Why Good Engineers Become Worse with AI (nidhish.dev)

A recent analysis highlights a paradoxical effect of AI tools on engineering performance, likening it to a phenomenon first observed by Francis Galton in 1886. While Large Language Models (LLMs) significantly enhance common engineering tasks, they can hinder creativity and innovation, causing proficient engineers to produce more average work when tackling novel challenges. This regression occurs because LLMs are designed to sample the most probable outputs—often leading to the omission of unique, innovative solutions in favor of safer, more common responses, ultimately resulting in structurally sound but technically flawed code. The study emphasizes that this issue stems from the models' inability to recognize the importance of specific contributions, such as critical lines of code deriving from less common mathematical formulas. Notably, while AI systems like OpenAI's reasoning model and DeepMind's AlphaProof Nexus have made significant strides in formal proofs and solving longstanding mathematical problems, they remain limited by their training data. Novel approaches within high-value sectors often go unrecognized as they remain unpublished, hidden behind corporate walls. This raises a critical warning for engineers: relying too heavily on AI tools may stifle their judgment and innovation, urging them to maintain a strong grasp of what uniquely drives their contributions.