AI Code Assistants Short-Circuit the Creative Process in Programming, Study Finds

Forum topic · 二一 · 2026-05-14

Summary

A study by Zeinabsadat Saghi, Run Huang, and Souti Chattopadhyay (arXiv:2605.13776) examined how LLM-assisted coding affects programmers' creative processes. Twenty programmers completed two sets of programming tasks, one with LLM assistance and one without, with the focus on ideation rather than code quality. Results showed that with LLMs, the idea-generation phase shortened significantly (p=0.0004) and 'aha' moments dropped significantly (p=0.002), as programmers rushed to execution. LLM-assisted code was more correct and runnable, though the total quantity of creative ideas was roughly the same. The authors argue the effect is not cognitive decline but friction removal: LLMs answer too readily, eliminating the productive struggle that drives divergent thinking and better design. They propose design implications—AI tools should support the creative process, e.g., by prompting users to consider alternative approaches before accepting a first solution. A participant described the experience as 'like taking the path of least resistance.'

"Creativity is fundamental to being human." So opens the paper, which then ran a simple but brutal experiment.

Twenty programmers each completed two sets of programming tasks — one with LLM assistance, one without. What was compared was not code quality (that's been studied to death) but the creative process itself.

Results

With LLMs, the idea-generation phase shrank significantly (p=0.0004). Programmers jumped to the "execution" stage much faster.
Creative moments dropped significantly (p=0.002). Those "wait, there's a better way to do this" flashes of insight became rarer.
LLM-assisted code was more correct and more likely to run — but the quantity of creative ideas was roughly the same.

The researchers identified four human–LLM collaboration patterns supporting different problem-solving strategies. But the overall trend is clear: LLMs let programmers produce functional code faster, at the cost of skipping the divergent-thinking phases that might have led to better designs.

One participant summed it up: "Like taking the path of least resistance."

Not making programmers dumber — just enabling shortcuts

The paper's finding is not that LLMs make programmers dumber. It's that LLMs are too good at handing you an answer. When you have a tool that随时 says "just write it this way," you rarely feel the urge to ask "is there a better way?" Human creativity is largely driven by friction — being stuck, being forced into detours, having to work around problems. LLMs remove much of that friction. But friction happens to be fertile ground for creativity.

Design implications

The design takeaway is clear: generative AI tools should be designed to "support the creative process" rather than merely "produce correct results." For example, after offering a solution, follow up with "have you considered an alternative approach?" Or require users to compare multiple ideas before accepting the first one.

Reference paper: Zeinabsadat Saghi, Run Huang, Souti Chattopadhyay. "'Like Taking the Path of Least Resistance': Exploring the Impact of LLM Interaction on the Creative Process of Programming." arXiv:2605.13776, 2026.