Two people, same tool, same twenty minutes, two different afterwards
One of them pushed back twice and rewrote the middle section. The other read it, thought it was fine, and sent it. The two documents are hard to tell apart. The two people are not.
That difference was measured this spring. Nearly two thousand adults in the United States and Canada worked through ten simulated work tasks — building plans from incomplete information, reading ambiguous data, explaining the reasoning behind a decision — using ordinary commercial AI tools. Afterwards, 58% agreed AI had done most of the thinking. Within that group, people reported lower confidence in their own independent reasoning and a weaker sense that the ideas were theirs. The people who modified, challenged or rejected what they were given reported the opposite on both counts.
Two things the study does not say, and both matter
It does not say anyone became worse at thinking. What was measured was how people rated their own confidence and ownership, not what they could actually do. And it cannot say which way the arrow points — people short on confidence may simply defer more readily. The researcher's own line on it is the careful one: the issue was not AI use itself, but the degree of passive acceptance.
Take the causal question off the table and something useful survives either way. Heavy use and passive use are not the same variable, and every conversation about AI at work treats them as one.
The variable that matters is invisible to the instruments most teams use
Adoption dashboards count sessions. Policies set out what you may use it for. Training explains what it can do. Not one of those distinguishes the person who argued with it from the person who did not. That distinction is the one the data lines up behind.
It is also the harder one to see. Acceptance leaves no trace. There is no record of the objection somebody did not raise, no artifact from the version that was never questioned. What lands in the folder is a finished piece of work either way — which is precisely why how much are people using it became the metric, not because anyone thought it was the right question, but because it was the one the tools could answer.
A one-question diagnostic that is harder to fake than a policy
Sit with a piece of AI-assisted work and ask the person who produced it where they disagreed with it. Not whether they checked it — where they pushed and what came back. An answer arrives immediately or it does not arrive at all, and the difference tells you which of the two twenty-minute tasks you are holding.
None of which is an argument for using it less. The people in that study who came out steadiest were not the light users. They were the ones who treated the first answer as a position rather than a result — and that is a habit that can be designed into how the work is done, rather than left to individual temperament.
Redesigning the work so the disagreement has somewhere to happen is what turns passive use into active use: Redesign the work so the disagreement has somewhere to happen →
Keep reading: The AI deskilling risk · Ask AI what you left ambiguous
This is the thinking behind a Trusted AI Culture.
Source: Sarah Baldeo, "Generative AI Reliance and Executive Function Attenuation: Behavioral Evidence of Cognitive Offload in High-Use Adults," Technology, Mind, and Behavior, American Psychological Association, 16 April 2026. The findings are correlational and the outcomes are self-reported.
Written with AI in the loop: my idea, AI drafted and sharpened, my judgment on the way out. Every word is mine to stand behind. — Darren