Plain-language definitions for the AI, data-centre, and digital-sovereignty terms used across the guide, each one linked to a source you can check.
AGI (Artificial General Intelligence)
A hypothesized future AI matching or exceeding human performance across most cognitive tasks. Whether it's achievable, when, and what "intelligence" even means in that sentence are all contested — the last one more than the field usually admits.
AIDA (Artificial Intelligence and Data Act)
Canada's attempted comprehensive AI law, Part 3 of Bill C-27. Died with the January 2025 prorogation. Everything in this book about "the regulatory vacuum" is the AIDA-shaped hole.
Algorithmic management
Running workers by software: AI-driven dispatch, scoring, scheduling, discipline. The boss is a dashboard; the appeal process is mostly vibes.
Alignment
Fine-tuning aimed at making a model's behaviour match human values — which immediately raises the question whose. Not solved; jailbreaks are the proof.
Attention
The transformer mechanism that weighs every previous token when predicting the next one. The reason your chatbot remembers the pronoun from three paragraphs ago — and the reason the chips run hot.
Behind-the-meter generation
A power plant built for one customer, bypassing the public grid — the architecture proposed for Wonder Valley, among others. As a category, faster to permit and harder for outside observers to track than grid-connected generation.
Bias (three senses)
Statistical: systematic mathematical error. Social: systematic favouritism for or against groups of people. Structural: a system reproducing existing power and resource distributions even when the math is "fair." Conversations that slide between these without flagging the switch are how everyone ends up shouting.
Bias-mapping
This book's core method: every source leans; the lean is nameable; naming it is not the same as dismissing the source. The goblin's whole personality, systematized.
CARE Principles
Collective benefit, Authority to control, Responsibility, Ethics — the Global Indigenous Data Alliance's framework for Indigenous data governance, built as a counterweight to "open data" defaults.
CLOUD Act
US law (2018) compelling US-headquartered companies to produce data they hold anywhere in the world under US legal process. The reason "the server is in Montréal" doesn't end the sovereignty conversation.
Colocation
Renting space, power, and cooling in someone else's data centre. Much of "Canadian" cloud capacity physically lives here.
Commissioned research
Studies produced by professional research firms for the entity being studied. Not corruption; not independent measurement either. Check who paid for the cake.
Common Crawl
A nonprofit archive of the scraped web since 2008; the starting flour in most foundation-model recipes. Contains the best technical documentation on Earth and also every forum argument you've ever regretted.
Deepfake
Synthetic media depicting a real, identifiable person doing or saying something they didn't. Bill C-16, tabled December 2025, would extend Canadian criminal law to the sexualized kind; most of the rest is legally on its own.
Embodied carbon
The emissions baked into making the hardware — mining, fabbing, shipping — before the first prompt runs. Routinely excluded from per-prompt math, which is exactly why the goblin asks about scope.
Emergence
Capabilities that show up at large scale that smaller models lack — or possibly an artifact of how we measure. Contested, like most exciting words.
Fine-tuning
The cheaper second stage of training that shapes a base model's behaviour for actual products. Pre-training decides what the model knows; fine-tuning decides how it acts in front of customers.
Foundation model
A large model trained on broad data that other applications build on — the substrate layer of the current AI economy, and the layer with the least training-data disclosure, per every edition of the FMTI.
Freely accessible vs. freely usable
The distinction the phrase "publicly available" is designed to blur. Readable is not licensable. Your car is publicly visible.
FMTI (Foundation Model Transparency Index)
Stanford's annual scoring of major developers' disclosure across 100 indicators. Measures what companies say, not whether it's true — a feature to keep in mind in both directions.
Hallucination
The field's word for a model confidently producing falsehoods. A revealing word choice: it implies a system that normally perceives reality and occasionally slips, when the system never perceives reality at all and frequently happens to match it.
Hyperscaler
The handful of giants (AWS, Microsoft, Google, and peers) operating cloud infrastructure at global scale — including most of the capacity "Canadian" AI actually runs on.
Indigenous data sovereignty
The principle that Indigenous peoples retain ownership and governance of data about themselves, their communities, lands, and knowledge — collective rights, organized at the level of nations, predating the AI debate by decades. See OCAP, NISR, CARE.
Inference
What happens each time you use a trained model. Tiny per query; multiplied by billions of queries a day. The per-query number and the total head in opposite directions — see Jevons.
Inferential extraction
What systems deduce about you from data that seemed harmless on its own. The data points were consented to; the introductions weren't.
Jevons' paradox
Efficiency gains that increase total consumption, because cheaper-per-use means used-everywhere. Watt's engine was the efficient one; Britain burned more coal. The single most load-bearing analytical tool in this book's environmental chapters.
Law 25
Quebec's modernized privacy law — automated-decision rights, GDPR-scale penalties, the strongest framework in Canada and the standing rebuke to everywhere else in Canada.
Likeness rights / NIL
A legal claim over your own name, image, likeness, and voice, enforceable against unauthorized use — the single mechanism where performer unions and deepfake-harm researchers converge from opposite directions.
LLM (Large Language Model)
A transformer trained on enormous text corpora to predict the next token — which, scaled up, produces something that looks impressively like understanding and mechanically isn't.
Model collapse
Degradation that can occur when models train on the outputs of earlier models — the photocopier photocopying photocopies.
NISR (National Inuit Strategy on Research)
Inuit Tapiriit Kanatami's 2018 framework for Inuit governance of research, including control over data. Predates the generative-AI moment; applies to it anyway.
OCAP® Principles
First Nations Ownership, Control, Access, and Possession of First Nations data — FNIGC's framework, est. 1998. The oldest and most operationally specific of the Indigenous data sovereignty frameworks. OCAP® is a registered trademark of the First Nations Information Governance Centre.
Parameter
One of the millions-to-trillions of adjustable numbers inside a model. "The model learned" means "the parameters moved."
PIPEDA
Canada's federal private-sector privacy law, vintage 2000 — old enough to have never met a smartphone, now refereeing foundation models.
Pre-training
The expensive first stage: building a base model from massive data. Establishes what the model knows, with no opinions yet about behaving in public.
PUE (Power Usage Effectiveness)
Total facility energy ÷ computing energy. 1.0 is perfect; the global average is ~1.5; only 22% of Canadian facilities will tell you theirs.
RLHF (Reinforcement Learning from Human Feedback)
Training a model to produce outputs humans rate highly. How chatbots got polite — and where a company's values quietly enter the product.
Sovereign cloud
Marketing term for Canadian-jurisdiction hosting. Answers whose law at the contract layer; says nothing about whose chips, whose capital, or whose data governance. See: maple leaf, invoice.
Synthetic data
Training data generated by other AI systems. Scales nicely; see model collapse for the catch.
TDM (Text and Data Mining) exception
A copyright carve-out permitting computational analysis of protected works — the centrepiece of the Geist/Craig position, the EU model, and the creator organizations' objections, all at once.
Token
The unit a language model actually reads and writes — word fragments, roughly. You think in ideas; the model bills in tokens.
Training
The energy-intensive process of building a model from data, as distinct from inference (using it). Conflating the two is the original sin of most AI environmental arguments, in both directions.
Transformer
The 2017 architecture (Vaswani et al., "Attention Is All You Need") under every current major model. One of its eight co-authors went home to Toronto and co-founded Cohere — Canadian content, certified.
Voluntary Code of Conduct
ISED's 2023 framework for responsible generative AI. Voluntary, as advertised: no enforcement, no penalties, and Microsoft had not signed as of June 2026. Real as a signal; not law.
WUE (Water Usage Effectiveness)
Litres of water per kWh of computing. Reported even less often than PUE, which takes doing.
C-16 (Protecting Victims Act)
Federal bill amending the Criminal Code, including provisions that extend the non-consensual-intimate-image offence to synthetic (AI-generated) sexual images. Introduced December 2025; received Royal Assent June 18, 2026, with most provisions in force July 18, 2026.
C-18 (Online News Act)
The 2023 law requiring large platforms to pay for the Canadian news they carry. Meta answered by blocking news; Google negotiated an exemption instead.
C-27 (Digital Charter Implementation Act)
The omnibus bill that carried AIDA plus a privacy-law overhaul. Died on the January 2025 prorogation, taking AIDA with it.
C-11 (Online Streaming Act)
The 2023 law bringing streaming platforms under the Broadcasting Act and CRTC oversight, Canadian-content rules included.
C-63 (Online Harms Act)
The Trudeau-era attempt at a comprehensive online-harms regime — a regulator, takedown duties, broad content categories. Died in 2025; the narrower C-16 is the successor on deepfakes.
CRTC (Canadian Radio-television and Telecommunications Commission)
The federal broadcasting and telecom regulator. Administers the Online News Act exemption and the streaming rules; one of the sectoral regulators doing AI-adjacent work without an AI-specific mandate.
OPC (Office of the Privacy Commissioner of Canada)
The federal privacy watchdog. Investigates, reports, and recommends — but under PIPEDA it largely cannot fine, which is the recurring "regulated but not enforceable" problem.
FNIGC (First Nations Information Governance Centre)
The First Nations organization that stewards OCAP® and First Nations data governance in Canada.
OSFI (Office of the Superintendent of Financial Institutions)
Canada's banking and insurance regulator — and quietly one of the first federal bodies to put AI model risk in writing, because money concentrates the mind.
ISED (Innovation, Science and Economic Development Canada)
The federal department that owns most of the AI-strategy file — the Pan-Canadian AI Strategy, the Voluntary Code of Conduct, and the AI for All launch.
Treasury Board of Canada Secretariat
The central agency that sets rules for the federal public service, including the Directive on Automated Decision-Making and the federal AI register.
Statistics Canada
The national statistical agency. Its surveys are the load-bearing source behind this book's AI-adoption and news-revenue numbers — government-operational data, lower lean, but watch the definitions.
Mila
The Quebec AI institute founded by Yoshua Bengio — one of Canada's three federally funded AI institutes, and one of the largest academic deep-learning communities anywhere.
Vector Institute
Toronto's AI research institute, associated with Geoffrey Hinton; one of the three national AI institutes under the Pan-Canadian strategy.
Amii (Alberta Machine Intelligence Institute)
The Edmonton-based AI institute, the third of Canada's federally funded trio, with deep reinforcement-learning roots.
CIFAR (Canadian Institute for Advanced Research)
The research organization that runs the Pan-Canadian AI Strategy on Ottawa's behalf and convenes much of the country's AI talent. The strategy's money is federal; CIFAR is the hand it flows through.
Cohere
The Toronto-founded foundation-model company — Canada's clearest "national champion" in frontier AI, co-founded by one of the Transformer paper's authors.
ITK (Inuit Tapiriit Kanatami)
The national representative organization for Inuit in Canada; author of the National Inuit Strategy on Research (NISR).
GIDA (Global Indigenous Data Alliance)
The international network that stewards the CARE Principles for Indigenous data governance.
ACTRA (Alliance of Canadian Cinema, Television and Radio Artists)
The union for English-language performers in Canada; an early mover on AI likeness and consent provisions in its agreements.
DGC (Directors Guild of Canada)
The union representing directors and other screen-craft workers; active in the parliamentary AI-and-creators debate.
CUPE (Canadian Union of Public Employees)
Canada's largest union; its Senate brief is the book's go-to for concrete worker-side AI-governance asks.
Unifor
Canada's largest private-sector union; its release counts are the sourced basis for the media-layoff figures in the news chapter.
CJC (Canadian Journalism Collective)
The non-profit that splits Google's ~$100M-a-year Online News Act payment among Canadian outlets — the cheque the law actually produced, after Meta walked.
CDMRN (Canadian Digital Media Research Network)
The research network that documented AI-impersonation and synthetic-news operations during the 2025 federal election.
DFRLab (Digital Forensic Research Lab)
The Atlantic Council unit whose analysis counted the "AI slop" YouTube network posing as Canadian news.
Duty to consult
The Crown's constitutional obligation to consult, and where appropriate accommodate, Indigenous peoples when a decision may affect their rights — the legal hinge in the data-centre consultation disputes.
Royal assent
The final step at which a bill passed by Parliament becomes law. "Tabled" or "passed third reading" is not the same as "in force" — a distinction this book keeps insisting on.
Prorogation
The formal ending of a parliamentary session. Bills still on the order paper die and must be reintroduced — which is exactly how AIDA and the Online Harms Act vanished in January 2025.
Fair dealing
Canada's copyright exception allowing limited use of protected works for purposes like research, criticism, and education. Narrower than US "fair use," and central to the text-and-data-mining fight.
Mark Carney
Prime Minister of Canada (as of 2026); launched the AI for All strategy. Roles rotate — treat the title as a snapshot.
Yoshua Bengio
Montreal deep-learning pioneer, Turing Award laureate, Mila founder, and chair of the International AI Safety Report.
Geoffrey Hinton
The "godfather of deep learning," associated with Toronto and the Vector Institute; 2018 Turing Award and 2024 Nobel laureate.
Philippe Dufresne
Privacy Commissioner of Canada (as of 2026), leading the OPC's AI-and-privacy work.