Source Library Appendix

The AI conversation will keep moving. Watch whether the Canadian newspapers' case against OpenAI is decided, settled or otherwise resolved. Watch which…

The AI conversation will keep moving. Watch whether the Canadian newspapers' case against OpenAI is decided, settled or otherwise resolved. Watch which compute projects come into service, what happens to the treaties Canada has signed, and how the next transparency assessment sets its criteria. In the Wonder Valley litigation, read what a ruling actually decides about the duty to consult. The habit worth keeping is the same one the book asks of you now: date the claim and read the receipt. The facts will move. The habit of asking who counted, what they could see and where their judgment came in will still be useful. It will also help when two people whose work you trust reach different conclusions. You can use the questions in a newsroom, at work or when a claim shows up in your own feed. How far they take you depends on the evidence you can get and the people you can ask. Take what helps you look a little closer. ---

The conversation you will encounter

Lire cette section en français

The Canadian AI conversation in the years following this guide's writing will continue to oscillate among recognizable patterns. The patterns will be present in news coverage, in policy debate, in workplace conversations, in academic exchange, in social media discourse, and in the broader cultural processing of what AI is and what it should be. Recognizing the patterns is the first portable skill.

The strategic-investment-versus-existential-risk oscillation. The Canadian AI conversation regularly cycles between framing AI as an opportunity Canada must seize before competitors do, and framing AI as a civilizational risk that requires precautionary restraint. Both framings are real positions with serious advocates. Both also tend to obscure the more granular policy questions the guide has documented. When you encounter either framing, the question to ask is what specific deployment, what specific policy, what specific institutional arrangement is being discussed, because the strategic-investment-versus-existential-risk oscillation rarely engages those specifics directly.

The corporate-ethics-statement pattern. Major AI corporations publish ethics frameworks, principles documents, and governance commitments at regular intervals. The frameworks read well. Their cycles often correlate with specific commercial pressures: preparing for product launches, responding to regulatory pressure, repositioning after public criticism. The Stanford FMTI documented (Chapter 18) that the companies producing the most ethics-statement material have, in aggregate, become measurably less transparent over the period when the statements were proliferating. This is not corruption; it is the structural feature of corporate ethics frameworks operating within commercial constraints. What matters about any such statement is what it actually constrains in operational practice that the corporation would not otherwise have done.

The Indigenous-rights-as-decoration pattern. Federal AI policy increasingly invokes Indigenous perspectives, often through brief acknowledgments or symbolic engagement with specific Indigenous voices. Substantive integration of Indigenous data sovereignty frameworks (OCAP, NISR, CARE) into working policy is much rarer. When acknowledgment language appears, the practical test is whether the policy engages the frameworks the acknowledged organizations have published, or whether the acknowledgment is the engagement.

The federal-strategy-versus-provincial-leadership pattern. Quebec has, across multiple AI-relevant policy domains (privacy, deepfakes, automated decision-making), moved earlier and more comprehensively than the federal government. The pattern is structurally significant for constitutional reasons (Chapter 17) and politically significant because it produces fragmented Canadian AI coverage. Federal framing that doesn't engage provincial variation invites a direct question: does it acknowledge what Canadian AI policy actually looks like for the 77% of Canadians outside Quebec?

The "AI for All" versus "AI for them" pattern. Federal AI strategy framing emphasizes broad benefit ("250,000 new jobs," "$200B economic growth," "AI literacy for all Canadians"). Civil society framing emphasizes specific affected populations (women workers facing displacement, racialized communities facing biased deployment, Indigenous communities facing data sovereignty violations, workers facing surveillance intensification). The two are not contradictory (each can be partly right), but they engage substantively different policy questions, and the thing worth pinning down for either is who specifically benefits, who specifically bears costs, and through what mechanisms.

The technical-fix versus structural-change pattern. Many AI policy proposals frame problems as engineering challenges with engineering solutions (better algorithms, better data, better audits, better metrics). The Lewis/Whaanga/Yolgörmez epistemology critique (Chapter 1, Chapter 15) suggests that some of the most consequential AI policy questions are structural rather than technical: whose interests AI is built to serve, whose values get encoded into systems, what foundations the systems rest on. When you encounter technical-fix framings, the test is whether the proposed fix addresses the surface symptom or the underlying structure. The distinction determines what the fix can accomplish.

These patterns will recur. Recognizing them is half the work. The other half is having portable tools for what to do once you recognize the pattern.


What to do when you encounter a specific AI claim

Lire cette section en français

Pick one claim you’ve heard recently. It could be about a tool at work, a promise from Ottawa or a number somebody sent you. Chapter 19 has the source categories and the seven questions. Here’s how to use them on the claim in front of you.

Step 1. Identify the source. A government announcement, a union brief and a company’s own report each give you a different starting point. Chapter 19’s “Bias-mapping methodology” section lists eight categories. Use them to decide what to look for in the evidence.

Step 2: Identify the claim's scope. What specifically is being claimed? Is the scope narrow (a specific deployment, a specific outcome, a specific policy provision)? Or is the scope broad (a general assertion about AI, about Canadian AI strategy, about AI's effects)? Broad-scope claims usually require more skepticism because they aggregate across many specific cases the speaker is unlikely to have engaged equally.

Step 3: Identify what is in scope and what is excluded. Most AI claims are technically defensible within their stated scope and selective in what they include. The corporate environmental claim that measures only inference and excludes training, embodied carbon, and induced demand is the documented example pattern (Chapter 8). The strategic-investment claim that measures jobs created without netting against displacement is another. When you can identify what is excluded, you can identify what the claim does and does not establish.

Step 4. Pick the questions that fit. What does the tool do? What data does it use? Who feels the effects, who sets the rules and who pays? Can anyone check the claim? Who got a say and what happens if something goes wrong? A claim may answer two of those questions and leave the ones you care about open.

Step 5: Triangulate across source categories. If the claim is corporate self-disclosure, what does the academic peer-reviewed literature say about the same topic? If the claim is government framing, what does civil society advocacy say about the same policy? If the claim is mainstream press critique, what do the underlying primary sources show? Triangulation across categories produces more reliable understanding than relying on any single category.

Step 6: Engage genuine disagreement honestly. Where the guide has documented rigorous voices disagreeing (Geist/Craig vs TWUC/ACTRA on copyright, Acemoglu vs Autor on labour, the algorithmic-fairness-literature internal debates, the AGI-timeline disagreement among senior AI researchers), the disagreement is genuine and the guide does not resolve it. When you encounter similar disagreement in future AI conversations, the guide's methodological commitment is to engage the disagreement rather than to collapse it toward one side. Recognizing genuine disagreement is itself analytical work.

Step 7: Form a working position without claiming to be conclusive. After the analytical work, you can hold a working position on the specific claim: what it establishes, what it doesn't, where you have confidence, where you have uncertainty. The working position is not a conclusion. It is the basis for further inquiry, further reading, further conversation. AI conversations are ongoing rather than concluded.

This seven-step practical application is the working version of the toolkit Chapter 19 assembled. Doing this for any specific claim takes time. Doing it well develops with practice. The guide hopes readers find the practice valuable enough to continue.


The conversations specific to your context

Lire cette section en français

Beyond the abstract analytical practice, AI conversations are contextual. The conversation in your workplace differs from the conversation in your community, which differs from the conversation in your professional or creative field. The guide offers brief guidance for several specific contexts readers will navigate.

In your workplace. If AI is used in screening, monitoring, decision support or customer service, Chapters 10 and 16 give you questions to ask. CUPE's Senate brief proposes notice, an explanation, human review, limits on biometrics and audits. In Ontario, start with the public job posting if it is covered. For an ongoing workplace tool, check the policies, agreement and laws that apply to that workplace.

In your professional or creative field. Different professional and creative communities face different AI integration questions. Writers, performers, musicians, journalists, photographers, designers, and other creative workers face the consent and compensation questions that Chapter 11 engaged. Healthcare workers, social workers, educators, and other professional service workers face the deployment and decision-support questions that Chapter 15 and Chapter 16 engaged. Workers in regulated industries (financial services, telecommunications, broadcasting) face the sectoral regulator and compliance questions Chapter 17 engaged. The specific organizations representing your professional or creative community are doing AI policy work; engaging them is how individual professional positions become coordinated political voice. ACTRA, the Writers' Union of Canada, the Directors Guild of Canada, the Canadian Medical Association, CUPE, Unifor, and many others have specific AI policy positions you can engage rather than develop from scratch.

In your community. If you live in an Indigenous community, the data sovereignty questions Chapter 9 engaged are not abstract. If you live in a province with strong privacy legislation (Quebec, BC, Alberta, Manitoba), your protections differ from those of Canadians elsewhere. If you live in a community where AI infrastructure is being proposed or built (data centres, hyperscaler regions, the Sturgeon Lake Wonder Valley project and successors), the local consultation processes and the broader sovereignty questions Chapter 6 and Chapter 9 engaged are immediate. The guide cannot provide community-specific guidance, but the analytical infrastructure travels.

In your personal decisions. You will make decisions about using AI products — chatbots, image generators, productivity tools, customer service interactions, social media platforms that incorporate AI, employment applications, healthcare interactions, financial services. The guide's analytical infrastructure helps you evaluate the specific deployments you encounter. The personal decisions aggregate into the broader political-economy that policy responds to. Individual choice does not substitute for political action, but individual analytical clarity contributes to it.

In civic engagement. You can engage in Canadian AI policy through specific mechanisms — submitting to government consultations (ISED, the OPC, the Privacy Commissioner, Treasury Board, provincial commissioners, sectoral regulators), supporting civil society organizations doing AI policy work, voting based on AI policy positions where they differ across parties, participating in academic and community AI literacy initiatives, and engaging directly with elected representatives at federal, provincial, and municipal levels. The political coalition for closing the AI policy gaps the guide has documented requires citizen engagement; the analytical infrastructure the guide has built supports that engagement.


The guide's limits — what you cannot expect from it

Lire cette section en français

Honest engagement requires naming what the guide cannot do. The chapter engages this directly.

The guide cannot keep up with specific facts. The guide was written in mid-2026. Specific facts will change — court decisions, regulatory developments, corporate announcements, policy shifts, new research findings, new institutional structures. Some of what the guide documents will be operationally outdated within months of publication; other parts will remain useful for years. The analytical infrastructure ages better than the specific facts. Readers who want current Canadian AI policy information should consult primary sources directly — the OPC, the Treasury Board, ISED, civil society organizations, academic researchers, investigative journalism, the Stanford FMTI annual editions.

The guide cannot substitute for domain expertise. The guide provides a general framework for thinking about AI claims. It does not provide specific expertise in healthcare AI, financial services AI, legal AI, educational AI, or any other domain that has its own specialized literature and practice. Where specific decisions require domain expertise, the guide points readers to that expertise rather than substituting for it.

The guide cannot resolve genuine intellectual disagreement. The chapter has named this repeatedly — Geist/Craig vs TWUC/ACTRA, Acemoglu vs Autor, the algorithmic-fairness-literature debates, the AGI-timeline disagreement. The guide presents both sides of these disagreements and asks readers to engage them. Readers who want the guide to tell them who is correct will not find that answer. The guide's methodological position is that the disagreements are productive when engaged honestly rather than resolved prematurely.

The guide cannot predict political outcomes. Chapter 20 mapped institutional pathways for closing documented policy gaps. Whether the political conditions for activating those pathways will emerge depends on factors the guide cannot predict — election outcomes, coalition coordination, public opinion shifts, international developments, technological change. The pathways are mapped regardless; whether they are used is contingent.

The guide cannot be value-neutral. The chapter has named the guide's lean in every chapter through the bias-label closing convention. The guide operates from a critical-engaged position — skeptical of corporate self-disclosure, sympathetic to Indigenous-led frameworks, attentive to worker concerns, willing to name government framing as government framing. Different readers will hold different positions; the guide presents its lean honestly while making the underlying analytical infrastructure portable across reader positions.

The guide cannot be the comprehensive Canadian AI public education resource. It is one contribution. Other contributions (from Indigenous-led organizations, from labour organizations, from creator organizations, from privacy commissioners, from academic researchers, from investigative journalism, from community education initiatives) are doing different and complementary work. Readers who want comprehensive Canadian AI literacy should engage multiple sources, not just this one.

The guide cannot save Canadian AI policy. Closing the documented gaps requires political coordination at a scale beyond what any single document can produce. The guide's contribution is to make the analytical infrastructure portable; what readers do with that infrastructure is theirs to determine.


What the guide has tried to do

Lire cette section en français

The guide opened with the claim that the word "intelligence" in artificial intelligence is doing more work than most people realize, and that the side that gets to define intelligence inside an AI system usually wins the argument about whether the system is working. Everything since has been an attempt to engage that work directly: rigorous voices presented in their own framings rather than as caricatures; corporate claims engaged on their merits while the patterns of selective disclosure got named; Canadian government AI policy treated as policy, substantive and real and partial and contested; the analytical infrastructure built to travel further than the conclusions it was demonstrated on; and, at every contested question, a refusal to hand you a resolution the evidence doesn't support. The toolkit was always the product. The conclusions were the demonstration. And the limits — one document about a moving field, a non-Indigenous-led book engaging Indigenous-led frameworks, an analysis that cannot resolve what reasonable people contest — are features of doing the work honestly, not failures to do the work comprehensively.


The benefit side, kept calibrated

Lire cette section en français

View receipt #127

Here’s what asking for the receipt looks like in a clinic. In a randomized trial across 16 Kenyan primary care clinics, researchers tested a language-model assistant embedded in the medical record system. The trial found no statistically significant reduction in treatment failure within 14 days. That was the outcome the researchers chose to test. If you hear that AI helped with a clinical task, keep asking what happened to patients. That is a different measurement, and in this trial the researchers actually took it.

A guide that spends this many pages on footprints, gaps, and overstatement owes you the other half of its own method, because the goblin's skepticism was never cynicism. The same questions that deflate a hype claim also protect a real one. Applied honestly, they show that AI demonstrably does some things well.

The strongest cases are narrow and measurable. AlphaFold predicted the structures of nearly every known protein and earned a 2024 Nobel Prize, though its makers are careful to call the outputs valuable hypotheses rather than experimental facts. AI weather models now beat the gold-standard European forecaster on its own benchmarks. Closer to home, an early-warning system at Unity Health Toronto was associated with a 26 percent drop in unexpected deaths on one hospital ward. These are not press releases; they are peer-reviewed results.

Run them through the toolkit anyway, because the method does not switch off for good news. Who counted, and how? The Unity Health result is a before-and-after study on a single ward, not a randomized trial, and the same category of tool produced the Epic sepsis model that missed roughly two-thirds of cases in external testing. Is the benefit demonstrated or extrapolated? A 2023 audit found data leakage inflating results across seventeen scientific fields, and in Canada a "funded" project means a compute bill was subsidized, not that the product worked. The honest position is not "AI is good," any more than it was "AI is bad." It is that some AI applications have earned their claims and some have not, and you now have the equipment to tell which is which.


What you carry forward

Lire cette section en français

Readers will carry different things forward from the guide. The chapter closes with what it hopes readers will carry forward, recognizing that what readers actually carry will be theirs to determine.

An understanding that AI is not magic. The pattern-matching framework from Chapter 2, the data-is-the-model framing from Chapter 3, the physical-substrate analysis from Chapter 4 — these establish that AI is industrial-scale pattern-matching on extracted data running on physical hardware that depends on specific supply chains and labour patterns. Once AI is demystified, the policy questions become legible. Magic resists policy; engineering does not.

An understanding that AI is not nothing. The capabilities are real. The deployments are real. The economic, social, and political stakes are real. The guide has refused the alternative magic of "AI is just autocomplete" with the same seriousness as the magic of "AI is general intelligence." What is there is genuinely powerful pattern-matching with substantial implications.

Questions you can carry. Where does the source lean? What can its evidence tell you? Who has a way to challenge the result? Keep the questions that help you get further into the claim.

A recognition that the Canadian AI conversation is contested. Reasonable people disagree about important things. The disagreements are productive when engaged honestly. The guide's bias-labelling methodology helps you track who is leaning which way and why; the analytical infrastructure helps you evaluate claims regardless of source lean.

A working understanding of where Canadian AI policy actually sits. Investment-focused federal strategy, sectoral-fragmented regulation, Quebec leadership ahead of Ottawa, post-AIDA regulatory gap, substantial transparency gaps relative to peer jurisdictions, specific institutional resources doing meaningful work within the broader gaps, documented policy mechanisms available for closing the gaps without political conditions currently aligned to activate them.

An honest sense of what the guide is and isn't. One contribution. Useful for a period of time. Replaceable by better analysis as the field develops. The guide's value is in helping readers engage AI conversations more productively; the guide's obsolescence is the natural endpoint of useful Canadian AI literacy work.

Most of all, the guide hopes readers will carry forward a recognition that AI is a Canadian conversation Canadians get to participate in. Canada cannot opt out of it (the technology and the deployment are already here) and cannot fully control it, since both are global. But it is positioned to engage substantively, drawing on institutional resources, political traditions, Indigenous-led frameworks, creator-rights and labour-protection traditions, and a privacy-and-civil-liberties inheritance that contribute distinctive perspectives the global conversation needs. Whether Canada engages substantively or watches it happen depends on what Canadians do. This guide has tried to help with the doing.


Closing

Lire cette section en français

This is where the guide ends. The conversation continues.

Maybe you read all twenty-one chapters. Maybe you came straight from the toolkit with one claim in mind. Either way, try the questions the next time someone hands you a number.

When you have the receipt, the next question is who can act on it. That might be a colleague, a union, a regulator or someone on your council. You’ll know a little more about what to ask them for.

🧌 GOBLIN CHECK — the last one. You don't need this book in your hands to do the work. Three questions fit in any pocket: Who counted that? What got left out of the count? Can I see the receipt? Ask them of the next AI claim you meet — including the claims this book made. Especially those.
ALIGNMENT — point the compass at this book. You have the questions now. The last move is the uncomfortable one: ask them of the guide in your hands. Where does it lean, what did it leave out, where are its receipts? A field guide that can't survive its own method hasn't earned your trust either. The goblin isn't in the margins anymore. It's you. Go hoard.

— Alex Yesilcimen and Brine Shadewater Labs Vancouver, Canada June 2026


Bias label for this chapter: closing methodological synthesis with explicit engagement of the manual's own limits and the conditions under which the analytical infrastructure travels. Author lean: methodological commitment to the manual's organizing principles of refusing premature resolution of genuine disagreements, presenting rigorous voices in their own framings, naming both the manual's substantive contributions and its honest limits; explicit recognition that closing the documented Canadian AI policy gaps requires political coordination the manual cannot produce but the analytical infrastructure can support; willing to close the manual at the same level of intellectual seriousness it opened with, including by refusing to write a triumphal conclusion. The chapter draws on the entire manual for source material; specific source citations appear in earlier chapters and in the Sources appendix.


End of manual.


🧌 GOBLIN CHECK — the last one. You don't need this book in your hands to do the work. Three questions fit in any pocket: Who counted that? What got left out of the count? Can I see the receipt? Ask them of the next AI claim you meet — including the claims this book made. Especially those.