San FranciscoGoogle says 75% of its new code is AI-generated.
Developers using mandated AI tools report skill atrophy — they can no longer reconstruct or debug their own code, multiple developers said.
One warned of unsortable tech debt if AI costs spike; executives report AI writing 75–90% of company code.
Sources: 404 Media
THE GENERATION EFFECT
Cognitive psychology has known since the 1970s that information you generate yourself is retained far better than information you merely read. The act of struggling to produce an answer — not the answer itself — is what builds the memory trace. Outsource the struggle, lose the trace.
THE GPS PRECEDENT
London taxi drivers who memorize the city's 25,000 streets — 'The Knowledge' — develop measurably enlarged posterior hippocampi. Studies since 2000 show GPS users navigate worse without the device than non-users, and their hippocampal activity during navigation drops. The tool that helps you arrive erodes the faculty that got you there.
THE CALCULATOR DEBATE, REVISITED
When calculators entered classrooms in the 1970s, educators warned of arithmetic atrophy. They were right — most adults today cannot do long division by hand — but the tradeoff was accepted because arithmetic is mechanical. The harder question for code is whether debugging and architectural reasoning are mechanical in the same way, or whether they require the same kind of internalized model that fluent arithmetic once provided.
WHY DEBUGGING IS DIFFERENT
Writing code and debugging code are asymmetric skills. Writing can be pattern-matching from examples; debugging requires holding the program's actual execution model in your head and reasoning about state. A developer who never wrote the code from scratch never built the mental model needed to find why it broke at 3am.
THE TECH DEBT TIME BOMB
Software systems accumulate complexity that only their authors fully understand. When the authors are LLMs and the human reviewers skim outputs without internalizing them, the institutional memory of why the code is shaped this way never forms. The bill comes due during incidents, migrations, or — as the article notes — when AI inference costs make regeneration uneconomic.
THE DEVALUATION FLOOR
Every previous wave of developer-productivity tooling — compilers, IDEs, Stack Overflow, Copilot — was predicted to deskill the profession and instead raised the ceiling on what a single developer could ship. The open question with LLMs is whether they raise the ceiling or lower the floor: do junior developers still develop into senior ones if they never struggle through the intermediate stage?