Building Canada's Early Warning System for AI Job Displacement

This executive summary was prepared for Mila, Quebec Artificial Intelligence Institute, as a companion to the full Policy Analysis Exercise. The executive summary and interactive materials are available at aiearlywarning.ca.

Executive summary

Evidence is emerging that generative AI is producing localized labour market disruptions among specific groups of younger workers, even as aggregate Canadian employment remains stable. A January 2026 Statistics Canada report found that employment for 15- to 29-year-olds in highly exposed, coding-intensive professions remained stagnant between late 2022 and late 2025, while employment for workers aged 30 to 49 in the same occupations grew. A November 2025 study by researchers at Stanford's Digital Economy Lab, using ADP payroll microdata combined with Anthropic's AI usage statistics, found a 16 percent relative employment decline among U.S. workers aged 22 to 25 in occupations where AI was being used to automate tasks.

These disruptions fall below the detection threshold of Canada's standard government surveys. The Labour Force Survey lacks the cell sizes to slice by occupation, age, and province simultaneously, and the Census runs on a five-year cycle. Canada's labour market institutions were built to detect shocks after they appear in aggregate data. They have no capacity to detect an entry-level hiring freeze in a narrow occupational cohort before it becomes a national problem.

In April 2025, Mark Carney's Liberal platform promised to track the labour market impacts of AI. Budget 2025 backed that promise by allocating $25 million for TechStat, a new program at Statistics Canada with the mandate to measure how AI is affecting the Canadian labour force. As of April 2026, TechStat has no formal tracking framework in operation. This briefing note proposes that the Government of Canada build an early warning and response system for AI-driven job displacement, housed within TechStat, and sets out what the system should measure, how Canada can obtain the data, and how the federal government and provinces should prepare to act on the signals it produces. It is written for Mila, which is well positioned to shape this system during the narrow window before TechStat's first-year work plan is finalized.

Who is at risk

Mapping the Stanford study's findings onto Canadian occupational classifications produces an illustrative cohort of approximately 66,705 workers aged 20 to 24 across seven occupations. Customer service representatives, software developers and engineers, and financial customer service representatives together account for roughly 78 percent of the cohort. Ontario hosts 43.4 percent of these workers (28,980), and Quebec accounts for 23.6 percent (15,740). The cohort represents approximately 1 percent of the 6 million Canadians that Statistics Canada estimates work in occupations with high AI exposure and low complementarity. This figure is a mapping exercise that demonstrates the methodology, and it is not the scope of the proposed system. The system is designed to detect AI-driven displacement wherever it appears in the Canadian labour force. The cohort is useful because it reveals the gap between where the risk of displacement is most likely to appear first and what Canada's statistical infrastructure can currently detect.

What to measure

Four data streams show promise in closing this gap.

AI usage statistics from frontier labs provide near-real-time visibility into how models are being used to automate or augment specific work tasks.

Task exposure metrics assess how susceptible each occupation is to AI automation as model capabilities change. TechStat can produce these internally, with no external data agreements required.

High-frequency administrative payroll data from commercial providers like ADP Canada and Dayforce confirms whether displacement has reached actual workers at the occupation-by-age-by-province granularity the LFS cannot produce. The Stanford study by Brynjolfsson and colleagues (2025) combined Anthropic's automation-augmentation usage data with theoretical exposure scores. They showed a correlation with observed employment declines among early-career workers while controlling for confounding variables that include firm-level shocks, post-pandemic labour adjustments, and the fallout from the 2021 to 2022 tech hiring boom across 3.5 to 5 million payroll records.

Firm-level AI adoption surveys help attribute observed employment changes to AI rather than to other economic forces.

TechStat should build this system to operate without frontier lab data, and create standardized plug-in points that labs can connect to on their own terms.

How to get the data

Each stream carries distinct acquisition challenges. AI usage statistics are the highest-value signal and the slowest to arrive. Only Anthropic and OpenAI currently publish usage data, and the international standardization effort is unlikely to produce a formal data-sharing commitment before early 2027. Open-weight models account for an estimated 20 to 30 percent of total AI usage, and when users run these models on their own hardware, no usage statistics are generated. Given these constraints, TechStat should build the system on domestic data streams first and treat lab telemetry as a plug-in upgrade, while accounting for the skew in the available AI usage data.

Payroll data partnerships with ADP Canada and Dayforce should be structured as voluntary agreements. Statistics Canada's 2018 attempt to compel financial data under Section 13 of the Statistics Act triggered public backlash and a Privacy Commissioner investigation. The Canadian Survey on Business Conditions, Canada's only high-frequency source of firm-level AI adoption data, has its final collection cycle scheduled for August 2026. If TechStat does not absorb the AI module before it sunsets, the data stream disappears with no replacement in place.

Preparing the response

Canada's COVID-19 experience demonstrated three structural vulnerabilities that could resurface in an AI displacement scenario. ESDC's legacy benefits infrastructure, built on COBOL mainframes, could not process the surge in claims or adapt to new eligibility criteria, and the government bypassed ESDC entirely and routed CERB through CRA. The inability to verify employment status in real time (CRA's most recent income data at the time of launch was from the 2018 tax year) contributed to $4.6 billion in confirmed overpayments and at least $27.4 billion in payments flagged for further investigation. CERB payments also triggered automatic claw-backs against provincial social assistance in several provinces, because there was no time to negotiate exemptions with 13 provincial and territorial social services ministries before launch. Recent efforts by the Forecasting Research Institute, Windfall Trust, and the Partnership on AI have begun identifying suitable policy responses, and none of these efforts have been translated into Canada's specific institutional context.

Recommendations for Mila

1. Brief TechStat on the indicator framework in this report and advocate for the data streams it will need, engaging in Q2 2026 before TechStat's first-year work plan is finalized.

2. Convene a Canadian AI labour displacement scenario exercise in October or November 2026 that produces a working document mapping plausible scenarios to candidate policy responses.

3. Consult provinces through bilateral outreach to Ontario and Quebec and a joint technical briefing with the Labour Market Information Council and communicate Canada's data requirements to the international AI usage data standard effort.

Recommendations for the Government of Canada

4. TechStat should secure three domestic data streams in 2026: an LFS digital platform supplement at quarterly frequency, absorption of the CSBC AI adoption module before its August 2026 sunset, and voluntary data-sharing partnerships with ADP Canada and Dayforce.

5. TechStat and ESDC's Labour Market Information Directorate should begin quarterly calibration sessions in mid-2026 to develop shared signal interpretation, escalation thresholds, and signal-to-response mapping.

6. ESDC should table two sequenced items at the FLMM's Changing Nature of Work and Skills Working Group: an awareness agenda item in late 2026 and a structured working session on means-testing treatment of future AI displacement benefits in mid-2027.

Explore the executive summary and interactive materials

The executive summary, the interactive data visualizations, and the proposed governance architecture for TechStat are available at aiearlywarning.ca. The full Policy Analysis Exercise is the underlying academic document.