Aleks data shows ChatGPT-era students 25% less likely to solve word problems without AI

For a manager, the useful detail in the UC Irvine study, as Entrepreneur reported it, is where the decline showed up. On the platform, with chatbots available, the measures a supervisor would track got better ^[3]. The drop appeared only in proctored exams with the tool removed ^[4], and only on problems that could be pasted into a chatbot ^[2]. A general slide in math ability should have hit the graphing tasks as well, and it did not ^[5].

Sina Rismanchian, the U.C. Irvine Ph.D. candidate who led the study, told The New York Times that AI lets students offload the step-by-step thinking that math problems require ^[6]. Speed comes the week the tool does. Weaker understanding builds over years, and the Aleks record holds about three years of post-ChatGPT data against roughly seven years before it ^[1].

The nearest this record comes to evidence about work is a separate multi-institutional experiment on adults, cited by Rismanchian and his co-authors. As participants grew used to letting language models produce the answers and skipped the "productive struggle" of working through each step, they became more likely to abandon hard questions they could still solve ^[7]. In a team, I would expect that to show up as escalations on problems the person could have handled.

Adam Green, a cognitive neuroscientist at Georgetown University, described the pull toward the shortcut in evolutionary terms. "Whatever you could do that exerted the least calories but didn't kill you was the best thing to do evolutionarily," he told the Times ^[9]. He also told the paper that "Part of how AI is making us dumber" is "undercutting the development of learning how to think in younger people that now have an alternative" ^[8]. Junior hires learning a role with an assistant already open are the workplace version of the group he describes.

Middlebury College's work suggests the pattern of use changes the outcome. Its July study of more than 200 undergraduates found that students who used AI like a tutor had higher test scores and higher-quality essays than students who had it draft entire essays ^[10]. Entrepreneur's account does not report how large the gap was.

The decision this quarter is what to measure. A team scored only on assisted output can post better numbers while unaided skill erodes, the same pattern the Aleks homework showed ^[3][4]. The consequence comes quarters or years later, when the tool is unavailable or wrong. In the Aleks data, the loss was found only by testing without the tool ^[4].