Software Developers Say AI Is Rotting Their Brains
AI coding tools are sharply dividing software engineers: some report dramatic productivity gains and use models as powerful assistants for debugging, boilerplate, and legacy systems, while others feel their skills, pride, and understanding are atrophying as they become editors of opaque machine‑generated code. Many worry that pressure for “velocity” encourages huge, low‑quality PRs, fragmented architectures, and shallow comprehension, turning programming from a craft into industrialized “slop” production. Supporters counter that when used with clear specs, testing, and critical review, LLMs can safely offload routine work, but even they acknowledge new risks around code review bottlenecks, long‑term maintainability, and the changing nature of the job.
Overall split: empowerment vs. erosion
- Many report feeling dramatically more productive with AI-assisted coding (especially agents), describing large speed-ups on some tasks and new ability to ship ideas quickly.
- Others find AI coding slower, more frustrating, or net‑negative once review and debugging are included, and some have stopped using it for code generation entirely.
- Several comments stress that effectiveness depends heavily on task type, codebase size/age, and how the tools are used.
Where AI helps vs. where it fails
- Works well for:
- Boilerplate, tests, simple scripts, glue code, internal tools.
- Debugging in messy legacy systems, “rubber‑ducking,” planning tasks, and quickly summarizing or traversing docs and wikis.
- Raising the floor for weak areas (e.g., frontend polish, sysadmin chores).
- Works poorly for:
- Large, long‑lived codebases with complex invariants and performance constraints.
- Frameworks with many breaking versions (e.g., Odoo), where models mix APIs.
- High-quality refactors or situations where code bloat and architectural clarity matter.
Quality, review burden, and velocity
- Big concern: explosive growth in code volume without proportional review capacity.
- Reviewers describe huge AI-generated PRs (thousands of LOC), often poorly tested, with duplicated functionality or unnecessary wrappers.
- Some fear devs will become full‑time reviewers for bots; others hope bots will eventually do review too.
- Several expect more errors in production because every line of code is a liability and AI is a “firehose.”
Skills, cognition, and “brain rot”
- One camp: AI lets you “outsource thinking but not understanding”; you still need strong intuition, judgment, and line‑by‑line review.
- Another camp: heavy reliance clearly atrophies mechanical coding skills and syntax recall; people report flunking basic interview tasks after months of agentic coding.
- Some describe a psychological shift: coding without AI feels pointless or inefficient, leading to dependency and reduced motivation.
- Others compare it to calculators or cars: capabilities atrophy, but that doesn’t automatically make the tool illegitimate.
Job satisfaction, craft, and industry direction
- Many who loved “crafting” software feel reduced to chatting with a bot, editing slop, and chasing velocity metrics; some plan to leave the field or already have.
- Others enjoy an “editor” role: using experience to guide agents while offloading rote work.
- There is anxiety that commercial development will become “industrial programming,” dominated by quantity over understanding, turning every new system into legacy from day one.
Evidence, hype, and organizational pressure
- Strong skepticism about claims of 5–50× productivity; commenters note lack of credible, reproducible studies and no obvious surge in high‑quality software.
- Some point out that large companies may be mandating AI usage and then touting adoption metrics.
- Several emphasize that current discourse underplays long‑term maintenance, tech debt, and human burnout from constantly “steering a rocket ship with chopsticks.”