessarily need to avoid AI because it uses energy. We should learn to use AI deliberately, choosing the least resource-intensive tool that can accomplish the purpose while still providing meaningful value to the organization and community. Think about when to use powerful tools carefully. You can use a fancy new state-of-the-art blowtorch to light a cigarette but most times a traditional pocket lighter works just fine.

6. Risk & Ethics Guardrails

A guardrail does not mean rejecting AI. It means using AI on our terms, with clear boundaries, human responsibility, community consent, and appropriate oversight. As AI becomes more powerful and gains access to more information, the need for strong guardrails becomes greater, not less. Below are suggestions for long-term AI-use:

• Never upload ceremonial, sacred, or restricted cultural knowledge to any public AI tool, regardless of the purpose — there is no version of “just for formatting” that makes this safe.

• Consent should be sought before using AI in ways that affect community data or knowledge. Public consent should never be assumed simply because a tool is convenient.

• Human and knowledge-keeper review before anything AI-assisted goes external to a funder, government, the public, or community.

• Ask vendors directly about data residency, training opt-out, and retention periods before adoption, and get the answer in writing.

• Reduce image/video generation Image and especially video generation carry a disproportionately higher energy cost per use and are not widely supported by the public because of the environmental cost. Use sparingly.

• Transparency & maintaining trust: let staff and community know where and how AI is being used in processes that affect them. Over-use and non-disclosed use of AI in communications has shown to erode trust within one another.

What can go wrong?

The guardrails above are the suggested rules. These are the specific risks if they're not followed:

• Extraction and Data Colonialism. Extraction is the taking of something valuable from a community, such as land, resources, labour, knowledge, or information, for the benefit of someone else, often without equivalent ownership or control remaining with the community. Data colonialism applies this same pattern to data: technology companies collect information about Indigenous people, communities, lands, cultures, and activities, then store, analyze, combine, or profit off of that information within systems they control. With AI, this can happen when community data is uploaded into their platforms. Once it’s entered into their tool, they are able to do whatever they want with it. Even if their current policies state they don’t claim ownership, they can update their policies without notice or consent. The concern is therefore not only whether data is private or secure, but who has the authority to collect it, control it, use it, benefit from it, and determine what happens to it over time.

• Surveillance and profiling. Big technology companies can learn about an organization not only from the information it enters into an AI tool, but also from how the organization uses the tool. An organization should ask not only, “What information are we giving this company?” but also, “What could this company learn about us from watching how we use its system?”, “How is the information collected on the inner workings of our organizations going to be used in the future? For the benefit of our people?” Historically this mass collection of information benefits very few.

• Community Consensus to use AI. Example: A community program has recordings and written notes from knowledge holders that were collected for a specific cultural or educational purpose. Staff later use an AI tool to transcribe and organize the material because it is faster than doing it themselves. No one asks the knowledge holders or community whether AI can be used for this new purpose. Even if the AI tool never publishes the material, the organization has changed how the knowledge is handled without obtaining consent for that use. As personal data becomes more and more valuable, this risk should be communicated to the community beforehand.

• AI dependence and over-reliance. Routine reliance on AI for writing carries a long-term risk. The communication skill required to write that grant narrative or that briefing from scratch atrophies over time. Are we able to discuss in detail and describe the purpose of documents that we are developing?  Do we trust that staff retains important information required for decision-making?

• Errors & Misinformation. AI can produce information that sounds accurate even when it is wrong. If no one checks the original source, a small mistake can make its way into reports, briefings, funding applications, or information shared with the community. The more an error is repeated, the harder it can be to correct.

7. A Good Mind Approach to Responsible AI Adoption

This is the process Kanikonriio AI has developed. People before technology, always.

1. Community and staff consultation: understand priorities, concerns, and comfort level with AI before any tool is chosen.

2. AI literacy and training: build shared understanding across staff of what these tools can and can't do, and where the risks are.

3. Policy development: write a data classification and AI-use policy grounded in the Good Mind checklist and OCAP, before broad adoption.

4. Small pilot: test one workflow in one program area, with clear data governance protocols, before expanding.

5. Community review: bring results back to community and leadership, adjust or discontinue based on what's learned, and only then scale.

8. AI Glossary

Plain-language definitions for terms used throughout this guide, useful for those who aren't AI specialists but need to follow the conversation.

9. Closing

Gratitude

This guide was written in the Good Mind, with gratitude to the knowledge keepers, community members, and leadership who continue to teach these principles, and in the hope that the tools described here are used to strengthen self-determination rather than erode it. Nya:weh for taking the time to read.