The Slow and Quiet Cognitive Atrophy of a Modern Software Engineer

Elmar Chavez names the danger after Pieter Bruegel’s The Blind Leading the Blind. He describes the prompt-accept-repeat workflow of AI-assisted coding as something that quietly removes the review step that used to catch mistakes and build understanding. The UI looks right and the AI-generated tests pass. In his account, that’s enough for engineers to accept a change without reading it. The same habit moves through the code review chain. A junior’s trust in AI output cascades into a senior’s trust in the junior, all the way to production.

Fichamento

#

• Skill decay from over-reliance on AI isn’t unique to software, but the “prompt, accept, repeat” workflow removes the review step that used to catch small mistakes before they accumulate.
• Doing a task yourself, rather than just directing an AI to do it, is what forces attention to edge cases, UX decisions, and which tests actually matter.
• Human-written code carries an owner. Someone who built it can fix it, and solves the same problem faster the next time because of that ownership.
• Unreviewed AI code moving through a PR chain, junior trusting the AI and senior trusting the junior, creates a false sense of trust that reaches production.
• His daily practice against atrophy is solving one coding challenge a day, recording himself explaining the solution out loud, then reviewing the recording for where he struggles or pauses.
• He caught AI recommending a deprecated method once, something his own prior study of the subject let him notice. A fully AI-directed workflow would have missed it.
• A debugging attempt handled entirely through AI prompting produced repeated lists of three to five pre-check steps that didn’t fix anything; doing the fix himself instead took 15-30 minutes and left him with an understanding that stuck.

Connects with I don’t like LLMs, where Martin Fowler names a related but different discomfort: an aversion to the Silicon Valley culture that shaped these models, distinct from the skill-decay argument here.