哈佛大学研究人员对超过700家软件开发企业进行了研究，分析了2021年至2026年3月期间超过700万名员工的3亿次工作事件数据，发现尽管AI编程助手能够生成大量功能代码，但人工代码审查流程成为了效率提升的瓶颈1。研究由哈佛大学研究员Fiona Chen和James Stratton主导，数据来自工程团队输出测量平台Jellyfish1。
研究结果显示，代码审查流程明显增长，拉取请求更可能需要修订，审查人员留下更多注释1。这意味着编码阶段因AI工具带来的效率提升被下游审查环节的工作量增加所抵消。研究团队得出的结论是，"几乎没有证据表明企业通过使用AI工具增加了软件输出或减少了就业"1。
Researchers at Harvard University conducted an extensive study of over 700 software development companies, examining more than 300 million work events including code submissions and pull requests from 2021 through March 2026, and found that while artificial intelligence coding assistants produce substantial volumes of functional code, human code review has become a significant efficiency bottleneck.1 The research, led by Harvard researchers Fiona Chen and James Stratton and drawing on data from Jellyfish, an engineering team output measurement platform, revealed that code review processes have expanded noticeably, with pull requests becoming more likely to require revisions and reviewers leaving more comments.1
Despite the productivity gains during the coding phase, these improvements are offset by the growing downstream review workload, resulting in no net increase in software output or reduction in employment at enterprises using AI tools.1 The study concluded that there is "almost no evidence that companies have increased software output or reduced employment by using AI tools."1
评论
还没有评论，欢迎留下第一条。