Addressing the data gap in AI and worker health

Picture a warehouse worker whose job tasks are organized by an intelligent agent. Or a call-centre employee whose conversations are prompted and scored by an ambient AI system in real time. Or a lawyer who increasingly relies on AI syntheses to stay on top of dense case law.

Is AI helping these people do their jobs? If so, at what cost to their physical or mental health? Right now, Canada’s labour market information can’t tell us.

For more than 30 years, the Institute for Work & Health has been building evidence on the health and safety of working people in Canada. Most recently, we have been examining how the changing nature of work—from pay adequacy to contract permanency—shapes workers’ health and experiences. Labour market information, including datasets like the Census of Population and Statistics Canada’s Labour Force Survey, are critical resources for examining the intersection of work and health. Our experience: Canada’s LMI landscape isn’t well-suited to this task.

Recently, with a multidisciplinary research team and cross-sector partners, we launched the Partnership on AI and Quality of work (PAIQ). Through this partnership, we’ve hit three main obstacles: finding and accessing data that let us examine how AI affects different dimensions of work and workers’ health and well-being.

First, data on health and safety, technology, and working conditions are siloed, making it difficult to see links between these domains.

Second, even when it is possible to draw those links, we are limited to indirect (or proxy) measures of occupational exposure to AI-related tools (most of which have been developed outside of Canada) to map AI capabilities onto job tasks. These measures provide a useful estimate of how much AI can augment or automate tasks in a given job. Importantly, these measures highlight patterns in the labour market that could benefit from additional explanatory research. Yet proxy measures provide an incomplete picture of how AI changes working lives for better or worse, or of what it means for workplace safety, equity, or perceptions of work.

Finally, some of the most impactful technologies, such as those involved in the management of workers, are overlooked in the available data and measures. Current measures of AI exposure may not account for different AI subtypes (for example, large language models or agentic AI) or their use in conjunction with other advanced technologies (for example, robotics, vehicles, or mobile devices). Specifically, no comprehensive measure exists to capture the impacts of different AI applications. One limitation is that existing measures cannot distinguish between AI systems that manage workers and those that autonomously perform work tasks or augment the work performed by human algorithmic management. Failure to address this methodological limitation may limit our ability to understand how diverse workers and occupations are impacted by AI and to see potential patterns in the labour market.

Based on these limitations, there are steps that can be taken to enhance LMI.

First, worker health ought to be part of LMI data collection activities. Assessing AI, work, and health concepts in multidimensional, population-level surveys could help us more fully unpack the impacts of AI on workers.

In addition, multidisciplinary research teams like ours should seek opportunities to link data across domains. This would allow more explicit connections to be made between LMI and worker health and safety outcomes.

Finally, rather than inferring AI exposure from job titles or task lists, we need LMI that asks Canadian workers how they are encountering AI. This includes what role AI plays in their work and how it changes their workload, perceptions of control, and feelings about work. Our use of LMI should be combined with primary data collection activities to offer a fulsome picture of the impacts that AI can have on working conditions and health outcomes for workers.

AI is already changing how work gets assigned, watched, and evaluated. We're just not measuring what that's doing to workers and their health. PAIQ is actively working to close that gap by directly asking workers about the impacts of AI, linking siloed data, and treating health as core component of LMI. Through our work, we will generate evidence to protect workers from AI harms while also promoting the technology’s benefits.