Abstract
The integration of Large Language Models (LLMs) into software engineering and critical infrastructure has been driven by vendor consensus and artificial benchmarks, masking a profound systemic failure. This meta-study synthesizes longitudinal raw telemetry, CI/CD artifacts, randomized controlled trials (RCTs), and neuroimaging to bypass subjective sentiment surveys and expose the net-negative productivity loops of LLM-driven workflows. We identify a fundamental ontological mismatch: LLMs are probabilistic next-token predictors optimized for linguistic fluency, not semantic comprehension or logical correctness. Applying these statistical text-generators to zero-error engineering environments has triggered a triad of structural deficits.
Economically, proprietary providers extract capital through opaque “idle meter” billing, background loops, and silent model downgrades, financially incentivizing failure-and-retry cycles over single-shot resolution. Technically, the illusion of rapid drafting is erased by a severe “review/rework tax.” Alarmingly, 24.2% of AI-introduced vulnerabilities survive into production due to automation bias and soft CI/CD gating. Cognitively, the outsourcing of productive cognitive friction to algorithmic generation measurably degrades human enactive neuroplasticity. EEG telemetry and longitudinal data reveal a widening skill gap between domain-master engineers and superficial “prompters,” accelerating a trajectory toward a macro-economic “knowledge-collapse steady state.”
We conclude that the current Generative AI paradigm constitutes a compounding global infrastructure risk. To mitigate this, we propose immediate institutional directives: the mandatory unbundling of API billing, strict human-expert hard-gating for critical systems, a transition to deterministic fixed-cost computing, and a systemic reallocation of research funding toward verifiable, neuro-symbolic AI architectures.
