AI-Driven Cognitive Deskilling and Automation Over-Reliance

Emerging evidence indicates that routine use of AI systems for cognitive tasks can degrade users' critical thinking skills, professional competence, and ability to detect errors — a pattern described as "cognitive deskilling."

In one clinical study, clinicians who used AI-assisted colonoscopy for several months showed approximately 6 percentage points lower adenoma detection rates when the AI assistance was removed, compared to their baseline before AI exposure. The finding suggests that sustained reliance on AI support can erode professional skills that are essential when the AI is unavailable or incorrect.

A study of 666 participants found that heavier AI tool use was associated with lower self-assessed critical thinking, mediated by cognitive offloading — the tendency to delegate cognitive work to external systems rather than engaging with it directly. While cognitive offloading can improve efficiency, the research suggests it may come at the cost of maintaining the reasoning skills that underpin autonomous decision-making.

Automation bias — the tendency to over-rely on automated outputs while discounting contradictory information — compounds these effects. In a randomized experiment with 2,784 participants, participants were significantly less likely to correct erroneous AI suggestions when doing so required extra effort or when they held favorable attitudes toward AI. This pattern has been documented across domains, from aviation monitoring to medical diagnostics.

The phenomenon extends to everyday AI use. In a study of 1,506 participants, those who used an opinionated AI writing assistant had both the opinions expressed in their text and their own subsequently reported opinions shifted toward those suggested by the model — often without realizing the shift had occurred. More broadly, analysis of ChatGPT usage patterns shows that a large share of interactions involve cognitively demanding activities such as writing, problem-solving, and information-seeking — precisely the tasks where delegation to AI risks skill atrophy.

The Canadian implications are significant. Health Canada has issued guidance on AI as a medical device, but does not address the deskilling risks to clinicians who become dependent on AI-assisted diagnosis. The TBS Directive on Automated Decision-Making governs federal AI use but does not require monitoring of public servants' decision-making competence over time. Canadian educational institutions are rapidly integrating AI tools without systematic assessment of effects on student learning and skill development. If AI systems become unreliable or are withdrawn, a deskilled workforce may lack the competence to compensate.