How Heavy AI Reliance Erodes Self-Belief at Work

I had a conversation last week with a designer who ships AI features for a living. After a heavy stretch of work with an AI assistant in the loop, the work was good and the deadlines all hit. But the day after, she opened a new task and felt a beat of hesitation. Not about the work. About herself. Like the part of her that knew she could do this had quietly stepped back.

That hesitation has a research name now.

What the research shows

Earlier this month, I came across a paper in Behavioral Sciences from Xuehan Zhu, Aiai Zhang, and Jiacheng Zhang. They surveyed 576 students who use generative AI daily, then tested a precise question. When people lean heavily on AI, does something happen to the felt sense of being capable? Researchers call that self-efficacy. The paper traces a path from the pull of instant ease, through what the authors call self-efficacy erosion, to a kind of functional dependency on the tools.

In plain language: when reliance on AI is driven by how easy it feels and by how superior the tool seems, a person’s belief in their own capability quietly slips. The slip is not loud. The work still gets done. But in the model, the slip mediates the move from heavy use to the kind of dependency where reaching for the AI feels less like a choice and more like the default.

A separate paper came out yesterday in Frontiers in Artificial Intelligence, naming a phrase I think we will hear more of: reskilling fatigue. The authors describe mid-career professionals whose roles, identity, and value rest on accumulated expertise, and whose relationship to learning has turned defensive. Their line stuck with me. “Learning has become more defensive against a looming sense of obsolescence. It has also become a means of self-preservation.” Different population, same neighbourhood. When AI moves into the work the self was built around, the part of the self that did the work takes a hit.

This is the self-belief slip the recent piece on copy-paste versus first-human-then-AI sense of authorship was already pointing at. One paper does not settle a science. Two careful papers in a row, on adjacent corners of the same question, are worth sitting with.

What it doesn’t tell us yet

The Behavioral Sciences paper has limits worth naming. The sample is Chinese university students aged 18 to 25, early adopters in a specific life stage and culture, with split-sample validation but a cross-sectional design. That means the model can map relationships, but it cannot lock down what happens first in the loop. The thing being measured is a perceived state, not a behavioural performance test. The reskilling fatigue paper is a perspective piece, not yet measured. So the bar is a careful study and a careful theory, not a settled finding.

One thing to notice in your work today

A small thing to carry. After a stretch of heavy AI assistance, ask one quiet question before the next task. Not what was produced. What is actually believed about capability tomorrow. If the answer feels smaller than it did a month ago, that is the slip the research is pointing at.

If you ship AI features, the same question lands on the design surface. A default that drops a user into instant ease can quietly change the shape of your workday and the shape of who they think they are while doing it. The honest practice is to notice both.