Is AI Making Developers Worse? FAQ

At the skills you offload to it - the evidence says yes. An Anthropic study found developers learning a new library with AI assistance scored 17% lower on comprehension, with debugging hit hardest. The skills you keep practicing are fine; the problem is that heavy agent workflows quietly stop you from practicing the load-bearing ones.

Because the parts of coding you delegated stopped getting practiced, and skills decay through disuse - the same mechanism as any unused muscle. This is skill atrophy, not a personal failing: reliable automation trains everyone who uses it to under-monitor and under-practice. The good news is it reverses the same way it developed, through reps.

When an AI writes the code, you get the output without the encoding - the mental work that would have wired it into memory never happened. The MIT "Your Brain on ChatGPT" study caught this directly: most participants couldn't quote from an essay they had just written with ChatGPT. Researchers call the accumulated gap cognitive debt.

Not dumber across the board - but on offloaded tasks, measurably less engaged. EEG data shows the weakest neural connectivity in people writing with ChatGPT, and two large surveys link heavier AI reliance to lower critical-thinking scores. The effects are task-specific and appear reversible, which is exactly why "dumber" is the wrong word and "in debt" is the right one.

It's the highest-risk group. Seniors who atrophy are losing skills they once built and can rebuild; juniors with AI from day one may never build them - they report higher productivity and lower confidence in their own skills than any previous cohort. A junior who debugs with AI as a second opinion learns; a junior who debugs by re-prompting does not.

The retention research points to one trait: cognitive engagement. Read and understand diffs before accepting them, form your own hypothesis before asking the AI to debug, ask follow-up questions about why a fix works, and keep some tasks fully manual - the way pilots keep hand-flying hours. Offload storage and boilerplate; keep diagnosis, comprehension, and recall.

No - abstinence trades a manageable problem for a competitive one. Nothing in the research argues against AI use; it argues against unconscious AI use. The studies consistently show the damage concentrates where engagement drops to zero, so the fix is choosing deliberately what to delegate, not deleting the assistant.

Honest tells: you read your own recent code and can't explain it, you paste errors into the chat before reading them, you can't name what your last accepted diff actually changed, and working without the assistant feels not just slower but disorienting. The METR study adds a warning: developers' self-perception was wrong by about 40 percentage points, so don't trust the feeling of "I'm fine" - test it.

Yes. Every mechanism in the research is disuse-based, and disuse-based decline rebuilds with use. Practical protocol: pick one debugging session a day to run unassisted, explain each accepted change in your own words, and code something small from scratch weekly. The MIT data suggests recovery lags - the debt lingers after you stop borrowing - so expect weeks, not days.

For the offloaded material, yes. Memory encoding requires processing, and delegated work skips the processing - the MIT study's most striking result was participants unable to recall text they had produced minutes earlier with ChatGPT. Information you actively work with, AI in hand or not, encodes normally.

The components are well-studied even where the label is new: cognitive offloading (Risko & Gilbert, 2016), automation complacency (four decades of aviation research), metacognitive laziness (Fan et al., 2024), and AI-specific results from MIT, Anthropic, METR, and Microsoft + CMU. Six serious studies in two years, all pointing the same direction, is past the "moral panic" threshold.