In March 2026, the Canadian Union of Public Employees told the Senate Social Affairs Committee it was seeing job losses among closed-captioning workers, medical transcriptionists, hospital dispatch workers and long-term care food-service workers. The brief also cited a Statistics Canada exposure estimate and interpreted it as a warning about work that could change or disappear. Keep the workplace reports and the exposure estimate in separate columns. At the June 2026 launch of AI for All, the federal government projected that AI adoption could help create up to 250,000 jobs by 2031. In a May 2026 speech, Bank of Canada External Deputy Governor Michelle Alexopoulos said the evidence available then did not point to widespread worker displacement from AI. A projection and an assessment of the evidence to date have different jobs. A company can eliminate a role while contractors elsewhere get paid to label data or review model outputs. The people doing those jobs may be an ocean apart. The question this chapter follows is who loses work, who gets new work, what it pays and who gets monitored on the job. You need different evidence for each answer. Start with what the Canadian figures can tell us. Then put Acemoglu and Autor beside each other and read CUPE’s five recommendations against Ottawa’s commitments. When you get to platform work and algorithmic management, keep asking who sets the terms and who can challenge them. ---
The Canadian empirical baseline
Lire cette section en français
Read what the AI hiring studies measured
In a voluntary 2025 Ontario survey, 19% of 446 respondents to a question about AI said it had reduced paid opportunities in their cultural field. That’s how those respondents saw their field. The survey doesn’t count how many of them lost work.
Statistics Canada found 26% fewer payroll jobs in film and video in November 2025 than in November 2022. A separate survey found post-production businesses spent 14.4% less on wages, salaries and benefits in 2025 than in 2023. One measure counts jobs and the other tracks compensation spending. Neither isolates AI’s role in the decline.
A job can disappear from your future without anyone being fired. Statistics Canada found that hiring among 15- to 29-year-olds slowed across occupations with both high and low potential exposure to AI. From the six months leading up to November 2022 to the six months leading up to August 2026, the hiring rate fell from 4.0% to 3.3% in lower-exposure occupations and from 2.5% to 1.9% in coding-intensive occupations. The coding decline started earlier in 2022. A slower hiring rate tells you the door opened less often. To explain why, you need evidence about what employers actually changed.
A U.S. Census working paper looked at the same question from a different angle. It used employer and employee records for 22- to 24-year-olds, grouping jobs by industry and state. The paper reports that adjusted employment fell 12% over the ten quarters after ChatGPT’s launch in the fifth of groups most exposed to AI. Fewer hires accounted for most of that decline. Statistics Canada measured hiring and separation rates across occupations, including for 15- to 29-year-olds, and found slower hiring even in jobs with low potential exposure. The U.S. result makes hiring worth watching. Neither study measures employers’ actual AI use, and the U.S. paper finds some trends began during the pandemic. The timing alone cannot tell you how much of the slowdown AI caused.
The Stanford–ADP study follows a different slice of the labour market: U.S. payroll records for workers aged 22 to 25, grouped by occupation. Its August 2026 revision reports a 19% employment shortfall in more AI-exposed occupations relative to where employment would have been if it had grown at the rate of less-exposed peers. That does not mean 19% of those workers were fired. The gap arose mainly through fewer hires. The authors find no evidence of widespread job displacement across the economy and describe their results as early indicators, not a causal estimate. The useful question is whether the first rung of the ladder is getting harder to reach. Answering it means tracking hiring while checking what else changed.
Three distinct primary sources establish what's actually documented about AI and Canadian work as of mid-2026.
Statistics Canada's analytical work. Statistics Canada has published multiple analyses of AI-related employment effects across 2024–2026, drawing on its Survey of Digital Technology and Internet Use, the Labour Force Survey, and specific analytical papers. The headline findings cited in subsequent advocacy and government documents:
- 31% of employees aged 18–64 in the Canadian study were in occupations classified as potentially highly exposed to AI with relatively low complementarity in May 2021. Another 29% were in potentially highly exposed occupations with high complementarity, and 40% were in occupations classified as less exposed.
- The study classifies occupations by their tasks and potential for AI to complement them. It measures exposure using a May 2021 snapshot. It gives no count of people who lost a job, moved to another one or needed retraining. The roughly 60% combined exposure figure and the federal 60% business-adoption target measure different things.
- In the same May 2021 estimates, 38% of employees in the Women+ category and 23% in Men+ were in the high-exposure, low-complementarity group. These are shares of employees in an exposure category. The study includes some non-binary people in each category.
- AI adoption among Canadian businesses has moved from approximately 12.2% in Q2 2025 to 19.2% by May 2026 (StatCan's most recent data at the time of this guide's drafting). The federal AI for All target of 60% adoption by 2034 implies a tripling over approximately eight years, which is rapid but not unprecedented for digital-technology adoption curves. The doubling-in-one-year trajectory matters: Statistics Canada documented adoption rising from 6.1% in Q2 2024 to 12.2% in Q2 2025, before the further rise to 19.2% by May 2026. The pace is accelerating, not steady.
GOBLIN FACTS — adoption doubled fast. Statistics Canada recorded business AI use rising from 6.1% in Q2 2024 to 12.2% in Q2 2025, then toward 19.2% by the guide's launch window.
- The sectoral distribution of AI adoption is highly uneven. StatCan's Q2 2025 data shows adoption concentrated in specific sectors: information and cultural industries at 35.6%, professional/scientific/technical services at 31.7%, finance and insurance at 30.6% — all substantially above the 12.2% aggregate at that point. These are the same sectors with high concentrations of mid-skilled and high-skilled white-collar work, which intersects directly with the documented displacement pattern in this chapter. Among Canadian firms already using AI, 25.7% had purchased cloud services or storage and 17.9% had purchased computing power or specialized equipment. AI adoption pulls capital toward infrastructure as it spreads, connecting directly to Chapter 6's provincial allocation analysis.
The gap between what AI can do and what it is doing. A complementary picture comes from the US, where Anthropic's Economic Index pairs a measure of what large language models can theoretically do, across the tasks that make up each job, against what the models are observed actually doing in work settings. The observed use is a small fraction of the theoretical capability, and both cluster in the same white-collar categories Canada's adoption data flags: computer and math, office and admin, business and finance. The uncovered space between the two is the part worth watching, because it is what narrows as capability, adoption, and deployment catch up to one another. The data is American (US tasks and employment weights), so read it as an illustration of the pattern rather than a Canadian measurement, but the shape of the gap is the clearest single picture of where AI is and isn't yet doing the work.
The Bank of Canada’s analysis. In May 2026, Michelle Alexopoulos said the evidence did not yet point to widespread worker displacement caused by AI. She also noted job cuts that employers attributed partly to AI and weak hiring in some exposed roles. That is a dated assessment of the evidence available to the Bank.
That broad picture leaves plenty of room for individual losses. It cannot tell you what happened to a particular captioner after a job ended. Statistics Canada’s newer hiring and separation study measures rates by occupation. Its separations include quits as well as layoffs, and it does not report what those workers earned in their next jobs.
🧌 GOBLIN CHECK — The unemployment rate can stay steady through someone’s worst year. If you want to know where the captioner landed, follow the captioner. Ask how long it took to find work, what the next job paid per hour and how many hours it offered.
The sectoral reports. CUPE’s submission identifies job losses among workers in closed captioning, medical transcription, hospital dispatch and long-term care food services. The brief gives examples rather than a tally of workers or a record of their next jobs.
- Federal public service. Chief Data Officer Stephen Burt said publicly that adopting AI could mean some federal job reductions. The Treasury Board AI Register records deployments. It does not report what happened to individual workers.
To learn what followed those workplace changes, we would need employers’ records or linked worker data. A record that someone left tells us little about their next contract, their hours or their pay.
Where the evidence lands. Canadian AI labour effects in mid-2026 consist of: documented sectoral displacement in multiple categories, no aggregate unemployment signal yet, substantial projected change across roughly 60% of the workforce, gendered displacement risk distribution, and rapid adoption-curve trajectory. None of these findings supports either "AI is causing mass unemployment" or "AI labour fears are unfounded." The honest reading sits between those poles, with specific empirical questions that the next sections engage.
The Acemoglu vs Autor academic disagreement
Lire cette section en français
Two economists at MIT, both widely respected, both publishing extensively on AI and labour, both rigorous, have reached substantially different conclusions about AI's likely macroeconomic and labour-market effects. The disagreement is genuine intellectual disagreement, not factional posturing. The guide presents both positions in their own framings, the way Chapter 11 presented Geist/Craig vs TWUC/ACTRA on copyright. Reading both is the chapter's central methodological commitment.
Daron Acemoglu's position. Acemoglu, co-recipient of the 2024 Nobel Memorial Prize in Economic Sciences (with Simon Johnson and James Robinson) for work on institutional economics, has published several papers and books over 2023–2025 making the case for modest AI macroeconomic impact with significant distributional concerns.
His key empirical and analytical claims:
First: Acemoglu's most-cited recent work, "The Simple Macroeconomics of AI" (NBER Working Paper 32487, April 2024; published in Economic Policy 40(121):13–58, 2025), provides a quantitative framework estimating AI's likely macroeconomic effects. The framework's central conclusion: AI will likely produce total factor productivity gains of approximately 0.5% over a ten-year period, substantially smaller than the productivity gains projected by AI-industry advocates. The analysis is built from task-level economics — what specific tasks AI can replace, what wage levels those tasks command, what share of the economy those tasks represent.
Second: Acemoglu's broader framework (developed with Pascual Restrepo at Boston University) emphasizes that historically, factory automation has resulted in higher local unemployment, not lower. The "automation as opportunity" framing that dominates corporate and government rhetoric is, in Acemoglu's analysis, empirically wrong for the populations that have actually experienced automation waves. The local effects are typically these: mid-wage workers displaced, the productivity gains captured by capital owners and managers, and low-wage employment growing as the displaced look for substitute work.
Third: Acemoglu, with Simon Johnson, made an extended historical argument in Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity (2023) that technological progress does not automatically translate into broad-based prosperity — it has done so in specific historical periods when worker power, regulatory frameworks, and political constraints redirected technological gains broadly. Without those institutional constraints, technological progress tends to concentrate benefits narrowly.
Fourth: Acemoglu, Johnson, and Autor jointly authored "Can we have pro-worker AI?" (CEPR Policy Insight 123, 2023) — the constructive structural-reform agenda. The piece argues that AI's labour-market effects depend critically on the policy choices governments make — across taxation, worker representation, training and education, and antitrust. The same AI deployed under different institutional conditions can produce dramatically different labour-market outcomes.
Acemoglu's bottom line, in brief: AI's macroeconomic impact will probably be modest in aggregate. The distributional effects will probably be sharp, concentrating gains with capital and certain skilled workers while displacing mid-wage workers and producing growth of low-wage employment as substitute. The outcomes will depend critically on policy choices Canada has not yet made.
David Autor's position. Autor, also at MIT and the field's leading scholar on labour-market polarization, with foundational work going back to a 2003 paper with Levy and Murnane that established much of the current empirical framework, has reached substantially different conclusions in his recent work on AI specifically.
His key empirical and analytical claims:
Autor's 2024 paper "Applying AI to Rebuild Middle Class Jobs" (NBER Working Paper 32140) makes the case that generative AI is meaningfully different from prior automation waves because it can augment mid-skill expert work rather than only displacing routine work. The previous automation wave (the 1980s–2010s deployment of computers and the internet) hollowed out the middle of the wage distribution by automating routine tasks, leaving high-skill cognitive work and low-skill physical work intact. Generative AI could reverse that hollowing-out by giving mid-skill workers access to expert capabilities that previously required years of specialized training. His more recent "Expertise" (NBER Working Paper 33941, with Thompson, 2025) provides the empirical follow-through: AI tools tend to improve the performance of less-experienced workers more than more-experienced workers, narrowing within-occupation skill gaps. If this pattern holds at scale, AI could function as a productivity-equalizer rather than as a wage-polarizer. And Autor has been explicit in public commentary that the augmentation-vs-displacement question is empirically contested — the outcome depends partly on policy choices, partly on how the technology develops, and partly on how labour markets and institutions respond. He is not claiming AI will automatically augment rather than displace. His claim is narrower: that the augmentation possibility is real, and more within reach than the AI-pessimist framing suggests.
Autor's bottom line, in brief: Generative AI has real potential to reverse decades of labour-market polarization by augmenting mid-skill expert work. The potential is not guaranteed; it depends on technological development paths and policy choices. But the technology is sufficiently different from prior automation waves that the historical pessimism may not transfer cleanly.
Where Acemoglu and Autor converge. Both economists emphasize that AI labour outcomes are not determined by the technology but by the institutional and policy context in which the technology is deployed. Both have collaborated on the "pro-worker AI" framework. Both reject the most extreme positions on either end (universal mass unemployment, universal broad-based prosperity).
Where they diverge. Acemoglu's analysis weights heavily the historical evidence that automation has tended to disadvantage workers absent institutional constraints. Autor's analysis weights heavily the structural difference between generative AI and prior automation waves. They are reading the same empirical and historical evidence and drawing different conclusions about which factors will dominate.
The guide's methodological position: the disagreement is genuine and the guide leaves it standing. Both Acemoglu and Autor have credentials, institutional positions, and published analyses that command serious consideration. The guide sets both positions side by side and asks readers to sit inside the disagreement rather than pick a winner to make the chapter tidier. The disagreement itself is the structurally important feature of the academic-economics conversation about AI labour effects, and Canadian policy that doesn't engage both seriously is reducing a real intellectual debate to a political choice.
The CUPE Senate brief — labour's primary documentation
Lire cette section en français
The Canadian Union of Public Employees represents, by its own count, 800,000 Canadian workers across healthcare, education, municipalities, libraries, universities, social services, child care, communications, utilities, emergency services, transportation, and airlines. CUPE's March 2026 submission to the Senate Social Affairs Committee is an 11-page document, concentrated rather than sprawling, anchored in primary research by senior researcher Sarah Ryan. It is the most-detailed Canadian labour primary documentation on AI policy currently available, and the guide draws on it extensively for this chapter.
The brief's structural argument, in CUPE's own framing: AI deployment in Canadian workplaces is happening rapidly, the regulatory framework governing it is substantially absent, the documented harms to workers are real, and the federal AI for All strategy commits to investment in AI development without committing to corresponding worker protections. The brief is policy-positive rather than rejectionist: CUPE explicitly endorses worker-augmentation AI applications and supports policy frameworks that distribute AI's benefits broadly. The lean is toward worker protection within a framework that takes AI as a real and continuing technological development.
CUPE's five specific recommendations, presented in their own framing:
Recommendation 1: Comprehensive AI legislation. CUPE calls for federal legislation that would establish mandatory transparency for AI deployment, restrictions on workplace surveillance, data protection requirements specific to AI processing, mandatory bias audits, and the institutional framework for ongoing AI policy oversight. The recommendation is explicitly framed as filling the gap left by AIDA's failure and as separate from the AI for All investment strategy.
Recommendation 2: Education, skills, and retraining infrastructure. CUPE calls for investment in training that workers can access as their jobs change. The brief cites an estimate of annual employer spending on training of $240 per employee. The exposure estimate can help identify occupations to study. To size a program, ask which workers face changes or layoffs, what training they need and whether they can afford the time away from paid work.
Recommendation 3: Labour market planning and inclusive consultation. CUPE calls for re-establishing the Canadian tradition of sectoral partnership tables (currently functioning meaningfully only in Quebec) that bring together unions, employers, and government to plan for technological transitions. The recommendation explicitly notes that the AI and Labour Advisory Council announced by Minister Solomon after labour union pressure (Chapter 5) is a consultation mechanism rather than a substantive planning body, and calls for the latter rather than only the former.
Recommendation 4: Public sector capacity and public digital infrastructure. CUPE calls for the federal government to build publicly-owned digital and AI technology applications that can be shared across the public sector at cost, as an alternative to the current pattern of federal departments procuring AI services from US hyperscalers and Canadian commercial vendors. This is the labour-side answer to the sovereignty question discussed in Chapter 9: not just which foreign clouds, but whether Canadian public services should be on foreign commercial clouds at all.
Recommendation 5: Public procurement requirements. CUPE calls for specific procurement framework requirements: transparency requirements for AI vendors selling to government; explainability requirements for AI systems used in significant decisions; inclusive governance requirements for AI development funded with public money; rights of access and control over data; restrictions on third-party data sales; environmental and human-rights impact assessments. The recommendation establishes specific operational criteria that current Canadian government AI procurement does not generally include.
The specific legal asks that connect to other chapters:
- Ban the use of biometric, facial recognition, and emotion recognition AI systems in workplaces (also discussed in Chapter 10).
- Prohibit significant employment decisions (hiring, promotion, discipline, termination, wage-setting) from being made based on AI output (the strongest single recommendation).
- Require notification when AI was used in any decision affecting a worker, with explanation of the result, and human review and appeal mechanism.
- Mandate bias and discrimination audits both before AI deployment and annually after.
- Stipulate that companies receiving public funding for AI development cannot use that funding while cutting jobs through AI deployment.
The bias label for these positions: civil society advocacy from Canada's largest public-sector union, with substantial primary research and documented constituency interest. The lean is toward worker protection. The underlying evidence (documented surveillance practices, documented job losses, documented training-investment gaps) is independently verifiable. The lean is real; the evidence is also real. Discounting the brief because it comes from a union would be as wrong as accepting it without noticing that it does.
A specific framing from CUPE's brief that the guide carries forward as a primary-source quotation: "Technology corporations have actively lobbied against regulatory safeguards despite documented harms." This connects directly to the proxy-lobbying mechanism discussed in Chapter 7. CUPE's brief cites a November 2025 Policy Options analysis titled "U.S. Proxy Lobbying Threatens Canada's Digital Sovereignty" as the documentation of how foreign tech corporations route influence through Canadian intermediaries to shape policy. The union's framing is that worker-protection legislation has not failed for lack of ideas; it has failed for lack of political space against organized industry resistance.
Who loses work, who gets it, and what it pays
Lire cette section en français
The chapter has presented the empirical baseline, the academic-economics disagreement, and the labour-side documentation. The structural finding the guide has been building toward across multiple chapters can now be set out directly.
Most Canadian discussion of AI and jobs collapses into a displacement-vs-augmentation framing. Will AI replace jobs (displacement) or make existing workers more productive (augmentation)? The framing implicitly assumes that the AI economy is a closed system in which the work that exists is either replaced or augmented. This framing is empirically incomplete.
The evidence gives us examples of job losses in some workplaces, low-paid contracting in parts of the AI supply chain and changes in how work is monitored. It does not link the Canadian workers who lost jobs to the people taking those contracts. A stable aggregate unemployment rate cannot make that link either.
Specific documented patterns across the guide's analysis:
Training data labour (Chapter 3). AI training depends on large-scale human labelling labour, primarily concentrated in the Global South, with documented investigations (TIME 2023, Washington Post 2023) reporting pay as low as roughly US$1.32 to US$2 per hour, typically without benefits, employment-status protections, or stable hours. The Sama/Kenya contracting for OpenAI's content moderation (documented in TIME, January 2023) is the most-cited specific case. Canadian AI deployment depends on this global labour pool even though the labour itself is not in Canada.
Content moderation labour. As AI deployment expands, the human content moderation required to keep AI outputs within acceptable bounds expands with it. The labour is performed under documented conditions of psychological harm exposure (graphic and traumatic content review) and at compensation levels that don't reflect the difficulty or psychological cost of the work.
Gig economy and platform labour. Canadian Uber drivers, DoorDash couriers, Amazon Flex drivers, and similar platform workers are subject to algorithmic management: AI-driven dispatch, performance scoring, dynamic pricing, and disciplinary action. The work category has grown substantially over the past decade, generally at compensation levels below the median Canadian wage and without conventional employment protections. Algorithmic management of gig labour is one of the AI economy's largest single labour categories, and it is a structural creation of AI-enabled platforms.
Algorithmic management of conventional workplaces (Chapter 10). Within conventional employment, AI surveillance tools (keystroke monitoring, communication monitoring, performance scoring, emotion recognition) intensify the management of workers' time and behaviour. The workers keep their jobs, but the monitoring reshapes the work itself.
What can happen to earnings after a layoff. In historical Canadian data through 2016, Statistics Canada found persistent earnings losses for many permanently laid-off workers with long tenure. That is evidence about layoffs generally. Annual earnings can fall because people spend longer without work or get fewer hours, as well as because a new job pays less per hour.
Follow the worker. Humlum and Vestergaard linked surveys to Danish administrative records. They found little change in average earnings or recorded hours over the first two years after ChatGPT’s launch, alongside moves by some chatbot adopters into higher-paying occupations. That study follows adopters in Denmark. The Canadian question here is where people who lost jobs tied to AI went next and what they earned. The sources we reviewed leave that question open.
EXAMPLE — the self-checkout. It didn't fire the cashier so much as redistribute the job: the store keeps fewer staff, and you became the unpaid one scanning your own groceries. Plenty of AI rollouts work the same way. The headline job doesn't vanish; the work quietly moves onto cheaper hands, or onto yours, and the productivity number still goes up.
Count jobs and count paycheques. To test whether people moved into worse-paid work, link the same workers’ previous and subsequent jobs and compare hourly pay, hours, benefits and time out of work. Then check whether AI drove the original job loss. Those are the receipts the claim needs.
ALIGNMENT — does it add to the worker, or stand in for them? The same tool can augment a job or hollow it out, and the rollout email almost never says which. Ask whether the AI makes the person better at the work, or makes the person optional. "Productivity" is the word used for both, right up until the layoff.
The federal *AI for All* labour commitments, evaluated
Lire cette section en français
The June 2026 AI for All strategy made specific labour-related commitments. Evaluating them against the analysis the chapter has developed:
The 250,000-jobs projection. The strategy projects up to 250,000 new jobs through AI adoption by 2031. The figure is a projection from the strategy's economic modelling, not a commitment to deliver specific numbers of jobs. The projection's confidence intervals, underlying assumptions, and sectoral breakdowns are not publicly disclosed in detail. Evaluated against the chapter's analysis: even taken as accurate, the 250,000 new AI-related jobs would need to be evaluated against the displacement of jobs in other sectors. The strategy's modelling does not publicly engage the displacement question with the same specificity.
The "Empowering Canadians" pillar, the strategy's worker-facing commitments. The strategy commits to AI literacy training, retraining programs, and workforce transition support. The specific funding levels are modest relative to the scale of projected workforce transition (60% of Canadian workers, per StatCan). The training delivery is partly through the three national AI institutes (Mila, Vector, Amii), which, as discussed in Chapter 7, are federally-aligned beneficiaries of AI for All rather than independent training providers. Evaluated against the chapter's analysis: the commitment is real but the scale does not match the projected transition magnitude, and the delivery infrastructure has structural conflicts of interest that the strategy does not address.
The 90,000 youth placements commitment. The strategy commits to 90,000 youth placements over five years, primarily through partnerships with industry. Set against this chapter's analysis, the worry is familiar: youth placement programs have historically produced mixed outcomes for participants, with industry-funded placements often functioning as low-wage labour pools rather than as career-development pathways. The specific design of the 90,000 placement commitment, including wage levels and pathway-to-permanent-employment provisions, is not detailed in publicly-available strategy documentation.
The AI and Labour Advisory Council. Minister Solomon announced this body after pressure from CUPE, Unifor, and other unions, the documented case discussed in Chapter 5 where labour pressure visibly shaped federal policy. Evaluated against the chapter's analysis: the Council is a consultation mechanism, not a substantive policy body. Whether it produces binding rules or only advisory recommendations is unresolved. CUPE's third recommendation (sectoral partnership tables with substantive planning authority) calls for something more than the Council currently is. The Council can become more substantive over time, but the current commitment is consultative.
The Treasury Board AI Register and federal public service AI. The parallel federal track (Chapter 10) acknowledges that AI deployment in the federal public service is occurring, including in ways that affect federal employment. Chief Data Officer Stephen Burt's confirmation that certain federal jobs will be cut is the operational acknowledgement of the strategy's labour implications. The chapter's analysis surfaces the obvious tension here: the federal government is both a major AI deployer and the entity setting AI labour policy, a dual role that produces conflicts of interest the current framework does not resolve.
What the strategy does not include, that the chapter's analysis suggests would matter:
- A binding commitment to bias-and-discrimination audits of federal AI deployment as it affects employment decisions.
- Mandatory consultation with affected workers before federal AI deployment.
- A binding commitment to fund retraining based on observed changes in tasks, layoffs and barriers to accessing training.
- Sectoral planning bodies with worker representation and real authority, established in binding form.
- A binding commitment to limiting AI in workplace surveillance and in significant employment decisions (CUPE's specific legal asks).
- A binding commitment to public procurement requirements ensuring that federal AI spending does not subsidize labour displacement at recipient companies.
The net assessment: AI for All includes labour commitments and leaves room for stronger protections. Exposure estimates help decide where to look. Evidence on actual layoffs, changed tasks, re-employment and pay is what would tell us which measures workers need and on what scale.
Where the AI adoption actually is: mines, farms, and forests
Lire cette section en français
The loudest AI conversation is about chatbots and frontier models. The largest AI deployment in Canada is a fleet of driverless mining trucks. If you want to see where automation is actually changing Canadian work, look away from the office and toward the resource economy, because that is where the technology has moved from pilot to payroll, and where the labour question is sharpest.
Mining is the clearest case. At its Base Plant oil-sands mine near Fort McMurray, Suncor now moves all of its ore with autonomous haul trucks, running roughly 120 of them by 2025, the largest such fleet at any single site in the world. The onboard systems optimize driving to cut wear and fuel. They also cut jobs: the program was projected to eliminate on the order of 400 net positions, operator roles that paid around $200,000 a year, which the union Unifor called a bigger threat to its members than any oil-price crash. This is the structural finding of this chapter in its most physical form, a real efficiency gain and a real displacement, in the same trucks.
Agriculture is the opposite shape: lots of promise, thin adoption. The federal strategy singles out precision agriculture as a flagship, and Protein Industries Canada has put millions into AI for crop genomics and supply chains. But on actual farms the numbers are small, with roughly a quarter of Canadian farms using auto-steer and only a few percent using drones, and the binding constraint is mundane: about a third of rural households lack reliable broadband, and precision agriculture needs real-time data. The tools that exist cluster on the largest operations that can afford them, which quietly accelerates farm consolidation.
Forestry and energy round out the picture. Natural Resources Canada is using LiDAR and machine learning to build forest inventories at a scale hand-surveys never could; Imperial's Kearl oil-sands operation credits process-control machine learning with record output, though the sources never cleanly separate the software's contribution from ordinary operational changes. (Wildfire detection and grid forecasting, covered earlier, belong on the same list.)
Two cautions hold this together. The first is the recurring one: in almost all of these cases, "AI" means machine learning, computer vision, and automation, not the generative kind the public debate fixates on, and the genuinely generative projects in the sector are still described as exploratory. The second is the marketing problem the rest of the book keeps meeting. Canada's securities regulators have warned issuers against unsubstantiated AI claims, and a widely cited count found that roughly 40 percent of European "AI startups" used essentially no AI at all. When a Canadian resource company says it is "powered by AI," the goblin's question is the same as always: powered how, by what, doing which part, and who checked? The adoption here is the realest in the country. So is the incentive to overstate it.
The working test
Lire cette section en français
The jobs edition of the standing test, compressed. For any AI-and-jobs claim, ask: scale (documented sectoral displacement and calm aggregate statistics can both be true — they usually are); wage and condition direction (a "jobs" count that doesn't say whether the new work pays less, monitors more, and protects nothing has skipped the actual question); source (employer, government projection, union brief, peer review — every one leans, and by now you can label the lean yourself); distribution (women, racialized workers, and specific regions and sectors carry dramatically different risk — aggregates launder this); policy contingency ("AI will do X" almost always means "AI will do X under current policy conditions," and the dropped qualifier is where the politics hides); and verification (self-reported by someone with skin in the game, or independently checkable?).
What the work shifts toward
Lire cette section en français
CHAPTER RECAP — you now have: - The Canadian baseline: Statistics Canada’s May 2021 exposure estimate, the Bank of Canada’s May 2026 assessment and CUPE’s workplace reports, each answering a different question. - Acemoglu and Autor give readers different ways to think about what AI might do to jobs. Check which tasks, workers and assumptions each argument has in view before carrying its conclusion into a particular workplace. - The CUPE Senate brief in its own framing — five specific recommendations, with the structural argument that AI deployment is rapid, regulation is absent, harms are real, and the federal strategy commits to investment without corresponding worker protections. - Why the existence of low-paid AI work and reports of job loss do not tell us whether the same workers moved between them, and what records would show where people landed and how their pay changed. - The AI for All labour commitments and the evidence needed to size retraining and other worker supports. - The working test for evaluating any AI-and-jobs claim: scale, wage and condition direction, source, geographic and demographic specificity, policy contingency, verification.
If a workplace tool causes harm, who can change the rules it operates under? Chapter 17 follows that question through Canadian departments, provinces and sector regulators.
The next time someone tells you AI is creating or destroying jobs, ask them to show their count. Then ask which workers it follows, what happened to their wages and hours, and how it ties the change to AI. A single headline number has a lot of work to do before it answers those questions.
Bias label for this chapter: empirical and analytical synthesis of AI and work. Author lean: toward measuring wages, conditions and worker control alongside job counts. Statistics Canada is primary for its exposure estimates, the Bank of Canada for its dated assessment, and CUPE for what the union reports and recommends. Research following workers can measure outcomes the exposure index cannot.
Primary sources cited or relied on in this chapter: Statistics Canada Survey of Digital Technology and Internet Use, Labour Force Survey, and analytical papers on AI and employment (2024–2026); Michelle Alexopoulos, Bank of Canada speech (May 2026); CUPE Senate Social Affairs Committee submission (March 2026, primary researcher Sarah Ryan); Daron Acemoglu, "The Simple Macroeconomics of AI" (NBER Working Paper 32487, April 2024; published Economic Policy 40(121):13–58, 2025); Acemoglu and Johnson, Power and Progress (2023); Acemoglu, Johnson, Autor, "Can we have pro-worker AI?" (CEPR Policy Insight 123, 2023); David Autor, "Applying AI to Rebuild Middle Class Jobs" (NBER Working Paper 32140, 2024); Autor and Thompson, "Expertise" (NBER Working Paper 33941, 2025); Treasury Board of Canada Secretariat AI Register; Chief Data Officer Stephen Burt public statements (January 2026); AI for All strategy launch documentation (June 2026); Policy Options "U.S. Proxy Lobbying Threatens Canada's Digital Sovereignty" (November 2025); TIME investigation of OpenAI/Sama contracting in Kenya (January 2023). Detailed citations in the Sources appendix.
🧌 GOBLIN CHECK — The unemployment rate can stay steady through someone’s worst year. If you want to know where the captioner landed, follow the captioner. Ask how long it took to find work, what the next job paid per hour and how many hours it offered.
Recap
- The Canadian baseline: Statistics Canada’s May 2021 exposure estimate, the Bank of Canada’s May 2026 assessment and CUPE’s workplace reports, each answering a different question.
- Acemoglu and Autor give readers different ways to think about what AI might do to jobs. Check which tasks, workers and assumptions each argument has in view before carrying its conclusion into a particular workplace.
- The CUPE Senate brief in its own framing — five specific recommendations, with the structural argument that AI deployment is rapid, regulation is absent, harms are real, and the federal strategy commits to investment without corresponding worker protections.
- Why the existence of low-paid AI work and reports of job loss do not tell us whether the same workers moved between them, and what records would show where people landed and how their pay changed.
- The AI for All labour commitments and the evidence needed to size retraining and other worker supports.
- The working test for evaluating any AI-and-jobs claim: scale, wage and condition direction, source, geographic and demographic specificity, policy contingency, verification.
Sources
- Statistics Canada Survey of Digital Technology and Internet Use, Labour Force Survey, and analytical papers on AI and employment (2024–2026)
- Michelle Alexopoulos, Bank of Canada speech (May 2026)
- CUPE Senate Social Affairs Committee submission (March 2026, primary researcher Sarah Ryan)
- Daron Acemoglu, "The Simple Macroeconomics of AI" (NBER Working Paper 32487, April 2024
- published Economic Policy 40(121):13–58, 2025)
- Acemoglu and Johnson, Power and Progress (2023)
- Acemoglu, Johnson, Autor, "Can we have pro-worker AI?" (CEPR Policy Insight 123, 2023)
- David Autor, "Applying AI to Rebuild Middle Class Jobs" (NBER Working Paper 32140, 2024)
- Autor and Thompson, "Expertise" (NBER Working Paper 33941, 2025)
- Treasury Board of Canada Secretariat AI Register
- Chief Data Officer Stephen Burt public statements (January 2026)
- AI for All strategy launch documentation (June 2026)
- Policy Options "U.S. Proxy Lobbying Threatens Canada's Digital Sovereignty" (November 2025)
- TIME investigation of OpenAI/Sama contracting in Kenya (January 2023).