Chapter 2 ended with the load-bearing claim of this guide: the model is the data. What an AI system "knows" is a compressed statistical reflection of its…

Chapter 2 ended with the load-bearing claim of this guide: the model is the data. What an AI system "knows" is a compressed statistical reflection of its training data. Change the data and you change the model. Change the data substantially and you get a different model entirely. The technical fact is not contested. The political and legal consequences of the technical fact are deeply contested, and the contest organizes most of the rest of the guide. This chapter goes inside the training data itself. What is actually in there? Where does it come from? How is it collected? What does it cost (in human labour, in environmental footprint, in consent, in compensation)? Who owns it? Who has access? Whose knowledge is included, on what terms, with what acknowledgment? Why is this the most opaque area of corporate AI disclosure, every year, in every transparency assessment? The chapter is shorter than Chapter 2 because the technical foundations are now in place. It is more politically pointed because the training-data question is where the political stakes of AI become specific. You will leave with: a working picture of what's actually inside the training corpora of major AI models; the distinction between "open" data and "freely scraped" data that public discussion routinely collapses; the Whaanga framing of AI training as a new colonial extraction operation, engaged substantively rather than decoratively; the working test for evaluating any claim about training data; and the conceptual setup for Chapters 9, 11, and 13, each of which deepens this material in its own direction. ---

Where the training data actually comes from

When OpenAI trained GPT-4, what specifically did they train it on? When Anthropic trained Claude, when Google trained Gemini, when Meta trained Llama, when Cohere trained the Command series — what data was actually in those training corpora?

In most cases, we don't fully know. The companies do not generally publish detailed lists of what their models were trained on. The Stanford Foundation Model Transparency Index has identified training data as the lowest-scoring subdomain across nearly every developer in every year of the Index's existence (2023, 2024, 2025). When the 2025 Index scored 13 major foundation-model developers, six of them (Amazon, Google, Midjourney, Mistral, OpenAI, xAI) did not disclose basic information about their training data composition. This isn't oversight. It's structural.

GOBLIN FACTS — opacity clusters around data. In the 2025 Foundation Model Transparency Index, six major developers did not disclose basic training-data composition information. That is why "what was it trained on?" remains the most important unanswered question in the room.

What we do know, from public statements, leaked documents, lawsuits, and the few model developers who have disclosed substantially (notably IBM with 95/100 on the 2025 FMTI, and Anthropic and AI21 in the upper tier), is that the training corpus for any major large language model typically contains:

The "public internet." A snapshot or rolling window of text content scraped from the web. The most commonly cited starting point is Common Crawl, a non-profit that maintains an open archive of web pages crawled since 2008. Common Crawl alone is hundreds of billions of words. Major model developers filter, deduplicate, and re-weight Common Crawl for quality and harm-reduction, but the unfiltered version is the starting point. Common Crawl includes news websites, blogs, forums, social media (where publicly accessible), product reviews, government websites, academic publications (where freely accessible), Wikipedia, Stack Overflow, Reddit, archived discussion boards, fan wikis, recipe sites, fan fiction archives, dating profiles (where indexed), and an enormous quantity of low-quality and machine-generated text. The internet is a strange corpus to train an intelligence on (it contains both the world's best technical documentation and the world's worst conspiracy theories), and the strange properties of internet text shape what AI models do.

Books and long-form text. Foundation models are trained on books — typically tens to hundreds of thousands of digitized titles. The Books3 corpus, documented in the training of Meta's Llama and alleged in litigation to have figured in the training of other major models, contained approximately 196,000 books, including substantial copyrighted material from contemporary authors. The corpus was assembled by scraping a "shadow library" that distributes books without licensing. Books3 was the subject of high-profile lawsuits including the Sarah Silverman case against Meta and OpenAI, and most major model developers have since changed their book-data practices, though what they have changed to is mostly not publicly disclosed.

Code repositories. Then there is the source code: foundation models ingest enormous quantities of it, primarily from GitHub, GitLab, and similar platforms. GitHub's terms of service permit collection for AI training (a position that was itself a major industry debate when GitHub's Copilot launched in 2021), but much of the code on GitHub was uploaded by developers under open-source licenses (MIT, Apache, GPL, BSD) that specify particular conditions for use. Whether AI training constitutes "use" under these licenses is contested.

Academic literature. Foundation models are trained on scientific papers from arXiv (the open preprint server), PubMed Central (open biomedical), and where accessible, paywalled journal content from publishers including Elsevier, Springer, Wiley, and Taylor & Francis. The major academic publishers have responded variously: some have signed licensing deals with model developers (Elsevier with Microsoft, for example), some are suing. The line between "freely accessible" and "freely usable" runs through this category.

News content. News articles enter the corpus too, often through Common Crawl (which captures the publicly-accessible web versions of news sites). This is the substance of the Canadian newspapers' lawsuit against OpenAI: approximately 16.1 million Canadian news articles, owned by Torstar, Postmedia, the Globe and Mail, the Canadian Press, CBC/Radio-Canada, and Metroland, which the plaintiffs allege were used to train OpenAI's models without licensing or compensation. Similar lawsuits are active in multiple jurisdictions including the United States (the New York Times v. OpenAI case).

Other public-internet-accessible content. Reddit comments (sold to Google for AI training in 2024 in a US$60M/year licensing deal — one of the few public licensing arrangements). Stack Overflow contributions (a high-quality technical Q&A corpus that the company has both licensed to AI developers and seen used without licensing, with ongoing disputes). Wikipedia (CC BY-SA licensed and openly used). YouTube transcripts (Google's training of Gemini reportedly used substantial YouTube video transcripts — owned by Google). Twitter/X content (used in xAI's Grok training as direct platform integration).

Multimedia data for multimodal models. Image generators (DALL-E, Midjourney, Stable Diffusion, Google Imagen) are trained on billions of image-text pairs, typically collected by crawling websites and using image captions, file names, and surrounding text as labels. The LAION dataset (Large-scale Artificial Intelligence Open Network, a German non-profit) maintained the most commonly-used training corpus for open image-generation research; LAION-5B contained 5.85 billion image-text pairs. The dataset was taken offline in December 2023 after researchers from Stanford's Internet Observatory documented 1,008 confirmed instances of child sexual abuse material in the dataset, with more than 3,200 further entries flagged as suspected. A revised, filtered version was released in late 2024. The structural problem is not solved: the public internet contains material that should never be in a training corpus, and scraping at scale picks it up.

Synthetic data. Increasingly, AI training data includes data generated by other AI systems: a model trained on outputs from a previous model. The technique allows training-data scaling beyond what human-generated content alone can supply, and is now standard practice for fine-tuning and alignment. The structural concern this raises: if successive AI generations are trained substantially on AI-generated content, errors and biases in earlier generations propagate forward, potentially amplifying. Academic research on "model collapse" (Shumailov et al., 2024 in Nature) suggests this is a real risk, though the consequences are still being characterized.

The training corpus for a frontier foundation model in 2026 is, by any reasonable count, in the tens of trillions of tokens — equivalent to a substantial fraction of the digitized written record of humanity. The composition is partly disclosed and partly not. The provenance of each piece is partly traceable and partly not. The consent given by the humans whose work is in the corpus is, in most cases, absent.

GOBLIN FACTS — tokens are the measuring tape. Foundation-model training is usually counted in tokens: chunks of text roughly smaller than words. A corpus in the tens of trillions of tokens is not a bookshelf; it is a statistical copy of a large share of digitized human writing.

---

The distinction public discussion collapses — "open" versus "freely scraped"

One of the most consequential confusions in AI policy discussion is the collapsing of two very different things: data that is freely accessible and data that is freely usable. The distinction is anything but academic: it underwrites the entire creator-rights conversation in Canadian AI policy and globally.

Freely accessible means the data is available without paywall or login requirement. Most of the public internet is freely accessible in this sense. You can visit a news article, read it, see the photographs, and process its content yourself without paying.

Freely usable is a much narrower category. The Wikipedia article you can read is also freely usable, because Wikipedia is published under Creative Commons Attribution-ShareAlike (CC BY-SA), which explicitly grants permission for reuse with attribution and same-license requirements. The arXiv physics paper you can read is also freely usable, because the authors agreed to its open distribution terms when posting. The MIT-licensed code on GitHub is freely usable, with the conditions specified in the MIT license. But the news article you can read on the Globe and Mail's website is not freely usable. The newspaper has published it for public reading. The newspaper has not licensed it for arbitrary downstream use, including the training of commercial AI systems.

EXAMPLE — the street-parked car. A car parked on a public street is freely visible. That does not make it freely usable. The same distinction is doing work when a website can be read by the public but has not licensed its content for commercial model training.

🧌 GOBLIN CHECK — "Publicly available" is the AI industry's favourite phrase, and the goblin would like you to notice that it means we could reach it, not we were allowed to keep it. Your car is publicly visible. This has historically not settled the question of who gets to drive it away.

The distinction matters because the AI industry's defense of training-without-permission rests substantially on conflating these categories. "We trained on publicly available data" — OpenAI's stated position, the position of multiple defendant companies in the active Canadian and US copyright lawsuits — is a true statement that elides the distinction between publicly accessible and publicly usable. The news articles, the books, the academic papers, the artist portfolios, the photographer's websites are publicly accessible. They are mostly not publicly licensed for AI training.

ALIGNMENT — available for what? "Publicly available" quietly answers the wrong question. A thing can be free to read and still not free to copy into a commercial training set. When a company says it trained on public data, ask: available to look at, or licensed to use? The gap between those two is where most of the copyright fight lives.

The contested legal question, in Canada and globally, is whether existing copyright law permits AI training as a fair-dealing or fair-use exception, or whether AI training requires explicit licensing of the kind the AI industry has mostly not pursued. The contested fact-question (what data was used) is partly hidden by corporate non-disclosure. The contested legal question (whether the use was permissible) depends on facts that are partly hidden. The opacity is structural to the conflict.

Two distinguishing categories of "publicly available" content the guide will track separately throughout:

Open and licensed for reuse: Wikipedia (CC BY-SA), arXiv preprints (typically CC BY or compatible), most open-source code (MIT, Apache, GPL with conditions), Project Gutenberg public-domain books, government works (varying jurisdictional rules), and similar deliberately-released content. Training on this category raises fewer issues, because the rights holders have, explicitly or by license, granted reuse permission. But even here, license conditions matter: CC BY-SA requires the same license to be applied to derivative works, GPL has copyleft provisions, MIT requires attribution. Whether AI models comply with these conditions when they output text resembling training-data material is a complicated downstream question.

Publicly accessible but not licensed for AI training: news articles, copyrighted books, commercial photography, professional artwork, scientific papers behind paywalls (the publishers' position) or in open-access journals (the publishers' more contested position about whether open-access permits AI training), social media posts that platforms have not licensed for downstream use (most of Reddit before 2024, most of Twitter/X before Musk, most of Facebook's content), and the long tail of content posted by individuals without explicit licensing intent. This is the contested category. The AI industry has generally trained on it. The rights holders are increasingly objecting.

A third category worth naming: data that has been proactively licensed for AI training, through deals between AI companies and content owners. Reddit-to-Google (2024, US$60M/year), the AP-OpenAI deal, the Financial Times-OpenAI deal, the various academic publisher-to-Microsoft agreements, the recent New York Times-Amazon agreement (after the Times sued OpenAI). These deals are growing as content owners realize they have leverage. The deals also confirm, by their existence, that the industry recognizes the underlying material was previously being used without licensing. The licensing market now emerging can reasonably be read as the AI industry hedging its own legal position — you don't usually pay for what you're confident you may take for free. (The industry would call it risk management and input procurement. Both readings fit the same facts; the guide's lean here is declared.)

---

The human labour layer

A specific dimension of training data that public discussion routinely misses: most large AI models depend on substantial human labour, often paid at very low wages, often in the Global South, often involving exposure to traumatic content. This is not a peripheral fact about how AI gets built. It is structural to current AI training practice.

Two specific layers:

Data labelling. For supervised learning tasks, including most computer vision and many specialized AI capabilities, humans must label training data. Image labelling tasks like "click on all the pixels showing a car" are familiar to anyone who has completed a CAPTCHA. The same task scaled to professional labelling for AI training is performed by hundreds of thousands of workers globally, primarily through platforms like Amazon Mechanical Turk, Scale AI, Surge AI, and Toloka, with concentration in Kenya, Philippines, India, and other low-wage jurisdictions. The documented wages run low: TIME's 2023 investigation of OpenAI's Sama contracting in Kenya reported take-home pay of roughly US$1.32 to US$2 per hour, and 2023 Washington Post reporting on Scale AI's Remotasks platform in the Philippines documented payments that arrived late, below posted rates, or in some cases not at all. The labour is precarious by design: paid per task, without benefits, and typically without the protections that come with employment status.

Content moderation for alignment. When OpenAI fine-tuned ChatGPT to refuse harmful requests, the training data for that refusal capacity was generated by human workers reviewing extremely disturbing content (descriptions of violence, abuse, child exploitation, terror) and rating model responses as appropriate or inappropriate. **The 2023 TIME investigation** documented OpenAI's contracting through Sama in Kenya, paying workers under US$2/hour to review graphic and traumatic content for the alignment training data that made ChatGPT safer for end users. Workers reported lasting psychological harm. Similar concerns have been documented elsewhere: content-moderation contractors working on Meta's platforms brought litigation in Kenya over trauma exposure and pay, and contract raters working on Google's AI products have publicly disputed their wages and conditions, though the specific arrangements vary by company.

The structural fact: the AI products marketed in Canada as cutting-edge automation are, at their training-data layer, built on substantial human labour. That labour is concentrated in jurisdictions with weaker labour protections, paid at wages substantially below what equivalent work would pay in Canada or the United States, and increasingly involving exposure to traumatic content. The "AI-replaces-workers" framing is partial. AI replaces some kinds of workers in some kinds of jobs while simultaneously depending on other kinds of workers in other kinds of jobs, with the second set largely invisible from the framing.

This connects to Chapter 16's labour analysis: the labour displacement story is more complicated than either the alarmist or the dismissive version allows. Some labour is being replaced. Other labour is being created, at lower wages and worse conditions, or intensified through algorithmic management. All of this is happening at once, and the full picture requires holding it together.

---

Whaanga's framing — AI as a new colonizer

The Indigenous Protocol and AI Position Paper (2020), introduced in Chapter 1 and deepened in Chapter 9, contains an essay by Hēmi Whaanga (Ngāti Kahungunu, Ngāi Tahu, Ngāti Mamoe, Waitaha; a Māori linguist and te reo Māori specialist, now Professor and Head of School at Te Pūtahi-a-Toi, Massey University in Aotearoa New Zealand) with a framing that bears specific examination in the data chapter. The essay's title: "AI: a new (r)evolution or the new colonizer for Indigenous peoples?"

Whaanga's argument, distilled: AI systems are "existentially dependent on sucking up vast amounts of data." When that data includes Indigenous knowledge, language, cultural practices, environmental information, and traditional knowledge (and most large foundation models do include this content, because they are trained on the broad internet, including any Indigenous-authored or Indigenous-described material that has been digitized), the result is structurally extractive in ways that map directly onto historical patterns of colonial knowledge extraction.

A non-Indigenous-led guide cannot speak on behalf of this framing. What it can do is present the operational implications. Three are direct:

The training-data layer. Most major foundation models have been trained on Indigenous content — Wikipedia articles about First Nations, Inuit, and Métis communities and history; news coverage of Indigenous events; academic literature on Indigenous languages and cultural practices; in some cases, Indigenous-authored material posted to the open web; in some cases, archival material of Indigenous knowledge collected by colonial-era researchers and digitized into accessible archives. There is no documented case of this training occurring with the consent of the communities whose knowledge and content was used. The FNIGC OCAP Principles, the ITK National Inuit Strategy on Research, and the GIDA CARE Principles all ask whether Indigenous data is being used with Indigenous ownership, control, access, possession, collective benefit, authority, responsibility, and ethics. The current state of foundation model training is that, on the public record, none of those questions were asked before training.

The model-output layer. When an AI model is asked to generate content about Indigenous topics, the model produces output shaped by what was in its training data. If the training data was disproportionately Western-academic-authored material about Indigenous peoples, rather than Indigenous-authored material by Indigenous peoples, the model's output will reflect that disproportion. The voice of the colonial archive can be amplified by AI even when the underlying intent is neutral information access, because the underlying training data was shaped by the colonial archive's accumulation patterns. The effect is measurable, not metaphorical.

The new-knowledge-creation layer. If AI tools are increasingly used in education, research, communication, and creative production, and if those tools have absorbed Indigenous content under conditions that violate Indigenous data sovereignty principles, then the use of those tools to create new content about Indigenous peoples or in dialogue with Indigenous communities perpetuates the underlying violation. Each new piece of content is built on a substrate that did not engage Indigenous frameworks for using Indigenous data.

This is the substantive engagement Whaanga's framing requires. The label "AI as new colonizer" is not rhetorical. It is a structural claim about how AI training data works in relation to Indigenous knowledge, what the alternative would look like (training models that engage the data sovereignty frameworks rather than ignoring them), and what the costs of not engaging are. The guide takes the claim seriously without claiming to speak for it, and asks readers, and Canadian AI policymakers, to do the same.

A specific Canadian connection. The Abundant Intelligences research network introduced in Chapter 1 (co-led from Concordia, with the Lethbridge pod and international partners) is doing concrete technical work on what AI training that did engage Indigenous data sovereignty frameworks might look like: small models trained on data with proper consent, models built using Indigenous epistemological assumptions, prototypes that demonstrate alternatives to the extract-and-scale-up paradigm. The work exists. The federal AI strategy could engage it. AI for All does not currently commit to doing so.

---

The cost layer — what training data extracts and what it returns

Putting the previous sections together produces a working picture of what AI training data actually represents at an aggregate level. The picture is not the marketing version.

What training data extracts:

The accumulated written, visual, and audio output of human beings over centuries, where digitized and accessible. The professional output of writers, artists, photographers, journalists, programmers, scientists, scholars, and amateur enthusiasts. The cultural and knowledge content of Indigenous peoples globally, accumulated through colonial-era research practices and contemporary online presence. The human labour of hundreds of thousands of low-wage workers, primarily in the Global South, performing labelling and content moderation. The physical-substrate materials (cobalt, lithium, coltan, gallium — Chapter 8's Owen et al. finding) extracted to build the compute infrastructure that processes the training. The electricity, water, and embedded carbon of the training itself.

What training data returns to the people whose work, knowledge, and labour are in it:

For Wikipedia contributors and open-source developers: in most cases, the attribution and license-compatibility that their licenses required. Mixed compliance.

For published authors and journalists: in some cases, recent licensing deals (the AP-OpenAI deal, the Reddit-Google deal, the New York Times-Amazon deal, the various academic-publisher deals). In most cases, no compensation, no consent, and ongoing litigation. Canadian creators have received less than US creators in licensing-deal terms because Canadian collective rights organizations are less consolidated and Canadian publishers have less leverage.

For Indigenous peoples whose knowledge is in the training data: in most cases, nothing. Neither consent mechanisms, nor compensation, nor governance over how the data is used, nor disclosure of what was used.

For data-labelling and content-moderation workers: in the documented cases (TIME's 2023 reporting on OpenAI's Sama contracting in Kenya, the Washington Post's 2023 reporting on Scale AI's Remotasks), pay as low as roughly US$1.32 to US$2 per hour, typically without benefits or employment-status protections, and in some cases lasting psychological harm.

For the communities living near the extraction sites of the physical-substrate materials: in many cases, environmental degradation, displacement, and disruption of traditional land use.

For the general public: access to AI products, sometimes free at the point of use, that generate value primarily for the corporations that produced the models.

The economics, in compressed form: the AI industry extracts substantial value from data, labour, knowledge, and material substrates contributed by populations who are mostly not compensated, who mostly did not consent, who mostly do not have governance over the resulting models, and who are largely invisible from the marketing narrative. The industry returns substantial value to investors, executives, and (in subset of cases) end users. None of this is concealed. It is simply how current AI gets built, plainly traceable by anyone who follows the supply chain from chip extraction through labour markets through training data through model output to where the revenue lands.

The guide does not argue that this structure is "the problem with AI" requiring a single fix. The guide does argue that any honest account of AI's environmental, social, and political stakes requires seeing this structure, and that policy responses (Canadian and global) that don't engage it are responding to a partial picture.

---

The working test for evaluating training data claims

Putting the chapter's material together, the practical test the guide asks readers to apply to any AI training-data claim:

Source. What data was specifically used? Has the developer published a comprehensive list, a representative sample, or only vague characterizations? If the latter, why?

Consent. Was the consent of the rights-holders sought before training? If yes, through what mechanism — explicit licensing, opt-out, or assumed under fair-use/fair-dealing? If no, what is the developer's stated legal position on why consent wasn't required?

Compensation. Have rights-holders been compensated for the use of their material? If yes, through what mechanism — direct licensing, collective licensing, or revenue-sharing? If no, what is the developer's position on whether compensation should be required?

Indigenous data. Has Indigenous content been included in training data? If yes, were OCAP/NISR/CARE Principles engaged? If no engagement, what was the developer's reason? Is there a clear separation in the training corpus between Indigenous content used with consent and Indigenous content used without?

Verification. Are the developer's claims about training data composition, consent, and compensation independently verifiable? Has a third party audited the training corpus? If not, the claims are corporate self-disclosure and should be read accordingly.

Output traceability. Can model outputs that resemble specific training-data material be traced back to that material? Current technical capability here is limited but improving. Where output-to-source attribution is possible, the developer's framing of "novel generation" versus "reproduction" can be tested empirically rather than rhetorically.

Most current AI products will fail most of these questions. The guide's recommendation is not that readers refuse to use AI products; readers can make their own choices about that. The guide's recommendation is that readers, and the policy-makers shaping Canadian AI strategy, read the failures as the structural conditions of how current AI is being built, and decide whether those conditions are acceptable or whether they should be changed.

---

The receipt on the raw material

CHAPTER RECAP — you now have: - A working picture of what's actually in the training corpus of major AI models — internet content (filtered and unfiltered), books (often acquired without licensing), code (under varying license conditions), academic literature (with growing licensing disputes), news (the substance of the Canadian newspapers' OpenAI lawsuit), social media (with growing licensing markets), multimedia data (including the LAION CSAM finding), and synthetic data (with model-collapse concerns). - The distinction between freely accessible and freely usable that public discussion routinely collapses — and the operational basis of the Canadian creator-rights conversation. - The human-labour layer of training-data preparation, including the low-wage data-labelling workforce in the Global South and the documented Kenyan content moderation contracting for OpenAI's alignment training. - Whaanga's "AI as new colonizer" framing engaged substantively — the training-data-layer, model-output-layer, and new-knowledge-creation-layer implications of training models on Indigenous content without engaging Indigenous data sovereignty frameworks. - The aggregate picture of what training data extracts (centuries of accumulated human output, Indigenous knowledge, low-wage labour, extracted materials, electricity and water) and what it returns to the contributors (in most cases, little or nothing). - The working test for evaluating any AI training-data claim: source, consent, compensation, Indigenous data engagement, verification, output traceability.

The next chapter (Chapter 4) takes the data picture this chapter has built and goes inside the physical layer — the chips that process the data, the racks they sit in, the cooling systems that keep them running, the supply chains that produce the chips, the materials extracted to build the chips, and Suzanne Kite's Lakota hardware-building protocol from the Indigenous Protocol and AI Position Paper as the structural counter-framing to industrial-scale hardware extraction.

You can now read any AI training-data claim with the structural equipment to ask the right questions. Most current claims will fail most of those questions. That failure is the structural condition of how current AI is being built, and engaging it honestly is the precondition for any serious Canadian AI policy.

---

Bias label for this chapter: structural-political analysis of AI training data, with explicit naming of the gap between corporate self-disclosure and operational reality. Author lean: skeptical of "publicly available" framings that elide the freely-accessible-vs-freely-usable distinction; sympathetic to creator-side and Indigenous-led critiques of training-without-consent; willing to name the human-labour layer in the Global South as structural to current AI rather than peripheral; explicit that the manual is not Indigenous-led and cannot speak on behalf of the Whaanga framing while engaging it substantively. Corporate self-disclosure (Stanford FMTI findings, lawsuit defendant positions) labelled and read accordingly. Indigenous-led primary sources (Whaanga essay, OCAP/NISR/CARE frameworks) treated as authoritative within their own frames.

Primary sources cited or relied on in this chapter: Stanford Center for Research on Foundation Models, 2025 Foundation Model Transparency Index; Canadian newspapers v. OpenAI (Ontario Superior Court CV-24-00732231-00CL); Books3 corpus litigation including Silverman et al. v. OpenAI; New York Times v. OpenAI; LAION dataset withdrawal documentation (Stanford Internet Observatory, December 2023); Shumailov et al., "AI models collapse when trained on recursively generated data" (Nature, 2024); TIME magazine investigation of OpenAI's Sama contracting (January 2023); Whaanga, "AI: a new (r)evolution or the new colonizer for Indigenous peoples?" (Indigenous Protocol and AI Position Paper, 2020); First Nations Information Governance Centre, OCAP Principles; Inuit Tapiriit Kanatami, National Inuit Strategy on Research (2018); Global Indigenous Data Alliance, CARE Principles; Reddit-Google licensing deal disclosure (2024). Detailed citations in the Sources appendix.

---