Stanford's Foundation Model Transparency Index scored 13 developers against 100 indicators in December 2025. The average was 40.69 out of 100, compared with 58 in the May 2024 edition. Stanford changed the criteria and the group of developers. The two results give us a place to start looking for gaps, with each score read against the edition that produced it. > GOBLIN FACTS — check the ruler. Stanford used 55 entirely new indicators in 2025. Only nine stayed the same. Training data remained an area with major disclosure gaps. The Treasury Board of Canada launched its public AI Register, giving readers a way to see some government deployments. Google’s August 2025 inference paper described how the company measured environmental impacts and said the results had not been independently verified. Data-centre reporting presents a different question again: which measurements are collected, who sees them and what can the public compare? The gaps look different depending on where you stand. A company can publish a carefully scoped result. A regulator can collect a number that never reaches the public. In Canada, privacy commissioners, the Auditor General, researchers and reporters do some of the work of checking. What each can see depends on who must disclose what, and whether anyone can verify it. This chapter uses the Stanford index to locate disclosure gaps, then follows selective disclosure through data centres, government systems, foundation models and automated decisions. It compares Canadian institutions with EU reporting requirements and ends with a working test readers can use on a transparency claim. ---
Why transparency matters in practice
Lire cette section en français
Finding a flaw and shipping a fix leave different paper trails. In a controlled Linux study published on November 4, 2026, OpenAI reported 948 working local privilege-escalation exploits in one run and about $1.2 million in API credits. Its pipeline also produced candidate patches for human review. Mozilla’s October 27 advisory identifies vulnerabilities fixed in Firefox 159 and credits OpenAI Codex Security for some reports. The Linux study gives a count from a specific test environment. The Firefox advisory tells readers which version contains the fixes. To know whether a particular system is protected, ask one more question. Has the update been installed?
OpenAI’s September 2026 Navier–Stokes announcement came with a paper and a formalization in Lean, a tool that checks proofs written in a precise language. The company reported a proof that a smooth three-dimensional flow, in its mathematical setup, can develop a singularity in finite time. The Clay Mathematics Institute said the problem appeared settled and pointed to its formal prize process. A headline tells you what was claimed. The proof lets you ask what was actually proved. The prize has its own rules about publication and acceptance by mathematicians.
Most discussions of AI transparency frame the question in democratic-process terms: citizens have a right to know how AI affects them, transparency enables informed public debate, accountability requires visibility. These are legitimate framings, but they are not the ones that drive most working transparency requirements. The practical framings are sharper.
Transparency is the precondition for most other AI policy work. Bias audits require knowing what training data the system used. Environmental analysis requires knowing what compute, water, and electricity it consumes. Copyright determinations require knowing what protected works were in the training set. Worker protections and civil-rights enforcement both require knowing when AI is being used in employment decisions and other significant decisions about individuals — and procurement oversight requires knowing what systems vendors are actually deploying in the first place. Without disclosure, the entire downstream apparatus of AI accountability cannot function.
These disclosure gaps change which questions people can answer. A list of training sources helps a creator investigate possible use of a work. Site-level records help neighbours check electricity demand and local water use. Where those records are incomplete, a modelled estimate or a corporate report can still answer part of the question. Ask which part.
Transparency is also the layer where the gap between corporate framing and practice is most easily observed. A corporation that publishes an AI ethics framework but does not disclose how the framework is put into practice has produced an unfalsifiable claim. The framework cannot be tested against what actually happens, because what actually happens is not visible. The guide's bias-mapping methodology treats this structurally: corporate ethics frameworks are corporate self-presentation that requires triangulation against independent evidence to be meaningful.
The chapter's frame on transparency, then: transparency is the operational infrastructure that makes everything else in AI policy possible, not a stand-alone value or a vague democratic good. Where transparency is weak, accountability is weak. Where transparency is strong, accountability becomes possible (though not automatic). The chapter examines transparency at the level it actually works: disclosure regimes, verification mechanisms, audit capacity, and the structural patterns of what gets disclosed and what doesn't.
The Stanford FMTI — the empirical baseline for industry transparency
Lire cette section en français
Source: Stanford’s December 2025 Foundation Model Transparency Index
The Stanford Center for Research on Foundation Models (CRFM) launched the Foundation Model Transparency Index in October 2023, with the explicit goal of providing systematic measurement of transparency across the foundation model industry. The Index has been published annually since, with the most recent edition (December 2025) scoring 13 major developers across 100 indicators organized into three domains: upstream (training data, labour, compute), model (capabilities, evaluations, mitigations, limitations), and downstream (distribution, usage policies, feedback, impact).
The 2025 Index's headline findings, in their own framing:
The edition averages. The May 2024 edition averaged 58 out of 100 across 14 developers. The December 2025 edition averaged 40.69 across 13. Read the roughly 17-point difference alongside the changes in indicators, companies and how reports were prepared.
Individual scores. Under the respective editions, OpenAI scored 49 then 35, Meta 60 then 31, and Mistral 55 then 18. IBM reached 95 in 2025. These are changes in scores. To see which disclosures changed, compare the underlying company reports and the criteria used for each item.
The training-data gaps. The 2025 study found limited disclosure about data acquisition and properties. It also flagged basic gaps in model information. Check the particular indicator and developer before drawing a conclusion about a dataset or model.
Who submitted a report. In May 2024, 14 of 19 invited developers submitted a report. In December 2025, 7 of 23 did, roughly 74% versus 30% of those contacted. Among the nine companies scored in both editions, self-submission went from nine to seven. The Stanford team prepared six of the 13 reports scored in 2025.
🧌 GOBLIN CHECK — Seven developers handed in their own report in 2025. The researchers wrote six more. When the goblin opens one, the first thing it checks is whose name is on the work.
What changed in participation. Fewer invited developers filed their own reports. The submission counts tell us who furnished a report for this exercise. To judge changes in public disclosure or a company's reasons for taking part, we need other evidence.
What the Index cannot capture. The Index measures disclosure: what the companies have published or made available. It does not measure whether what is disclosed is accurate. The Index does not perform independent verification; it assesses what the companies say about their AI rather than independently auditing it. A perfect 100 on the Index would mean the developer has published comprehensive information; it would not necessarily mean the information published is accurate. This is a methodological feature, not a flaw: independent verification at the scale of the entire foundation model industry would be impossible without substantial new regulatory infrastructure. The Index does what it can do.
In 2025, 55 indicators were entirely new and only nine were unchanged. Stanford also changed the group of developers and collected evidence for companies that did not submit a report. A year-to-year disclosure comparison needs a consistent set of companies, questions and evidence rules.
ALIGNMENT — disclosed, or verified? A figure a company publishes about itself and a figure an outsider checked are not the same kind of fact, even printed in the same font. Most AI transparency is the first kind: homework the student graded. Before you lean on a transparency number, ask who, if anyone, was allowed to audit it.
The Canadian connection. Cohere does not appear among the 13 developers scored in December 2025. To assess its transparency, go to the company's own disclosures and look for independent evidence that can check them.
The bottom line. The 2025 edition documents substantial gaps in what readers can learn about major models. Follow those gaps back to specific questions about the data, the systems and the evidence. The published scores and the rate of submitting a report each describe one part of that picture.
The corporate selective-disclosure pattern
Lire cette section en français
Beyond the aggregate trend the Stanford FMTI documents, a specific pattern is worth foregrounding because it shapes how readers should interpret corporate AI disclosures generally: companies disclose where the numbers favour them and stay opaque where the numbers don't.
Three documented examples illustrate the pattern.
Google's August 2025 AI inference paper. Discussed in Chapter 8 as the centerpiece corporate environmental disclosure, the paper provides unusually detailed methodology for measuring energy, carbon, and water consumption per Gemini text prompt, finding that the median text prompt uses approximately 0.24 watt-hours of energy, 0.03 grams of CO₂-equivalent, and 0.26 millilitres of water. The methodology is sophisticated; the senior authors (Amin Vahdat and Jeff Dean) are credible engineers. The paper is also a worked example of selective disclosure. The median is the friendliest possible measurement (heavy users, long prompts, image generation, video generation, and agentic workloads consume orders of magnitude more). The scope is inference only (training and embodied carbon are excluded). The comparison ("less than nine seconds of TV") is friendly framing. The verification status is explicit: footnote 2 of Google's blog post announcing the paper states "The data and claims have not been verified by an independent third-party", a disclaimer the technical paper itself does not carry.
The paper is methodologically real and the disclosure is meaningful. It is also the kind of disclosure that selectively measures the favourable subset, excludes the unfavourable subsets, frames the result in friendly comparison, and operates without verification. Reading it as comprehensive AI environmental measurement would be a mistake: it measures a specific narrow scope carefully, and the scope was chosen strategically.
EXAMPLE — the label with the sugar left off. A company can print the calories, the protein, the fibre, and simply leave off the sugar. Every number is accurate; the label still misleads. Corporate AI disclosure can work the same way — the figures you're shown are real and chosen, and what's missing from the label is often the part you came to check.
Meta's open-weight disclosure pattern. Meta releases the model weights for its Llama series, a substantial transparency commitment that the closed-weights developers (OpenAI, Anthropic, Google) do not match. Meta is also one of the lowest-scoring developers on the Stanford FMTI's overall index, because while Meta discloses weights extensively, it discloses very little about training data composition, training labour conditions, or downstream usage impact. Open weights does not equal transparency. Meta has made a specific transparency commitment that flatters open-source narrative while remaining opaque on dimensions that would matter for copyright litigation (training data) and labour analysis (data labelling and content moderation).
The pattern is not unique to Meta. It is structurally common: companies disclose extensively in the dimensions where disclosure produces commercial or reputational benefit, and selectively or not at all in dimensions where disclosure produces legal risk or competitive disadvantage. Reading any single corporate transparency commitment as comprehensive transparency requires examining what is and isn't disclosed, not just what the framing claims.
Anthropic's safety-and-alignment disclosure. Anthropic publishes substantial documentation of its safety-and-alignment work: Constitutional AI methodology, Responsible Scaling Policy, red-team findings, evaluation results. This is real transparency in a domain where most competitors disclose less, and it is also selective. The training data composition is not publicly disclosed in detail; the commercial customer base and revenue model are not disclosed; the corporate governance is partly visible (public-benefit corporation status, board composition) and partly not (investor relationships, the trade-offs made in acquisition or major-investment discussions).
The pattern repeats the others: extensive disclosure in safety domains, where it produces reputational benefit and informs public debate, and selective disclosure in commercial domains, where it could affect competitive position or legal exposure. The disclosure is selective rather than dishonest, patterned in ways the framing does not always foreground.
Put together. Corporate AI transparency in 2026 is a patchwork — a patterned set of disclosures, each chosen for its own reasons, often honest within its scope, and leaving substantial dimensions undisclosed. The guide's bias-mapping methodology asks readers to track what is disclosed and what is not, what is verified and what is not, what scope is chosen and why. Where verification is absent and scope is strategically narrow, even technically honest disclosure can produce misleading aggregate impressions.
Canadian AI transparency, layer by layer
Lire cette section en français
Canada’s privacy commissioner told the federal AI transparency consultation to make information understandable and useful when people need it. Imagine finding out, after the fact, that software helped decide whether you got a service. A notice buried in the privacy policy may have been available the whole time. It still leaves you asking what the tool did and who can explain the decision. The commissioner’s submission is advice to government. The test it offers is a practical one. Could you use the disclosure to do anything?
The chapter now maps Canadian AI transparency systematically across the layers the guide has documented in earlier chapters. The picture is fragmented, with substantial gaps relative to peer jurisdictions and substantial real institutional work happening within those gaps.
The data centre / infrastructure layer. Chapter 6 separates NRCan’s Canadian market estimate from Uptime’s 2023 survey of corporate sustainability reporting. Neither establishes the share of Canadian sites publishing PUE or water figures. The EU collects annual data from covered sites and releases aggregates. Canada’s September 2026 development principles ask proponents to measure and report water use and communicate local impacts. They do not set up an equivalent annual collection or public database.
What a reader can see by facility. The public sources reviewed here do not provide a comprehensive, comparable Canadian inventory of electricity use, water inputs and efficiency metrics by site. A utility’s connection decision, an operator’s sustainability report and a measured annual water total answer different questions. Chapter 20 asks what public facility-level reporting could add.
The foundation model layer. As documented in Chapter 11 and across this chapter: the largest foundation model developers operate with substantial opacity on training data, declining engagement with transparency benchmarks, and selective disclosure in scopes that flatter results. Canada has no foundation model transparency law. The European Union AI Act's general-purpose AI provisions establish disclosure requirements for foundation models deployed in the EU market; Canadian foundation model developers (Cohere) selling internationally must comply with EU requirements but have no equivalent Canadian obligation.
The government disclosure layer. The Treasury Board AI Register launched in November 2025 with more than 400 systems reported by 42 institutions, from research through deployment. The launch scope excludes AI embedded in low-risk commercial products and covers federal institutions. For security-related or contractor-operated tools, check the individual entry and relevant procurement or oversight records before drawing conclusions from a gap.
The algorithmic decision layer. Under Québec's private-sector law, a person can get notice and request information about a decision based exclusively on automated processing of their personal information. They can submit observations to someone who can review it. Ontario's employment standards require disclosure of AI used in screening, assessment or selection in covered publicly advertised job postings. These protections apply in specific settings. Check the decision, the organization and the rule together.
The workplace AI surveillance layer. As documented in Chapter 10 and Chapter 16: substantial workplace AI surveillance (keystroke monitoring, communication monitoring, biometric tracking, emotion recognition) is deployed in Canadian workplaces without specific disclosure to affected workers, without independent audit, and without specific regulatory requirements for either disclosure or audit. The CUPE Senate brief's recommendation 4 (workplace AI surveillance prohibitions and audit requirements) explicitly addresses this gap. The federal AI for All strategy does not commit to closing it.
The AI environmental impact layer. As documented in Chapter 8: corporate AI environmental disclosures (Google's August 2025 paper as the most-detailed example) are selective, unverified by independent third parties, and operate within scope choices that flatter results. Canadian AI providers (Telus, Bell, Cohere) make AI environmental claims that are corporate self-disclosure without third-party verification. The EU AI Act's environmental disclosure requirements for high-risk AI provide a comparator; Canada has no equivalent.
The Indigenous data sovereignty layer. As documented in Chapter 9: federal AI deployment does not generally engage OCAP, NISR, or CARE Principles. The transparency dimension of this gap is direct — when AI is trained on or processes Indigenous data, there is no current Canadian framework requiring engagement with Indigenous data sovereignty principles, and no transparency framework requiring disclosure of when Indigenous data is included in training sets or processed by deployed AI systems.
The cumulative finding. Canadian AI transparency operates at substantially lower levels than peer jurisdictions across multiple layers. The pattern is consistent: voluntary or sectoral frameworks where peer jurisdictions have mandatory comprehensive frameworks, scope limitations that exclude the categories where transparency would matter most, and absence of mandatory third-party verification.
The EU comparator and what it would mean for Canada
Lire cette section en français
The European Union has been steadily building the most comprehensive AI transparency framework currently in operation. The chapter engages it directly because it represents the working alternative to Canada's current approach.
The EU AI Act transparency provisions (relevant subset of the broader Act, which covers risk classification, mandatory conformity assessment, and other requirements):
General-purpose AI model providers (foundation model developers) must publish detailed documentation about training data composition (without violating trade secrets or competitive sensitivity), compute used for training, methodologies for testing and evaluation, energy consumption during training, and impact assessment results. The disclosure is required regardless of where the provider is headquartered if the model is placed on the EU market.
High-risk AI deployers must publish documentation about the AI system's intended purpose, the system's logic and the main parameters used, the training and validation data and methods, and the system's accuracy and limitations. Affected individuals have rights to know when they have been subject to high-risk AI and to receive explanation.
General transparency provisions require that users of AI systems are notified that they are interacting with AI in many contexts, that AI-generated content be labelled in many contexts, and that deepfakes generally be disclosed as artificially generated.
The European AI Office coordinates implementation and enforcement, with substantial powers for investigation, requesting documentation, and penalizing non-compliance. Fines can reach 7% of global annual turnover (or €35 million) for the most serious violations.
The EU’s recast Energy Efficiency Directive (2023/1791) and Delegated Regulation (EU) 2024/1364 require annual submissions from operators of covered centres with installed IT power demand of at least 500 kW. They report inputs such as total site and IT energy, total and potable water inputs, and waste heat reused. Annex III defines the derived indicators, including PUE and water usage effectiveness (WUE). Under Article 5 and Annex IV, the Commission and relevant national authorities have access to individual submissions. The public database presents aggregates at Member State and EU level. Article 12 of the directive separately calls for publication of Annex VII information, subject to protections for trade secrets and confidentiality. Article 12(2) exempts certain defence and civil protection centres from that publication duty.
The General Data Protection Regulation (GDPR) provides the foundational data protection layer that the AI Act builds on, including the right to be informed about automated decision-making and the right to explanation for individuals subject to consequential automated decisions.
The Digital Services Act (DSA) provides additional transparency requirements for online platforms, including AI-related transparency for content moderation, recommender systems, and platform algorithm impact.
What this means for Canadian comparison. The EU framework is comprehensive — covering foundation models, deployed AI, data centres, automated decisions, and platform AI — with mandatory disclosure, independent verification mechanisms, central coordination, and substantial enforcement penalties. Canada's framework, by contrast, is fragmented, mostly voluntary, lacking central coordination, and with limited enforcement penalties. The two frameworks reflect fundamentally different choices about how AI deployment should be governed.
The EU framework has critics. It has been characterized as overly burdensome by industry submitters, as inadequate by civil society in some dimensions, as inconsistent in implementation across member states, and as constraining EU competitiveness against the US and Chinese AI sectors. The critiques are real and the framework is not perfect. It is also the most-comprehensive AI transparency framework currently in force, and the comparison is empirically illuminating regardless of how readers evaluate the trade-offs.
The Canadian alternative path. Canada could choose to adopt EU-equivalent transparency requirements — through federal legislation, federal-provincial coordination agreements, or sectoral regulatory implementation. The federalist constraint from Chapter 17 limits direct federal authority over much of the AI deployment that the EU AI Act regulates, but federal-provincial coordination could achieve comparable coverage. The political will to pursue this path has not emerged in Canada to date. Whether it could emerge is a contested political question; that the EU framework provides the comparator is empirical.
The institutional resources Canada has
Lire cette section en français
The chapter has documented Canada's transparency gaps relative to peer jurisdictions and to best practice. The chapter must also engage the institutional resources Canada does have, which are substantial even within the broader gap.
The Office of the Privacy Commissioner of Canada is one of the most-effective AI-policy-relevant institutions in Canada. The Office has issued substantive AI guidance interpreting PIPEDA for AI applications, conducted high-profile investigations (the joint Clearview AI investigation with provincial counterparts being the most-documented), engaged extensively in international privacy coordination, and produced substantial policy advice on AI privacy matters. The Privacy Commissioner Philippe Dufresne has been one of the most-public Canadian voices calling for substantial PIPEDA modernization including AI-specific provisions. The institutional capacity is real; the statutory framework constraining the institution is dated.
Provincial privacy commissioners have done parallel work within their respective jurisdictions. The Quebec Commission d'accès à l'information has substantial enforcement authority under Law 25 and is the most-empowered Canadian privacy regulator. The BC Office of the Information and Privacy Commissioner, the Alberta Information and Privacy Commissioner, and the Manitoba Ombudsman (which includes privacy oversight) each have meaningful AI-relevant capacity. Ontario's Information and Privacy Commissioner is an independent officer of the Legislature with real authority over the public sector and health privacy, and has issued substantive AI guidance. What Ontario lacks is a general private-sector privacy law of its own, which leaves the private sector in Canada's most populous province to the 2000-vintage federal PIPEDA.
The Auditor General of Canada has begun examining federal AI use systematically. The Office of the Auditor General's reports include findings on AI deployment in specific federal departments, and the institutional model of independent audit of federal operations is well-established and respected. Expanded AG focus on AI specifically would be one institutional pathway for increased federal AI transparency without requiring new legislation, since the AG's mandate already authorizes the analysis.
Statistics Canada's AI and Technology Measurement Program, funded at $25M over six years through the 2025 federal budget, is the primary federal source of empirical data on Canadian AI deployment and labour-market effects. The funding is modest relative to the analytical scope but represents substantive institutional commitment. StatCan's AI-related work has been the empirical foundation for several findings the guide cites (the 31% labour displacement finding from Chapter 16, the AI adoption rate trajectory from Chapter 5, the workforce composition analysis).
The judiciary. Canadian courts are increasingly receiving AI-related cases. The Canadian newspapers v. OpenAI case (Chapter 11) is the most-significant currently active, but the volume is growing across copyright, employment, privacy, and consumer-protection contexts. Judicial transparency mechanisms — published rulings, available court documents, precedent analysis — provide one of the few mechanisms producing systematic public information about how specific AI deployments are operating in Canada.
Academic researchers. Canadian academic AI researchers — at Mila, Vector, Amii, individual universities, and through collaborative networks — produce substantial peer-reviewed work that contributes to AI transparency in domains where corporate disclosure is limited. The Abundant Intelligences program (Chapter 1 and Chapter 15) is one example. Independent academic researchers like Sasha Luccioni (Hugging Face Montréal, foundational work on AI environmental measurement), Carys Craig and Michael Geist (copyright and AI from different positions), and Suzie Dunn (deepfakes) provide the academic-research layer that the corporate self-disclosure layer cannot.
Civil society organizations. CUPE, Unifor, ACTRA, the Writers' Union of Canada, the BC + AI coalition, the Centre for International Governance Innovation, the Canadian Civil Liberties Association, and many others produce substantive analysis and documentation that adds to the broader transparency landscape. The organizations' lean is generally toward worker, creator, civil-libertarian, or consumer protection — the lean is real and the guide's bias-mapping methodology asks readers to track it. The lean is also where these organizations contribute most to transparency: documenting what corporate disclosure does not.
Investigative journalism. Canadian and international journalism has produced substantial reporting on AI deployment patterns, corporate practices, and policy debates. The 2025 Canadian election deepfake study (Chapter 13), the Stanford FMTI annual releases, the TIME investigation of OpenAI's Sama contracting (Chapter 3), the various reports of corporate AI ethics issues, all contribute to the public transparency landscape that the institutional regulatory framework does not provide.
The institutional balance sheet. Canada has substantial institutional resources for AI transparency work — multiple regulators with real expertise, a judiciary engaging AI cases, academic research capacity, civil society documentation, and investigative journalism. The resources are working within statutory and political constraints that limit their cumulative effect to substantially less than what comprehensive AI transparency would require. This is structural, not a criticism of any specific institution. Each institution is doing meaningful work within its mandate. The mandates, taken together, do not yet add up to the transparency framework that AI deployment at current scale requires.
The convergent finding — transparency is the precondition
Lire cette section en français
Try enforcing a rule when you can’t see the activity it covers. A creator needs to know what went into training. A community needs to know what a facility draws from the grid. A worker needs to know which decision the tool helped make. In each case, disclosure gives someone a place to start asking for an answer.
Working through this connection across earlier chapters:
Copyright (Chapter 11) depends on knowing what was in the training data. Without that knowledge, the rights-holder claims that are being litigated cannot be proven systematically. The case-by-case litigation that is currently happening is producing precedent but is a slow, expensive, and partial substitute for what transparency would enable: systematic identification of unauthorized uses and systematic remedy.
Environmental impact (Chapter 8) depends on knowing what compute, water, and electricity AI systems consume. Without mandatory reporting at the data centre layer and verifiable methodology at the model layer, the environmental analysis cannot be evidence-based. The corporate self-disclosure pattern produces unverifiable claims; the regulatory framework that would produce verifiable claims does not exist in Canada.
Bias and discrimination (Chapter 15) depends on knowing what training data the system used, what evaluation it underwent, and what outcomes it has produced in deployment. Without this disclosure, bias audits are limited to what can be reverse-engineered from outputs — a substantial methodological constraint. The algorithmic fairness literature engages this directly; the regulatory framework that would require disclosure does not.
Workplace AI deployment (Chapter 10 and Chapter 16) depends on knowing what AI is being used in employment decisions. Without disclosure to workers, the rights to challenge AI-driven decisions cannot be exercised. CUPE's specific recommendation 4 (notification, explanation, and review) directly addresses this transparency gap.
Indigenous data sovereignty (Chapter 9) depends on knowing when Indigenous data is being used in AI training and deployment. Without disclosure, OCAP, NISR, and CARE Principles cannot be put into practice: the Indigenous communities cannot exercise the data governance the frameworks call for if they do not know when their data is being used.
Sovereignty (Chapter 9) depends on knowing what foreign jurisdictions have what access to what Canadian AI infrastructure. The CLOUD Act analysis depends on knowing which providers operate under which jurisdictions; the supply chain analysis depends on knowing where AI hardware is designed, fabricated, and deployed; the data flow analysis depends on knowing where Canadian data is stored and processed. Each of these depends on disclosure that the current Canadian framework does not require.
Deepfakes (Chapter 13) depend on knowing when content is AI-generated. The disclosure requirements that the EU AI Act establishes for synthetic content provide the operational mechanism for the broader policy goals around deepfake regulation; Canada has not established equivalent requirements.
The institutional point. Each of the policy areas the guide has engaged depends on disclosure infrastructure that Canada has not built. The transparency gap is not separate from the other policy gaps; it is the common structural precondition for closing them. Canadian AI policy that addresses the substantive issues without addressing the transparency infrastructure is attempting to enforce rules where the underlying activity is not visible.
This is the reason the chapter argues that comprehensive Canadian AI transparency requirements (through federal legislation, federal-provincial coordination, or substantial sectoral regulatory expansion) would not just improve one dimension of AI policy. It would provide the operational infrastructure that makes the rest of AI policy possible. Chapter 20 returns to this with specific path-forward analysis on what comprehensive Canadian AI transparency requirements could look like and how they might be implemented through the institutional pathways Chapter 17 mapped.
The working test
Lire cette section en français
One last run of the standing test, transparency edition, kept short because Chapter 19 is about to assemble all of these into the single portable version anyway. For any AI transparency claim, ask: scope (what's disclosed, and what's pointedly not?); verification (third party, or self-marked homework?); audit access (can anyone outside the building examine the underlying data and methodology?); comprehensiveness (does it cover what matters, training data, environmental impact, deployment outcomes, supply chain, or is it selective in patterned ways?); comparability (formatted to be compared, or formatted to prevent it?); and operational consequence (does the disclosure let anyone actually do anything, be informed, challenge, seek remedy, or is it transparency-as-aesthetic rather than transparency-as-infrastructure?).
Why this one comes first
Lire cette section en français
CHAPTER RECAP — you now have: - Stanford's 2025 Index documents substantial gaps. Compare the published scores with their edition rules and distinguish the submission counts from disclosure itself. - The corporate selective-disclosure pattern named with documented examples — Google's August 2025 paper (sophisticated methodology, selective scope, unverified), Meta's open-weight disclosure (substantive in one dimension, opaque in others), Anthropic's safety transparency (substantial in safety, selective in commercial dimensions). - The Canadian transparency gaps across infrastructure, models, public-sector deployment, automated decisions, workplace surveillance, environmental impacts and Indigenous data governance, with the difference between regulatory collection and public access made explicit. - The EU comparator framework — AI Act transparency provisions, EED data centre reporting, GDPR foundational rights, DSA platform requirements, with the European AI Office providing coordination — as the operational alternative to Canada's current approach. - The institutional resources Canada has — OPC, provincial commissioners, Auditor General, Statistics Canada, judiciary, academic researchers, civil society organizations, investigative journalism — and the structural finding that the resources are operating within constraints that limit their cumulative effect below what comprehensive AI transparency would require. - The useful disclosure depends on the question. A creator may need to know what went into training. A community may need a facility’s electricity and water figures. A worker may need to know where a tool entered a decision. Indigenous data governance asks who can decide what is collected, used or shared. - The working test for evaluating any AI transparency claim: scope, verification, audit access, comprehensiveness, comparability, operational consequence.
Chapter 19 puts the working questions together. Keep them handy for the next claim that comes with a confident number and no obvious source.
You can now read any AI transparency claim with the structural equipment to recognize what is and isn't being disclosed, what verification is and isn't present, and what the disclosure does and doesn't enable. The transparency conversation in Canadian AI public discourse oscillates between corporate framing of voluntary commitments (which the guide treats as one input among others) and demands for comprehensive disclosure (which the guide recognizes as the operational precondition for most other policy work). The institutional infrastructure that would close the gap exists in pathways the guide has mapped; the political conditions for activating those pathways have not yet emerged.
Bias label for this chapter: institutional and structural analysis of Canadian AI transparency, with explicit comparative international framing and explicit connection to policy preconditions in earlier chapters. Author lean: skeptical of corporate self-disclosure unverified by independent third parties; sympathetic to comprehensive mandatory transparency frameworks as the operational precondition for most other AI policy work; willing to credit Canadian institutional resources within the broader gap rather than dismissing them entirely; explicit that the manual treats this as comparative empirical analysis rather than as policy advocacy for any specific regulatory framework. Independent academic measurement (Stanford FMTI, Luccioni research network) treated as primary on industry transparency patterns. Corporate self-disclosure (Google, Anthropic, Meta sustainability reports and ethics frameworks) labelled and read accordingly. Government framing (Treasury Board, ministerial statements) treated as primary on what has been committed to. Civil society and academic critique treated as primary on identified gaps.
Primary sources cited or relied on in this chapter: Stanford Center for Research on Foundation Models, 2025 Foundation Model Transparency Index (December 2025); previous editions of the Index (2023, 2024); Google Cloud Blog "Measuring the environmental impact of AI inference" (Vahdat & Dean, August 21, 2025); Meta Llama series release documentation; Anthropic Constitutional AI and Responsible Scaling Policy documentation; Treasury Board of Canada Secretariat AI Register (November 28, 2025); Natural Resources Canada Best Practice Guide for Canadian Data Centres (November 2024); Office of the Privacy Commissioner of Canada AI guidance documents and joint findings on Clearview AI (2021); European Union Artificial Intelligence Act (effective August 1, 2024); European Union Energy Efficiency Directive 2024/1364; European Union General Data Protection Regulation (GDPR); European Union Digital Services Act (effective 2024); CUPE Senate Brief (March 2026); Quebec Law 25 documentation. Detailed citations in the Sources appendix.
🧌 GOBLIN CHECK — Seven developers handed in their own report in 2025. The researchers wrote six more. When the goblin opens one, the first thing it checks is whose name is on the work.
Recap
- Stanford's 2025 Index documents substantial gaps. Compare the published scores with their edition rules and distinguish the submission counts from disclosure itself.
- The corporate selective-disclosure pattern named with documented examples — Google's August 2025 paper (sophisticated methodology, selective scope, unverified), Meta's open-weight disclosure (substantive in one dimension, opaque in others), Anthropic's safety transparency (substantial in safety, selective in commercial dimensions).
- The Canadian transparency gaps across infrastructure, models, public-sector deployment, automated decisions, workplace surveillance, environmental impacts and Indigenous data governance, with the difference between regulatory collection and public access made explicit.
- The EU comparator framework — AI Act transparency provisions, EED data centre reporting, GDPR foundational rights, DSA platform requirements, with the European AI Office providing coordination — as the operational alternative to Canada's current approach.
- The institutional resources Canada has — OPC, provincial commissioners, Auditor General, Statistics Canada, judiciary, academic researchers, civil society organizations, investigative journalism — and the structural finding that the resources are operating within constraints that limit their cumulative effect below what comprehensive AI transparency would require.
- The useful disclosure depends on the question. A creator may need to know what went into training. A community may need a facility’s electricity and water figures. A worker may need to know where a tool entered a decision. Indigenous data governance asks who can decide what is collected, used or shared.
- The working test for evaluating any AI transparency claim: scope, verification, audit access, comprehensiveness, comparability, operational consequence.
Sources
- Stanford Center for Research on Foundation Models, 2025 Foundation Model Transparency Index (December 2025)
- previous editions of the Index (2023, 2024)
- Google Cloud Blog "Measuring the environmental impact of AI inference" (Vahdat & Dean, August 21, 2025)
- Meta Llama series release documentation
- Anthropic Constitutional AI and Responsible Scaling Policy documentation
- Treasury Board of Canada Secretariat AI Register (November 28, 2025)
- Natural Resources Canada Best Practice Guide for Canadian Data Centres (November 2024)
- Office of the Privacy Commissioner of Canada AI guidance documents and joint findings on Clearview AI (2021)
- European Union Artificial Intelligence Act (effective August 1, 2024)
- European Union Energy Efficiency Directive 2024/1364
- European Union General Data Protection Regulation (GDPR)
- European Union Digital Services Act (effective 2024)
- CUPE Senate Brief (March 2026)
- Quebec Law 25 documentation.