盲目信任AI输出的开发模式面临长期困境

自2025年初以来，开发者在使用大型语言模型时出现了一种"关闭大脑"的现象，即不加批判地假设LLM的输出正确 1。这种做法本质上形成了一个"肉类代理"循环——人工负责验证和纠正LLM生成的代码错误 1。随着LLM能力的不断改进，这一模式的实际效果有所提升 1。

然而，这种工作方式存在根本性的困境 1。即使LLM最终能自主生成高质量代码，采用这种验证模式的员工仍将面临被淘汰的风险，因为公司可以完全用LLM替代这一人工环节 1。这意味着对于那些依赖盲目信任AI输出的工作者而言，不存在他们能够安全立足的技能阶段 1。

A recurring pattern has emerged among developers using large language models since early 2025, characterized by an uncritical acceptance of LLM-generated output without verification 1. This tendency toward mental disengagement reflects a broader shift in how people interact with AI systems as their capabilities have improved 1.

The underlying work model, sometimes described as "meat proxy" development, involves developers in a cycle of human validation and correction of LLM errors 1. While this approach has gained some effectiveness as LLM capabilities have strengthened, the author argues the strategy offers no long-term security for workers 1. Even if language models eventually reach a point where they can independently produce high-quality code, employees who have adopted this passive validation approach will become redundant, since companies could simply automate the entire process and eliminate the human component entirely 1.

The core tension is that improvement in LLM performance paradoxically undermines rather than protects the position of workers relying on this dependent model 1. There is no threshold of AI capability at which such an approach becomes sustainable for human employment 1.