It Will Take Decades to Undo the Damage Done by "AI" (andrea.corbellini.name)

Many software firms are aggressively pushing large language model (LLM) “generative AI” tools into engineers’ workflows—sometimes coercively—because they produce convincing outputs and promise productivity gains. Real-world experience and targeted studies show a striking pattern: senior developers see little or no benefit, while junior developers get large productivity boosts. That boost masks a problem—new engineers increasingly rely on LLMs for most tasks, advancing in output but not in underlying skill. LLMs excel at pattern-matching and reapplying prior solutions from training data, but they do not genuinely invent or reason about novel problems. As a result, code produced or iteratively “fixed” by juniors and LLMs can be brittle, unmaintainable, and prone to subtle flaws that more experienced engineers must later detect and repair. The significance for the AI/ML community is profound: deploying powerful but limited tools without complementary education, oversight, testing, and human-in-the-loop practices risks creating a long-term skills gap and mounting technical debt. The author warns that within a decade or two retired senior developers may be needed to rescue an unmaintainable landscape—akin to today’s COBOL resurrections—unless curricula and industry practices adapt or truly generative, creative AI emerges. Practical implications include prioritizing fundamentals in education, enforcing rigorous code review and verification, and reframing what we call “AI” to avoid overreliance on brittle pattern-based systems.