哈佛大学近期针对 718 家企业开展的一项研究借助了 Jellyfish 分析平台，研究发现：尽管 AI 编程助手与智能体能够提升代码产出量，但并不会转化为完工工作量的同步增长，部分原因在于 AI 生成的代码加大了开发后期的工作量，尤其是代码评审环节。
研究数据的时段为 2021 年 1 月至 2026 年 3 月，覆盖 700 余家企业、70 万名员工，合计约 3 亿个工作事件。数据来源于多个渠道，包括人力资源信息、代码版本控制系统、工单管理系统、谷歌日历等日程工具，以及 GitHub Copilot、Cursor、Claude Code 等生成式 AI 工具。
研究人员采用三项指标衡量编码产出：新增代码行数、代码提交记录、合并请求。
AI 助手只会小幅提升编码产出：代码行数增加 12%，提交记录增加 9%，合并请求增加 5%；其中仅提交记录的增加具有统计显著性。AI 智能体则能在三项指标上带来显著的大幅增长：代码行数增加 30%，提交记录增加 20%，合并请求增加 23%。
论文作者基于上述数据，以每名工程师每月解决的 Jira 工单和大型需求数量评估软件交付成果；工单被视为独立的工程任务单元，大型需求对应规模更大的项目或功能。研究发现，编程助手并没有提升工单解决数量，而 AI 智能体带来了 12% 的增长，但增幅远低于代码行数 30% 的预估涨幅。
为解释这一落差，研究团队考察了完整开发流程——编码之外还包含多个环节：功能与范围规划、代码评审、测试和部署。
我们基于 Jellyfish 数据推算工时分配，并通过面向 100 名 Prolific 平台工程师的调研验证推算结果：工程师仅将 30% 至 40% 的时间用于写代码，其余时间都投入到配套的任务中。当 AI 提升编码效率时，后续配套任务就可能成为瓶颈。
具体来看，合并请求的平均评审时长增加了 49%，单条合并请求的评论数量增加了 53%，需要修改的合并请求占比翻了一倍。除此之外，参与代码评审的人员占比增加了 14%，即便引入 AI 辅助评审也没有改善这一状况。作者同时指出，合并请求相关工作量上涨并不能用单次合并请求体量变大来解释。
Reddit 用户 Prime_1 在评论这项研究结果时表示：“每一行代码都是潜在风险”。用户 markehammons 进一步延伸了这个观点，强调了 AI 幻觉带来的额外风险：
当你面对一个不可靠的合作对象时，风险更是加倍，而智能体比大多数合作对象都更不可靠。它们可能在毫无恶意的情况下做出最糟糕的事情，只是一心想要完成指派任务，却会做出各种离谱的操作。
Reddit 用户 enigmamonkey 还进一步提出，故障平均恢复时长或许会成为评估 AI 对软件开发影响的一项愈发重要的指标：
如果工程师对自己交付的代码的编写量越来越少、理解也越来越浅，我猜想故障会变得更加频繁，排查修复也更加困难。尤其是代码库持续扩张的时候——而代码库必然会扩张。
Jellyfish 是一家软件公司，提供用于衡量和分析软件工程绩效的平台。它接入了来自 GitHub 和 Jira 等工具的数据，帮助企业了解开发者生产力和软件交付情况。
查看英文原文：https://www.infoq.com/news/2026/10/jellyfish-ai-productivity/
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
A recent study by Harvard University on 718 businesses, which benefited from the Jellyfish Analysis Platform, found that although AI programming assistants and AI agent were able to increase code output, it did not translate into a synchronized increase in the completion workload, partly because the codes generated by AI increased the workload of developing post-production, particularly the code review chain.
The time frame for studying the data was January 2021 to March 2026 and covered over 700 enterprises and 700,000 employees, totalling about 300 million work events. Data are derived from a variety of sources, including human resources information, code version control systems, schedule tools such as the Worksheet Management System, Google Calendar, and generated AI tools such as GitHub Copilot, Cursor, Claude Code.
Researchers measure output using three indicators: new code lines, code submission records, consolidation requests.
AI assistants only add a small amount of coding output: a 12 per cent increase in code lines, a 9 per cent increase in submission records and a 5 per cent increase in consolidation requests; only an increase in submission records is statistically significant. AI Z42Z was able to bring significant growth in three indicators: 30 per cent in code lines, 20 per cent in submission records and 23 per cent in consolidation requests.
Based on the above data, the author of the paper assesses the results of the delivery of the Jira worksheet, which is addressed monthly by each engineer, and the large demand quantity software; the worksheet is considered to be a stand-alone engineering task module with large demand corresponding to larger projects or functions. The study found that the programming assistant did not raise the number of worksheets, while AI Z42Z brought about a 12% increase, which was well below the projected increase of 30% for code lines.
To explain this gap, the research team examined the full development process, which includes multiple components: functional and scope planning, code evaluation, testing and deployment.
We extrapolate the hours allocated based on Jellyfish data and validate the results by researching 100 Prolific platform engineers: engineers spend only 30 to 40 per cent of their time writing codes, and the rest of their time on accompanying tasks. When AI enhances coding efficiency, follow-up missions can become bottlenecks.
In particular, the average length of evaluation of consolidation requests increased by 49 per cent, the number of comments on individual consolidation requests increased by 53 per cent and the number of consolidation requests requiring revision doubled. In addition to this, the share of participants in code reviews increased by 14 per cent, and even the introduction of AI-assisted reviews did not improve the situation. At the same time, the authors note that the increased workload associated with consolidation requests cannot be explained by the larger volume of individual consolidation requests.
Readdit user Prime 1, commenting on this finding, states: “Every line code is a potential risk”. User Markehammons further extends this view by highlighting the additional risks posed by AI hallucinations:
When you face an unreliable partner, the risk is doubled, while AI agent is less reliable than most partners. They may do the worst without any malice, but with the sole intention of fulfilling their assigned tasks, they will do all sorts of crazy things.
Reddit user enigmamonkey further suggests that the average recovery time may be an increasingly important indicator of the impact of AI on software development:
If engineers were to produce less and less understanding of the codes they had delivered, I would assume that the failure would become more frequent and that the searching would be more difficult to repair. Especially when it continues to expand -- and it will necessarily expand.
Jellyfish is a software company that provides a platform for measuring and analysing software engineering performance. It provides access to data from tools such as GitHub and Jira, and helps enterprises understand developers ' productivity and software delivery.
See the original English language version at https://www.infoq.com/news/2026/10/jellyfish-ai-policy/
Auto-fetched and translated by AI Atlas. All rights belong to the original author.