If you remember one thing from this chapter, make it this: the word "intelligence" in "artificial intelligence" is doing more work than most people realize, and it has never been a neutral technical term. Almost every public argument about AI, whether it's what AI can do, what it threatens, or what Canada should build, rests on assumptions about what intelligence is. Those assumptions are contested. The contest is not new. And the side that gets to define "intelligence" inside an AI system usually wins the argument about whether the system is working. This chapter does three things. It explains, in working terms, what is and isn't happening inside the AI systems that have suddenly arrived in Canadian conversations about jobs, sovereignty, the environment, and democratic life. It places that technology in its actual history, which is much longer than the 2022 ChatGPT launch most coverage starts from. And it sets up the methodology the rest of the guide will use: how to read sources that disagree, how to spot the bias in your own preferred sources, and how to hold contested questions as contested rather than collapsing them into your existing position. This is the longest chapter. The rest of the book gets shorter and more focused once these foundations are in place. ---
AI is older than the news cycle suggests
When Canada's Treasury Board launched its public register of federal AI use in November 2025, the press release contained a sentence worth re-reading: "AI has been in use in the Government of Canada for decades, with systems dating as far back as 1994."
Nineteen ninety-four. Before broadband internet was a household concept. The federal government has been using AI techniques, at first rules-based expert systems, then statistical learning, eventually the neural networks that dominate today, for more than thirty years.
What changed in late 2022 wasn't the arrival of artificial intelligence in Canada. It was the arrival of a particular kind of AI, large generative language models, in the consumer market, in a form that could be used by anyone with a browser and a credit card. The technology that became ChatGPT had been in development inside research labs (including Canadian labs in Toronto, Montreal, and Edmonton) for at least a decade before it became a consumer product. The shock of late 2022 was less about technical novelty than about distribution: a kind of AI that had previously been the property of researchers and corporations was suddenly the property of everyone.
This matters because the framing of AI as a sudden 2022 arrival distorts almost every downstream question. If AI is new, the response can be cautious experimentation, public consultation, slow uptake. If AI is the latest stage in a thirty-year continuum of automation that has already reshaped Canadian work, services, and government, which is closer to the truth, the questions become different. They become questions about which AI techniques are arriving in which contexts, with what particular risks and benefits, not "should Canada do AI."
A working definition the guide will use throughout: AI is the umbrella term for techniques that get computers to do tasks that previously required human judgment. Some of those techniques are old (decision trees, expert systems, statistical regression). Some are new (transformer-based language models, diffusion image generation). They share the umbrella, but they are not the same thing, and the umbrella term hides important differences. We'll come back to this.
---
A short tour of what's actually happening inside
You don't need to understand the math to read the rest of this guide. You do need a working mental model. Here's one.
Machine learning is the family of techniques that lets a computer find patterns in data without those patterns being explicitly programmed. Show a system a million labelled photographs of cats and dogs, and it will learn, through a process of statistical adjustment across thousands of internal parameters, to distinguish them on new photographs it hasn't seen. The system doesn't "know" what a cat is in any sense recognizable to a person who has held a cat. It has built a statistical map of pixel patterns that correlate with the human-supplied label "cat."
This is the first conceptual move readers need: machine learning systems do not "understand" in any sense that matches the everyday human meaning of understanding. They produce outputs that correlate with what understanding would produce, on the kinds of inputs they were trained on. Sometimes the correlation is excellent. Sometimes it breaks in ways that are revealing: a state-of-the-art image classifier will confidently identify a school bus as an ostrich if you alter a handful of pixels invisible to the human eye. The system is doing pattern-matching of remarkable sophistication. It is not doing what your brain is doing.
Deep learning is a subset of machine learning that uses artificial neural networks: layered mathematical functions inspired (loosely) by the structure of biological neurons. A neural network with one or two layers is "shallow." One with dozens or hundreds is "deep." The shift to deep networks in the 2010s, combined with the availability of vast training datasets and specialized hardware (graphics processing units, or GPUs, originally designed for video games), is what enabled most of the AI capabilities Canadians now interact with.
Large language models (LLMs), the technology behind ChatGPT, Claude, Gemini, Cohere's products, and the federal government's expanding AI deployments, are a specific kind of deep learning system. They are trained on enormous quantities of text (most of the public internet, large portions of digitized books, code repositories, and other written material) to predict, given some input text, what text would plausibly come next. That prediction task, when scaled up enough, produces a system that can write essays, answer questions, summarize documents, write code, translate languages, and do hundreds of other things that look impressively like understanding.
Two facts about LLMs that the rest of the guide depends on.
First: LLMs are trained on text without the consent of most of the humans who produced that text. This is the empirical claim at the centre of the Canadian newspapers' November 2024 lawsuit against OpenAI, the Writers' Union of Canada's revised author contracts, the Alliance of Canadian Cinema, Television and Radio Artists' (ACTRA) collective agreements with explicit AI prohibitions, and most of the global creator-rights movement around AI. The legal status of training-without-permission is contested in Canada (we'll spend Chapter 11 on this). The empirical fact of it is not. The Canadian news case alone covers approximately 16.1 million owned and licensed works.
Second: LLMs do not have access to facts in the way a database does. They have access to statistical patterns about how words and concepts have been used together in their training data. When an LLM tells you that the capital of Canada is Ottawa, it is not consulting a fact table. It is generating the most statistically likely completion of "the capital of Canada is ___" based on millions of times that pattern appeared in its training material. Most of the time, that produces correct answers. Sometimes, when the training data is sparse, contradictory, or out of date, it produces confident-sounding falsehoods. The technical term in the field is "hallucination," which is itself a revealing word choice: it implies a system that normally perceives reality and occasionally mis-perceives, when the more accurate description is a system that never perceives reality at all and sometimes happens to produce outputs that match it.
You can absorb the rest of the guide without remembering the technical vocabulary. The two conceptual moves are the load-bearing ideas: these systems are pattern-matchers rather than understanders, and they are trained on material whose ownership is contested.
---
The contested word
Now we come to the harder claim.
Throughout the AI research community, "intelligence" has a specific working definition, most commonly attributed to a 2007 paper by Shane Legg and Marcus Hutter: "Intelligence measures an agent's general ability to achieve goals in a wide range of environments." This is the operating definition behind almost all serious work on what's called Artificial General Intelligence (AGI), the hypothesized future AI system that would match or exceed human performance across most cognitive tasks.
In 2024, a research paper published in the peer-reviewed journal AI & Society did something unusual. Jason Edward Lewis, Hēmi Whaanga, and Ceyda Yolgörmez — the principals of the Abundant Intelligences research program, a major Indigenous-led AI research initiative co-led from Concordia University in Montréal and Massey University in Aotearoa New Zealand, with pods including the University of Lethbridge, Bard College, and the University of Hawai'i West Oahu — traced where Legg and Hutter's foundational definition actually came from.
Legg and Hutter's 2007 paper surveys some seventy definitions of intelligence and synthesizes its own. Among the sources in that survey, and the one Lewis, Whaanga, and Yolgörmez follow down the citation trail, is a 1994 Wall Street Journal statement called "Mainstream Science on Intelligence," assembled by Linda Gottfredson. The statement was published as a defence of the IQ research community after Richard Herrnstein and Charles Murray's The Bell Curve generated public controversy. It contains lines like: "Members of all racial-ethnic groups can be found at every IQ level. The bell curves of some groups (Jews and East Asians) are centered somewhat higher than for whites in general. Other groups (blacks and Hispanics) are centered somewhat lower than non-Hispanic whites."
This is one of the documents sitting in the citation lineage of the working definition of "intelligence" used in serious AGI research. Lewis and colleagues argue it is not an incidental one. The Lewis-Whaanga-Yolgörmez paper quotes it directly. Their argument is not that AGI researchers are personally racist. Their argument is sharper: the field's foundational definition of its central concept incorporates a particular, historically situated, and empirically contested conception of intelligence — one shaped by the IQ-research tradition that produced The Bell Curve — and that conception is not flagged or examined when AGI research builds on it.
The paper's opening claim, which I want to reproduce because it organizes everything that follows: "The artificial intelligence industry-academic complex does not have an ethics problem, it does, however, have an epistemology problem."
Translate that out of academic register: the issue isn't that AI researchers are insufficiently ethical. It's that the foundational concepts AI is built on — what counts as intelligence, what counts as a goal, what counts as success in achieving it — are not the neutral technical primitives they are presented as. They are inheritances from particular intellectual traditions — mostly Western, mostly twentieth-century — and largely indifferent to the question of whether other traditions have alternative concepts that might do the work better.
This is the conceptual move that organizes the rest of the guide. None of it means AI doesn't work; it plainly does, in the narrow technical sense of producing useful outputs on the tasks it's trained on. Nor does it mean Indigenous and non-Western traditions reject AI. Lewis and his colleagues are explicit that they are not in the "decomputerization" camp; they want to engage AI from different foundations, not refuse it. What it means is that **when someone tells you AI is or isn't intelligent, you should ask which definition of intelligence they're using, and where that definition came from.
ALIGNMENT — whose definition? When someone tells you AI is or isn't "intelligent," the argument is already a fight about a word. Ask which definition they're using and where it came from. A surprising number of public AI arguments are really arguments about a definition nobody put on the table.** Almost no one in public discourse asks that question. The guide will ask it repeatedly.
A line worth carrying with you, from the same body of work: the opening epigraph of the 2020 Indigenous Protocol and AI Position Paper attributes to Olana Kaipo Ai (Kānaka Maoli) the formulation "Aloha is the intelligence with which we meet life." You don't have to adopt that as your operating definition. You do have to notice that it is one, and that for the Hawaiian creators of the Position Paper, it does the work that "general ability to achieve goals in a wide range of environments" does for AGI researchers, while making completely different commitments about what intelligence is for.
---
How to read this guide
The guide you're reading is not neutral. No source on AI is. What it tries to do, and what it asks you to do alongside it, is to be honest about which way each source leans, and to put leaning sources next to each other so the contest stays visible.
Here is the methodology, in plain terms.
Every source you read on AI has a position relative to commercial, political, or institutional interests. Government press releases lean toward justifying the policy they're announcing. Corporate sustainability reports lean toward the numbers that flatter the corporation. Union statements lean toward worker concerns. Indigenous-led research leans toward Indigenous epistemologies. Academic peer-reviewed work leans toward what the discipline has agreed counts as rigour. Industry-commissioned research firms (a category we'll come back to) lean toward the conclusions their commissioning client would prefer. The lean is not corruption. It's the way human institutions work. The problem is when readers absorb a leaning source as if it were neutral.
The guide will label every source as it cites it, using eight categories:
- Government framing (press releases, ministerial speeches, strategic announcements) leans toward justifying current policy.
- Government operational (technical guides, registers, statistical reports) leans less, though it pays to watch the scope and definitions.
- Government leaked/draft documents (working papers not meant for public release) are useful as evidence of internal disagreement, and lean according to who leaked them and why.
- Civil society / advocacy sources (unions, Indigenous organizations, public-interest non-profits) lean toward the constituency the organization represents.
- Academic peer-reviewed work (journal articles, conference proceedings) leans toward what the field counts as rigour, which is itself a position.
- Corporate self-disclosure (sustainability reports, technical blog posts, voluntary transparency) leans toward the company's interests even when technically honest.
- Corporate via commissioned-research firm (economic-impact reports, market projections) leans toward the commissioning client while wearing the costume of third-party analysis.
- Mainstream press critique (investigative journalism, opinion analysis) leans toward the outlet's editorial position.
Through the rest of the guide, when we cite a source, you'll see the lean named. When two sources disagree, we'll show you both leans and let you do the work of weighing them.
EXAMPLE — three reviews of the same restaurant. The owner's write-up, a rival's one-star pan, and a regular's scribbled notes all lean, and you'd be foolish to trust any one of them on its own. Read together, the leans start to cancel and the real place comes into focus. That's the whole bias-mapping move: don't throw out the slanted source, line it up against the others.
A practical demonstration. When Google publishes that the median Gemini text prompt uses 0.24 Wh of energy and 0.26 mL of water, that is corporate self-disclosure from a company with an interest in defending against the "AI is environmentally catastrophic" narrative. The underlying methodology is technically sophisticated; the senior authors (Amin Vahdat, Jeff Dean) are credible engineers; the numbers may well be accurate within their stated scope. And the framing is strategic: choosing the median (the friendliest statistic), choosing inference (excluding training and embodied carbon), choosing the comparison "less than nine seconds of TV" (a flattering frame). None of those observations cancels the others out. The honest reading is: "Google's data appears methodologically careful within its scope, and Google has chosen the scope to flatter the result." That's a more useful frame than either "Google is lying" or "Google has settled the question."
🧌 GOBLIN CHECK — Google says a Gemini prompt costs less energy than nine seconds of TV. Three small questions, doing heavy lifting: median prompt (the friendliest statistic in the drawer), inference only (training not included; batteries also not included), and verified by, per footnote 2 of Google's own blog post announcing the paper: nobody. The number can be honest and the frame can still be a salesman. Both at once. That's the entire methodology of this book, arrived one chapter early.
A second practical demonstration. When the Canadian Union of Public Employees (CUPE, 800,000 members by its own count) tells the Senate Social Affairs Committee in March 2026 that AI deployment in workplaces requires comprehensive consultation, transparency, and limits on algorithmic management, that is civil-society advocacy from an organization whose members include exactly the workers most affected by AI deployment. The lean is toward worker protection. The underlying data CUPE cites is independent and rigorous: Statistics Canada's analysis that 31% of Canadian workers are in jobs highly exposed to AI with low complementarity (the jobs most likely to be transformed or displaced; "dramatically change or disappear" is the interpretive gloss, not StatCan's wording); the doubled risk for women-dominated jobs; the documented cases of closed-caption workers, medical transcriptionists, and hospital dispatch workers already losing jobs. The lean is real; the evidence is also real. The honest reading is: "CUPE has worker-protection interests and the empirical case it builds is substantiated by independent data." Discounting their submission because they are a union would be as wrong as accepting it without noticing they are one.
GOBLIN FACTS — labour risk has a number. CUPE's brief leans toward worker protection, but the 31% exposure figure it cites comes from Statistics Canada. The goblin rule is simple: notice the lean, then check whether the evidence survives it.
Throughout the guide, you'll practice this kind of reading. By Chapter 21, you'll have a portable toolkit for it that you can use on any AI claim you encounter, including this guide itself. The guide calls this method the bias-mapping methodology throughout, and it is the foundation tool the rest of the analytical infrastructure depends on.
---
Three concepts that organize what follows
The rest of the guide is structured around three big questions and one through-line. Here are the concepts you'll need to track.
Training versus inference. This is the most important technical distinction for understanding AI's environmental, economic, and rights debates. Training is the (very expensive, very energy-intensive) process of building an AI model from data. Inference is what happens every time you use the model: you type a prompt and get a response. Training a major large language model can consume tens of millions of dollars of electricity over months. Each subsequent prompt consumes a fraction of a watt-hour. The environmental and copyright debates often conflate these; the labour debate often ignores them entirely. Holding them separate will make most of what follows clearer.
Direct versus indirect effects. When you read that AI is becoming more efficient (that a Gemini prompt now uses 33 times less energy than it did a year ago, per Google's August 2025 measurement), your instinct may be to conclude that AI's environmental footprint is shrinking. The Canadian-led research of Sasha Luccioni (at Hugging Face Montréal), with Emma Strubell (Carnegie Mellon) and Kate Crawford (Microsoft Research / USC), published at the 2025 ACM FAccT conference, makes the opposite case: efficiency gains in AI are systematically reinvested into more AI use, larger models, broader deployment, and induced demand in other sectors. The economic name for this is Jevons' paradox: efficiency improvements often increase total consumption rather than decreasing it. Chapter 8 is where this gets its full treatment. For now, just hold that the per-query number going down and the total environmental footprint going up are not contradictions; they're the same phenomenon viewed at different scales.
Sovereignty in three layers. Canadian AI conversations use the word "sovereignty" in at least three distinct senses, and the conversations often slide between them without anyone noticing. National sovereignty is the question of whether Canadian data, Canadian models, and Canadian AI infrastructure are under Canadian rather than foreign (mostly U.S., increasingly Chinese) jurisdiction. Personal sovereignty is the question of whether individuals have meaningful control over how their data and likenesses are used. Indigenous data sovereignty is the question of whether Indigenous peoples retain ownership, control, access, and possession of data about themselves, their communities, their lands, and their knowledge. That question has its own decades-long framework (the First Nations Information Governance Centre's OCAP® Principles, the Inuit Tapiriit Kanatami National Inuit Strategy on Research, and the international Global Indigenous Data Alliance's CARE Principles) that predates and exceeds the current AI debate. These three are not the same question. A "sovereign cloud" deployment from Telus or Bell answers the first one, partially, while doing nothing about the second or third. The guide will name which sovereignty is at stake every time the word appears.
ALIGNMENT — sovereignty compass. When a claim says "sovereign AI," ask which direction the compass points: national jurisdiction, personal control, or Indigenous data governance. A claim can align with one layer while leaving the others untouched.
And the through-line. The single most important finding in our source library, the one that connects environment, sovereignty, intellectual property, and ethics into one geography rather than four separate topics, is this: a 2023 peer-reviewed study by Owen and colleagues found that 54% of the world's energy-transition-mineral extraction projects — mines producing cobalt, lithium, copper, rare earths, and other materials computing hardware also depends on — are located on or near Indigenous peoples' lands; 69% once peasant lands are included; and 62% of the projects on those lands sit in high-water-risk locations. One precision note the guide owes you up front: Owen et al. counted projects, not material volumes, and their study is about the energy transition broadly, not computing specifically. The guide extends the finding to AI hardware because the mineral families overlap; the extension is the guide's move, flagged here so you can weigh it. That finding was cited in the Luccioni/Strubell/Crawford paper to make a specific point: AI's environmental footprint isn't only a question of electricity and water at the data centre. It's also a question of where the physical substance of AI comes from, who is displaced when it is extracted, and which ecosystems are degraded by its extraction. When we get to Chapter 9 (sovereignty) and Chapter 15 (ethics), the same Owen et al. number will appear again, doing different work. The chapters of this guide are not independent topics. They are different views onto the same underlying geography.
GOBLIN FACTS — the mineral map is part of the AI map. Owen et al. counted 54% of energy-transition-mineral projects on or near Indigenous peoples' lands. The guide extends that geography to AI hardware because chips, servers, batteries, and power systems draw on overlapping mineral families.
---
What you've got now
CHAPTER RECAP — you now have: - A working mental model of how machine learning, deep learning, and large language models actually function, enough to follow the rest of the guide without needing technical background. - An understanding that AI in Canada has a thirty-year history rather than a three-year history, and that the current moment is the consumer-market arrival of a specific kind of AI rather than the arrival of AI itself. - The conceptual move that "intelligence" is not a neutral technical term, that its definition in AI research has a specific and contestable history, and that engaging that history is the beginning of clear thinking about AI rather than a digression from it. - A methodology for reading sources: every source leans, the lean is not corruption, the lean is something you can name and account for, and the guide will name it every time. - Three concepts to track through the rest of the book: training versus inference, direct versus indirect effects, three layers of sovereignty. - The through-line that connects the chapters: AI is a physical thing made of contested material, drawn from contested land, training on contested data, deployed by contested institutions, and the chapters of this guide are different views onto that one underlying reality.
The next chapter goes deeper on how machines actually learn: specifically, what happens inside a transformer-based language model when it does what looks like understanding. If you want the deeper technical layer, read Chapter 2 next. If you want to follow the Canadian story, skip ahead to Chapter 5. The guide is designed to work in both orders.
---
Bias label for this chapter: educational synthesis with explicit Indigenous-led epistemological framing in Section 3. Methodology in Section 4 disclosed. Where contested claims appear, leans labelled. Author lean: critical-engaged with current Canadian AI policy, sympathetic to Indigenous data sovereignty frameworks, skeptical of corporate self-disclosure without independent verification.
Primary sources cited or relied on in this chapter: Treasury Board AI Register (November 28, 2025); Lewis, Whaanga & Yolgörmez, "Abundant intelligences: placing AI within Indigenous knowledge frameworks," AI & Society 40(1):2141–2157, 2024; Indigenous Protocol and AI Position Paper (2020); Google Cloud Blog "Measuring the environmental impact of AI inference" (Vahdat & Dean, August 21, 2025); Luccioni, Strubell & Crawford, "From Efficiency Gains to Rebound Effects: The Problem of Jevons' paradox in AI's Polarized Environmental Debate," FAccT 2025; CUPE Senate Brief, March 2026; Owen et al. (2023); Canadian newspapers v. OpenAI (Ontario Superior Court CV-24-00732231-00CL, 2024–2025). Detailed citations in the Sources appendix.
---
🧌 GOBLIN CHECK — Google says a Gemini prompt costs less energy than nine seconds of TV. Three small questions, doing heavy lifting: median prompt (the friendliest statistic in the drawer), inference only (training not included; batteries also not included), and verified by, per footnote 2 of Google's own blog post announcing the paper: nobody. The number can be honest and the frame can still be a salesman. Both at once. That's the entire methodology of this book, arrived one chapter early.
Recap
- A working mental model of how machine learning, deep learning, and large language models actually function, enough to follow the rest of the guide without needing technical background.
- An understanding that AI in Canada has a thirty-year history rather than a three-year history, and that the current moment is the consumer-market arrival of a specific kind of AI rather than the arrival of AI itself.
- The conceptual move that "intelligence" is not a neutral technical term, that its definition in AI research has a specific and contestable history, and that engaging that history is the beginning of clear thinking about AI rather than a digression from it.
- A methodology for reading sources: every source leans, the lean is not corruption, the lean is something you can name and account for, and the guide will name it every time.
- Three concepts to track through the rest of the book: training versus inference, direct versus indirect effects, three layers of sovereignty.
- The through-line that connects the chapters: AI is a physical thing made of contested material, drawn from contested land, training on contested data, deployed by contested institutions, and the chapters of this guide are different views onto that one underlying reality.
Sources
- Treasury Board AI Register (November 28, 2025)
- Lewis, Whaanga & Yolgörmez, "Abundant intelligences: placing AI within Indigenous knowledge frameworks," AI & Society 40(1):2141–2157, 2024
- Indigenous Protocol and AI Position Paper (2020)
- Google Cloud Blog "Measuring the environmental impact of AI inference" (Vahdat & Dean, August 21, 2025)
- Luccioni, Strubell & Crawford, "From Efficiency Gains to Rebound Effects: The Problem of Jevons' paradox in AI's Polarized Environmental Debate," FAccT 2025
- CUPE Senate Brief, March 2026
- Owen et al. (2023)
- Canadian newspapers v. OpenAI (Ontario Superior Court CV-24-00732231-00CL, 2024–2025).