Canadian workers and AI skill erosion: what IBM’s survey shows

IBM’s October 2, 2026 Canadian release reports that 62% of surveyed Canadian employees worry about AI eroding their skills. It also reports a gap in priorities: 68% of Canadian human-resources executives identify supervising, validating and overriding AI output as essential, compared with 29% of employees who rank judgment as important. These are respondent views, rather than a direct test of declining ability.

Start with the sample and the question

IBM and Oxford Economics conducted two surveys between April and June 2026. The Canadian portions included 75 executives and 452 full-time employees; the global totals were 1,500 executives and 8,800 employees. The executive sample represented relatively large organizations. Readers should not assume that every Canadian occupation, small business or employment situation is equally represented.

One widely repeatable number needs its denominator: IBM says 69% of Canadian employees who worry about erosion believe some skills have already eroded. That is a subgroup, not 69% of all Canadian workers. The release and IBM’s global report belong to the same research program; citing both does not create two independent replications.

A separate study asks a related question

A CHI 2025 paper by Microsoft Research authors surveyed 319 knowledge workers. It found an association between greater confidence in generative AI and lower reported critical-thinking effort. The authors also describe thinking shifting toward verification and integration of output. This is a separate study, but it too relies on reported experience; it is not a Canadian longitudinal measurement of lost skills.

Together, the studies motivate a practical question: when an assistant supplies the first answer, does the worker still practise identifying the problem and checking the result? Neither study establishes that every use of AI causes deterioration. Task design, experience and the way people review output may matter.

Keep judgment visible in the workflow

Our recommendation is to define acceptance criteria before opening the assistant. For a research summary, write down the decision it must support, the evidence needed and the uncertainty that would change the recommendation. After generation, inspect the underlying sources rather than accepting a plausible paragraph as verification.

Managers can reserve some practice without AI and compare how staff handle realistic errors. A useful exercise is to ask a colleague to identify a missing assumption in a draft, then explain a correction in their own words. Measure repair time and final quality alongside time saved. These are suggested training practices, not interventions whose effectiveness was tested in IBM’s survey.

What stronger evidence would look like

A convincing follow-up would measure the same skills repeatedly, describe task exposure and compare comparable workers under clearly specified conditions. It would distinguish short-term effort savings from durable skill change. For now, the headline is concern and a gap in priorities, with a reason to make checking and accountability part of everyday AI use.

Sources and editorial method

We use the linked original statements and reports to distinguish documented facts from our analysis. Company announcements describe their authors’ plans and claims; they are not independent verification of future outcomes. This article is dated and will be revised if material evidence changes. Read our editorial standards and corrections process.

IBM Canada: October 2 study release ↗ — Canadian results, subgroup denominators and methodology; accessed October 2, 2026.
IBM: Designing the Thinking Organization ↗ — Global report from the same research program; not independent corroboration.
CHI 2025: generative AI and critical thinking ↗ — Separate survey research, published April 2025; self-reported effort and associations.