You will encounter AI claims for the rest of your life. They will come from corporations selling AI products, from governments announcing AI strategies…

You will encounter AI claims for the rest of your life. They will come from corporations selling AI products, from governments announcing AI strategies, from regulators publishing AI rules, from academics making AI predictions, from union analyses of AI workplace impact, from Indigenous-led frameworks engaging AI epistemologies, from journalists reporting on AI deployments, and from activists making AI-related demands. The claims will not stop, and most of them will be honest within their scope while leaning in patterned ways toward the interests of whoever is making them. This chapter exists to give you portable tools for reading those claims — tools that work across the AI conversations you will encounter beyond this guide. It is structured as a toolkit rather than a synthesis. The previous sixteen chapters did the analytical work; this one extracts the moves that travelled across them and makes them explicit so you can apply them yourself. The argumentative substance is in the earlier chapters; the readable infrastructure for evaluating AI claims independently is here. > EXAMPLE — the stat your uncle forwards. The family chat lights up: AI is going to create (or destroy) a million jobs, source a screenshot. Before you fire back, run the three pocket questions — who counted it, what got left out, can you see the receipt? Nine times out of ten the link is a press release from whoever benefits. The toolkit isn't only for policy papers. It's for Tuesday. You will leave the chapter with: the bias-mapping methodology as a complete framework with the categories defined and worked examples provided; the seven working tests from the contested-questions chapters consolidated into a single portable structure; the through-line analytical moves (three sovereignty layers, Jevons' paradox, the epistemology framing, the four-category actor map, the convergent transparency precondition) explicit as tools rather than as embedded arguments; and worked examples showing the toolkit in action on the kinds of claims you will actually encounter. ---

One: The bias-mapping methodology — the foundation tool

Chapter 1 introduced the bias-mapping methodology that has organized every chapter since. The methodology is the foundation tool because everything else in the toolkit depends on the reader being able to recognize the lean of a source before evaluating the claim the source makes.

The categories, as the guide has used them throughout:

Government framing (press releases, ministerial speeches, strategic announcements) tilts toward justifying current policy. Watch the gap between commitments made and the mechanisms to deliver them, and the gap between an announcement and the document it summarizes. Government operational material (technical guides, registers, statistical reports, audits) runs lower than framing, but the scope and definitions are where the action is: watch the categorical judgements about what counts (the Treasury Board AI Register's "low-risk commercial products" exclusion is the documented example). Government leaked or draft documents reveal a lean set by who leaked them and why, and earn their value as evidence of the gap between internal candour and public framing. The CBC leak of the AI for All draft is the case in point.

Civil society / advocacy (unions, Indigenous organizations, public-interest non-profits, professional associations) leans toward the constituency the organization represents, but the underlying evidence is often independently strong even when the framing is advocacy-positioned; that evidence is what you evaluate. Academic peer-reviewed work leans toward what the discipline counts as rigour, which is itself a value commitment; watch the specific methodological choices, the disciplinary scope, and the limits of what the methodology can address.

Industry / corporate self-disclosure (sustainability reports, technical blog posts, voluntary transparency frameworks, ethics commitments) leans toward the company's interests even when technically honest. Watch what is in scope and what is excluded, what verification status is named, what comparison frames are chosen. The commissioned-research-firm variant (economic-impact reports and market projections delivered by professional research firms to clients) leans toward the commissioning client while packaged as third-party analysis; the commissioning relationship itself is usually disclosed in fine print and rarely foregrounded. Mainstream press critique leans toward the outlet's editorial position, but the specific evidence a piece is built on is often independently verifiable regardless of the framing.

The methodology's core methodological move. Every source has a lean. Recognizing the lean is not the same as dismissing the source. The lean is what you account for; the underlying evidence is what you evaluate. Sources from different lean categories can all be partially correct simultaneously, and triangulating across categories often produces more reliable understanding than relying on any single category.

A worked example. In June 2026, three sources made claims about Canadian AI workforce effects: the federal government projected 250,000 new AI-related jobs over five years; the Bank of Canada reported no widespread AI unemployment signal yet; and the CUPE Senate brief documented specific sectoral displacement and projected 31% of Canadian workers in jobs likely to dramatically change. The federal projection is government framing (lean: toward justifying the strategy). The Bank of Canada finding is government operational (lean: lower, but watch the macro-vs-sectoral framing). The CUPE documentation is civil society advocacy (lean: toward worker protection, with substantial independent evidence behind it). A reader using the bias-mapping methodology would not collapse these into a single answer; they would recognize that all three can be partially correct (the projected new jobs may materialize while sectoral displacement also occurs and aggregate unemployment doesn't immediately spike), and would form a more accurate working picture than any single source provides.

🧌 GOBLIN CHECK — Triangulation, goblin edition: the government tells you the jobs created, the union tells you the jobs lost, the central bank tells you the average didn't move. Nobody is lying. Each is holding one corner of the receipt. Your job is to tape it back together.

---

Two: The working tests, consolidated

The contested-questions chapters (3, 8, 9, 10, 11, 12, 13, 14, 16) each closed with a working test specific to the chapter's topic. The tests are structurally similar because they share underlying analytical logic. This section consolidates them into a single portable framework.

The questions that work across most AI claims:

On the underlying activity. What specifically is the AI doing? Where is it deployed? Who deploys it? Who owns the deploying entity? What category of work does it perform?

On the data. What data was used to train or operate the AI? What was the consent status of the data subjects? What compensation, if any, was provided to data sources? Is the training data composition publicly disclosed?

On the impact. What outcomes is the AI producing? Are the outcomes evenly distributed across affected populations? Is there documented bias, displacement, surveillance intensification, or other measurable effect? On whom?

On the institutional context. What jurisdiction governs the AI deployment? What regulators have authority? What disclosure requirements apply? What audit or appeal mechanisms exist?

On the financial substrate. Who pays for the AI deployment? Who profits? Where does the capital come from? Where does the revenue flow?

On verification. Are claims about the AI independently verifiable? Or are they self-reported by the deploying entity? If verified, by whom and through what methodology?

On consent and recourse. Have affected individuals or communities given meaningful consent to the AI deployment? Do affected parties have recourse if the AI produces harm? Through what mechanism?

These seven questions can be applied to virtually any AI claim. The chapter's earlier working tests can be derived from these seven for specific contexts — Chapter 3's training-data test, Chapter 8's environmental test, Chapter 9's sovereignty test, Chapter 10's privacy test, Chapter 11's copyright test, Chapter 13's deepfake test, Chapter 15's ethics test, Chapter 16's labour test, and Chapter 18's transparency test are all specific applications of this general framework to particular AI claims.

A worked example. When AI for All committed to the "Sovereign Foundations" pillar in June 2026, applying the framework:

Underlying activity: The pillar commits to scaling Canadian AI champions (primarily Cohere), building domestic AI infrastructure (Bell AI Fabric), and supporting sovereign AI capability. The commitment is real but partial. It does not address most of the underlying dependencies (US-designed chips, Taiwan-fabricated chips, Dutch lithography equipment).

Data: The Canadian AI workloads will use training data that is mostly not Canadian-sourced. Most large foundation models are trained on the global public internet, which is dominantly English-language and US-cultural. Canadian sovereignty over compute infrastructure does not translate to Canadian sovereignty over training data.

Impact: The pillar projects 250,000 new AI-related jobs and $200B in economic growth. The projections are projections, not commitments. The displacement that may accompany the growth is not separately projected.

Institutional context: The federal government leads, with ISED as the policy department. Provincial governments have substantial jurisdiction over AI in healthcare, employment, and education that the federal pillar cannot reach. Indigenous data sovereignty frameworks exist but are not integrated into the pillar.

Financial substrate: Public funding ($300M+ in compute access, $200M+ across the research institutes, $240M for Cohere's compute build-out). Private capital is substantially American (Cohere's investors include US-headquartered NVIDIA, Oracle, Salesforce, Cisco).

Verification: The pillar's claims are corporate-government self-disclosure. The Stanford FMTI does not currently scope to Cohere. Independent verification of the operational sovereignty claims is limited.

Consent and recourse: Canadian citizens have not been substantially consulted on the strategy beyond the formal AI for All development process. The strategy does not include affected-citizen consent mechanisms or specific recourse provisions.

Applying the seven-question framework to AI for All's "Sovereign Foundations" produces a substantially more granular picture than the strategy framing provides.

ALIGNMENT — the three questions, pocket-sized. Strip the toolkit down and three questions do most of the work: who counted it, what got left out, and can you see the receipt? They fit on the back of a transit pass, and they work on this guide too. Point them at me whenever you like. The pillar is real, partial, dependent on substrates outside Canadian sovereignty, projecting outcomes that are not commitments, operating without comprehensive verification, and engaging consent only at the highest level of consultation. This is the picture the framework produces. It is more useful for evaluating the pillar than either accepting it as comprehensive or dismissing it as inadequate.

---

Three: Through-line analytical moves — the deeper toolkit

Beyond the bias-mapping methodology and the working tests, the guide has been building through-line analytical moves that travel across multiple chapters. These are the more sophisticated tools the toolkit offers — moves that change how you frame AI questions, not just how you evaluate specific claims.

The three sovereignty layers (Chapter 9). When you encounter any claim about AI sovereignty, ask which layer is being engaged. National sovereignty (jurisdiction, control over infrastructure)? Personal sovereignty (individual control over data and decisions)? Indigenous data sovereignty (collective Indigenous rights over data about Indigenous communities)? A claim that addresses only one layer is not the same as a claim that addresses all three. Most current Canadian AI sovereignty claims engage national sovereignty only, sometimes only partially. BC + AI's maple-leaf test from Chapter 5 is the operational check for whether a sovereignty claim is doing what the framing implies.

Jevons' paradox (Chapter 8). For any claim that AI efficiency improvements reduce environmental impact, the question is whether the improvement is measured per-unit or in aggregate. Efficiency gains in AI have been systematically reinvested into more AI use, larger models, broader deployment, and induced demand in other sectors. Per-query efficiency can improve while aggregate footprint grows; both can be true at once, and corporate claims that infer the second from the first are usually misframing rather than dishonest. The same logic applies to productivity, labour-saving, and cost-reduction claims. Efficiency at one scale does not equal benefit at another.

The epistemology-not-ethics framing (Chapter 1, Chapter 15). Faced with an AI ethics framework or a bias-and-fairness claim, ask whether it engages the system's foundational assumptions or only its outputs. The Lewis/Whaanga/Yolgörmez analysis distinguishes ethics framings (incremental improvement within current foundations) from epistemology framings (what the foundations are doing and whose interests they serve). Most corporate and government AI ethics work operates in the first register. Indigenous-led frameworks, the broader algorithmic accountability literature, and the constructive alternative programs (Abundant Intelligences and related work) operate in the second. Recognizing which one is in play changes what the claim is actually doing.

The four-category actor map (Chapter 7). Claims about "Canadian AI development" almost always concern one of four actor types: foundation model developers (in Canada, primarily Cohere), foreign hyperscalers (AWS, Microsoft, Google through their Canadian regions), Canadian telecommunications incumbents (Bell, Telus, Rogers), and federally-aligned research institutes (Mila, Vector, Amii, CIFAR, CAISI). The four have different commercial interests, different relationships to the federal strategy, different jurisdictional positions, and different transparency obligations. A claim that doesn't specify which is ambiguous in ways that often serve a specific actor.

The convergent transparency precondition (Chapter 18). For any domain-specific AI policy claim — copyright, environment, labour, sovereignty, deepfakes, privacy — ask whether the underlying disclosure infrastructure exists to make the policy work at all. Most areas depend on disclosure Canada has not built. Transparency is the precondition for most of them: copyright enforcement needs training-data disclosure, bias work needs algorithm disclosure, sovereignty analysis needs supply-chain disclosure. The precondition is the structural test for whether any other policy claim can do the work it implies.

The commissioned-research-firm pattern (Chapter 7). Economic-impact claims about AI investments — jobs created, GDP contribution, productivity gains, regional benefits — should prompt the question of who produced the underlying research. Increasingly, these narratives come from professional research firms (Public First is the documented Canadian-relevant example) working for the entities being measured. The same firm sometimes produces studies for more than one side of the same conversation: for the foreign hyperscalers whose footprint is being measured, and, by citation, for the federally aligned institutes that lean on those numbers. Reading the output as commissioned research rather than independent measurement changes how the numbers should be weighted.

The structural-bias frame (Chapter 15, Chapter 16). AI bias claims engage one of three senses of bias, and conversations often slide between them: statistical bias (mathematical deviation from expected values), social bias (favouritism for or against demographic groups), and structural bias (reproduction of broader power and resource distributions). The three are related but not identical. Statistical bias yields to technical fixes; social bias yields to specific audits and adjustments; structural bias requires structural change that the technical fixes cannot reach.

---

Four: Worked examples — the toolkit in action

The toolkit is more useful when readers see it applied to specific examples. This section provides three worked examples of the kinds of claims you will encounter.

Example 1: A corporate AI sustainability report claiming the company has achieved "carbon-neutral AI."

The bias-mapping move: corporate self-disclosure. Lean: toward favourable environmental framing. Watch: what is in scope, what verification exists.

The seven-question framework applied: - Underlying activity: What specifically is being measured — the company's data centres, its AI training, its AI inference, the embodied carbon of its chip supply chain, the carbon impact of induced consumption from AI products? The scope choice substantially determines the result. - Verification: Has the carbon-neutral claim been verified by an independent third party? Through what methodology? Many "carbon neutral" claims rely on renewable energy certificates (RECs) that have been criticized in the academic literature (O'Brien 2024 documented in Chapter 8) as substantially overstating actual emissions reduction. - Scope: Does the claim address Scope 1 (direct emissions), Scope 2 (purchased energy), Scope 3 (supply chain and use)? Most corporate carbon-neutral claims address Scopes 1 and 2 partially and Scope 3 minimally.

The through-line moves applied: - Jevons' paradox: Is the per-unit efficiency improvement being presented as if it implies aggregate emissions reduction? The Luccioni/Strubell/Crawford framework (Chapter 8) applies here. - Transparency precondition: Are the underlying figures publicly disclosed at sufficient granularity for independent verification?

The synthesis: corporate "carbon-neutral AI" claims are usually selective in scope, partially verified, and operating in a framework where the rebound effects of expanded AI deployment are not counted as the company's emissions. The claims are often technically defensible within their stated scope while misleading about aggregate environmental impact.

Example 2: A federal government announcement that an AI strategy will create 100,000 new jobs in a specific sector.

The bias-mapping move: government framing. Lean: toward justifying the strategy.

The seven-question framework applied: - Underlying activity: What does "AI-related job" mean in the projection? Does it include workers directly building AI products? Workers using AI tools in their existing roles? Workers in expanded industries that AI-enabled productivity might support? - Impact: Is the 100,000 figure offset by projected displacement in other sectors? Government strategy projections often present gross creation without netting against displacement. - Verification: Is the underlying modelling published, with assumptions and confidence intervals? Or is the number presented without methodology? - Institutional context: Does the strategy include enforceable mechanisms ensuring the projected jobs materialize, or is the figure a projection that the government commits to pursuing without specific delivery requirements?

The through-line moves applied: - Bias mapping: Government strategy projections lean toward justifying the strategy. Independent academic-economics analysis (Acemoglu, Autor from Chapter 16) provides the counterweight. - The four-category actor map: Which actors actually employ the projected workers? Are the foundation model developers (Cohere) creating direct employment at scale, or is the employment in adjacent industries that the strategy hopes will expand? - Commissioned-research-firm pattern: If the projection was produced by a research firm working for the government or for institutes that benefit from the strategy, that affects how it should be weighted.

The synthesis: government job-creation projections from AI strategies are projections, not commitments, frequently presented gross without netting against displacement, often produced through methodologies whose assumptions aren't fully disclosed, and with limited mechanisms for accountability if the projections don't materialize. The projections may be honest within their assumptions; the assumptions are usually optimized for politically-useful results.

Example 3: A civil society organization's report on AI surveillance in Canadian workplaces.

The bias-mapping move: civil society / advocacy. Lean: toward worker protection.

The seven-question framework applied: - Underlying activity: What specific surveillance practices are documented? Are they from primary research, from member testimony, or from secondary reporting? Are the cases identified by employer, by sector, by region? - Impact: What documented harm is the report establishing? Are the harms quantified, or are they presented qualitatively? Is there a causal chain from surveillance to specific outcomes? - Source: Who funded the report? What was the methodology? Was it peer-reviewed?

The through-line moves, here, run differently than in the first two examples — the lean is the report's strength, not its weakness. Bias mapping: the lean is real (toward worker protection), and it is also where the organization contributes most, documenting what corporate disclosure does not. Structural-bias frame: is the report addressing statistical bias in workplace AI, social bias, or structural bias (the broader power-and-resource implications of workplace AI surveillance)? The structural frame produces different policy implications. Transparency precondition: does the report identify what disclosure infrastructure would be needed for the documented harms to be systematically addressed, or does it focus on the documented harms without the structural transparency analysis?

The synthesis: civil society reports on workplace AI surveillance are advocacy-positioned but often substantively rigorous. The underlying documentation is frequently independently verifiable. The framing tends toward worker protection, which is the constituency the organization represents. Discounting the report because it comes from an advocacy organization would miss the substantive evidence; accepting it without noticing the lean would miss the framing choices. A careful read engages both.

---

Five: Limits of the toolkit

The toolkit is not unlimited. Honest engagement requires naming what it can and cannot do.

What the toolkit can do. Recognize the lean of a source. Apply systematic questions to AI claims across domains. Identify scope choices, verification status, and what is and isn't disclosed. Connect specific claims to broader analytical frames (Jevons' paradox, the epistemology framing, the three sovereignty layers, etc.). Recognize patterns of corporate selective disclosure, government framing, commissioned-research substitution for independent analysis, and the structural-bias dimension that technical fixes don't address.

What the toolkit cannot do. Resolve genuine intellectual disagreements among rigorous voices (Acemoglu vs Autor, Geist/Craig vs TWUC/ACTRA, the algorithmic-fairness-literature internal debates, the AGI-timeline disagreement among senior AI researchers). The toolkit can help readers recognize that the disagreement is genuine and engage both sides; it cannot tell readers which side is correct.

What the toolkit asks of readers. Time and attention. The framework produces better reading of AI claims, but it requires actual reading — checking the methodology, examining the scope choices, verifying the verification status, triangulating across source categories. The toolkit doesn't substitute for engagement; it makes the engagement more productive.

What the toolkit cannot substitute for. Domain expertise on specific AI applications. A medical AI system raises different questions than an employment AI system, which raises different questions than a creative AI system. The toolkit provides the general framework; the specific analysis comes from domain experts. That division of labour (portable framework here, specific expertise elsewhere) is the toolkit's limit and also its honest design. The guide's commitment throughout has been to refusing to resolve genuine intellectual disagreements prematurely, and the toolkit operationalizes that: where rigorous voices disagree (Acemoglu vs Autor; Geist/Craig vs TWUC/ACTRA; the algorithmic-fairness debates), it helps readers engage the disagreement rather than collapse it toward one side.

---

Carrying the toolkit out

CHAPTER RECAP — you now have: - The bias-mapping methodology as a complete framework with the eight categories defined operationally and the worked example demonstrating how triangulation across categories produces more reliable understanding than relying on any single category. - The seven working questions consolidated from the contested-questions chapters, with the demonstration that the chapter-specific working tests (training data, environment, sovereignty, privacy, copyright, deepfakes, ethics, labour, transparency) are all specific applications of the general framework. - The through-line analytical moves explicit as tools rather than as embedded arguments: the three sovereignty layers, Jevons' paradox, the epistemology-not-ethics framing, the four-category actor map, the convergent transparency precondition, the commissioned-research-firm pattern, the structural-bias frame. - Three worked examples (corporate sustainability claim, government job-creation projection, civil society surveillance report) demonstrating the toolkit in action across different source types and different policy domains. - The toolkit's limits named honestly: what it can do (recognize leans, apply systematic questions, identify patterns), what it cannot do (resolve genuine disagreements, substitute for domain expertise), and what it asks of readers (time and engagement, not just acceptance).

The next chapter (Chapter 20) takes everything the guide has developed and engages specific policy mechanisms for closing the gaps the guide has documented. The institutional pathways from Chapter 17, the convergent transparency precondition from Chapter 18, the specific gap documentation from Chapters 8 through 15 — Chapter 20 returns to these with the question of what could actually be done, by whom, through what mechanism, with what political conditions required.

You can now read AI claims with portable analytical equipment that travels with you beyond the guide. The toolkit is not a substitute for thinking; it is an infrastructure for thinking that the guide has been building chapter by chapter. The infrastructure works because the underlying analytical moves work — and the underlying moves work because the guide has tested them across multiple specific domains, demonstrated their applicability across multiple kinds of claims, and refined them through the worked examples the contested-questions chapters provided. The toolkit is the guide's primary contribution to public AI literacy in Canada. Carry it with you.

---

Bias label for this chapter: pedagogical toolkit assembly, with explicit commitment to portability over comprehensiveness. Author lean: methodological commitment to refusing premature resolution of genuine intellectual disagreements; sympathetic to readers' need for portable tools that work beyond the specific cases the manual has engaged; willing to name the toolkit's limits honestly rather than claim comprehensive coverage; explicit that the manual's value is in making analytical frameworks portable rather than in providing domain-specific expertise the manual cannot offer. The chapter draws on every previous chapter for source material; specific source citations appear in those chapters and in the Sources appendix.

---