Do AI Coding Tools Actually Make Developers Faster? The Data Says It Depends

AI 编程工具真的能让开发者更快吗？数据表明这取决于具体情况

Artificial intelligence is changing the way software is developed. Today, developers can ask an AI coding assistant to generate a function, explain an error, write a test, refactor code, or suggest a solution to a problem that might otherwise take several searches and attempts. This naturally raises an important question: Do AI coding tools actually make developers faster? The simple answer is: sometimes. 人工智能正在改变软件开发的方式。如今，开发者可以要求 AI 编程助手生成函数、解释错误、编写测试、重构代码，或针对原本需要多次搜索和尝试才能解决的问题提供方案。这自然引出了一个重要问题：AI 编程工具真的能让开发者更快吗？简单的回答是：有时可以。

Research on developer productivity presents a more complicated picture than the popular idea that AI automatically makes programmers more productive. Some studies have found significant productivity gains, while another controlled study found that experienced developers actually took longer when AI tools were available. The difference suggests that the real question may not be whether AI makes developers faster, but when and under what conditions AI makes developers faster. 关于开发者生产力的研究呈现出的图景，比“AI 能自动提高程序员生产力”这一普遍观点要复杂得多。一些研究发现了显著的生产力提升，而另一项对照研究却发现，在可以使用 AI 工具时，经验丰富的开发者反而花费了更多时间。这种差异表明，真正的问题或许不在于 AI 是否能让开发者更快，而在于 AI 在何时、何种条件下才能让开发者更快。

Evidence That AI Can Increase Developer Productivity

AI 能提高开发者生产力的证据

One of the strongest pieces of evidence comes from a set of field experiments conducted at Microsoft, Accenture, and an anonymous Fortune 100 company. The researchers studied 4,867 software developers and evaluated what happened when developers were given access to GitHub Copilot. Across the three experiments, developers with access to the AI coding assistant completed 26.08% more tasks than those without access. The researchers also found that less experienced developers experienced greater productivity gains. This is important because the study was conducted in real workplace environments rather than only asking developers whether they felt more productive. It provides evidence that AI coding assistants can improve measurable output in some professional software-development settings. But that is not the whole story. 最有力的证据之一来自在微软、埃森哲和一家匿名财富 100 强企业进行的一系列实地实验。研究人员调查了 4,867 名软件开发者，评估了他们在获得 GitHub Copilot 使用权限后的表现。在三项实验中，拥有 AI 编程助手权限的开发者比没有权限的开发者多完成了 26.08% 的任务。研究人员还发现，经验较少的开发者获得的生产力提升更为显著。这一点很重要，因为该研究是在真实的工作环境中进行的，而不是仅仅询问开发者是否感觉自己效率更高。它证明了 AI 编程助手可以在某些专业软件开发场景中提高可衡量的产出。但这并非故事的全部。

When AI Made Developers Slower

AI 何时让开发者变慢了

In July 2025, METR published a randomised controlled trial involving experienced open-source developers. The researchers studied 16 developers completing 246 tasks in mature repositories that they already had experience working with. The result was surprising. When developers were allowed to use early-2025 AI tools, they took 19% longer to complete the tasks compared with when AI tools were not allowed. Before the experiment, developers expected AI to reduce their completion time by 24%. Even after completing the tasks, they believed AI had made them about 20% faster. This creates an interesting gap between perceived productivity and measured productivity. A developer can feel faster because an AI assistant quickly produces code, explanations or suggestions. But producing code faster is not necessarily the same as completing a software-development task faster. The developer still has to understand the generated code, check whether it works, test it, identify mistakes, integrate it into an existing codebase and sometimes spend additional time correcting the AI’s suggestions. 2025 年 7 月，METR 发布了一项涉及资深开源开发者的随机对照试验。研究人员观察了 16 名开发者在他们已有经验的成熟代码库中完成 246 项任务的过程。结果令人惊讶：当允许开发者使用 2025 年初的 AI 工具时，他们完成任务的时间比不使用 AI 工具时反而长了 19%。实验前，开发者预期 AI 能将完成时间缩短 24%。即使在完成任务后，他们仍认为 AI 让他们的速度提升了约 20%。这在“感知生产力”与“实际测量生产力”之间造成了一个有趣的差距。开发者可能会因为 AI 助手快速生成了代码、解释或建议而感觉自己变快了，但生成代码的速度快并不等同于完成软件开发任务的速度快。开发者仍然需要理解生成的代码、检查其是否有效、进行测试、识别错误、将其集成到现有代码库中，有时还需要花费额外时间来纠正 AI 的建议。

So Which Study Is Right?

那么，哪项研究是正确的？

I don’t think these findings necessarily contradict each other. They studied different developers, environments, tasks and AI-tool contexts. The Microsoft-led research examined thousands of developers working in company environments and found a substantial increase in completed tasks. The METR study focused on experienced open-source developers working on repositories they already knew well. The difference is important. A developer working on repetitive tasks may benefit greatly from AI-generated boilerplate, documentation, tests or routine code. On the other hand, a developer working inside a complex and familiar codebase may spend considerable time reviewing, correcting and integrating AI-generated suggestions. In other words, the value of AI may depend as much on the task as on the tool. 我认为这些发现并不一定相互矛盾。它们研究的是不同的开发者、环境、任务和 AI 工具背景。微软主导的研究考察了在企业环境中工作的数千名开发者，发现任务完成量有显著增加。而 METR 的研究则聚焦于在自己非常熟悉的仓库中工作的资深开源开发者。这种差异至关重要。从事重复性任务的开发者可能会从 AI 生成的样板代码、文档、测试或常规代码中获益良多；另一方面，在复杂且熟悉的现有代码库中工作的开发者，可能需要花费大量时间来审查、纠正和集成 AI 生成的建议。换句话说，AI 的价值不仅取决于工具，同样也取决于任务本身。

METR itself later cautioned against treating its 2025 result as a universal statement about software development. In February 2026, the organisation reported that a newer experiment was affected by selection effects and concluded that its newer data was not reliable enough to estimate the current productivity effect precisely. It also noted that AI may be speeding developers up more in early 2026 than its earlier study estimated. That is a useful reminder: AI development is moving quickly, so productivity research can become outdated as tools and developer workflows change. METR 后来也提醒人们，不要将其 2025 年的结果视为关于软件开发的普遍结论。2026 年 2 月，该组织报告称，一项更新的实验受到了选择效应的影响，并得出结论：其较新的数据不足以精确评估当前的生产力影响。它还指出，在 2026 年初，AI 对开发者的提速效果可能比其早期研究估计的要大。这是一个有益的提醒：AI 的发展日新月异，随着工具和开发者工作流的改变，生产力研究很快就会过时。

What This Means for Developers

这对开发者意味着什么

For me, the most useful lesson is that using AI is not the same as becoming a better developer. AI can reduce the time required to perform certain tasks, but developers still need the technical knowledge to judge whether the output is correct. For example, an AI tool can generate a piece of code in seconds. But if I cannot explain what that code does, identify its assumptions, test it properly or recognise when it is wrong, then the speed I gained may come with a larger problem later. This is especially important for developers who are still learning. The temptation is to ask AI for the complete solution whenever we encounter a difficult problem. That can produce an immediate result, but it can also remove some of the thinking required to develop strong programming skills. 对我而言，最有用的启示是：使用 AI 并不等同于成为一名更好的开发者。AI 可以减少执行某些任务所需的时间，但开发者仍然需要具备判断输出结果是否正确的技术知识。例如，AI 工具可以在几秒钟内生成一段代码。但如果我无法解释这段代码的作用、识别其假设、进行适当测试或识别其错误，那么我所获得的“速度”可能会在未来引发更大的问题。对于仍在学习阶段的开发者来说，这一点尤为重要。每当我们遇到难题时，很容易倾向于直接向 AI 索要完整解决方案。这虽然能产生即时结果，但也可能剥夺了培养扎实编程技能所需的思考过程。

A better approach is to use AI as a development partner rather than a replacement for development knowledge. I can use it to: explain an unfamiliar error; suggest different approaches to a problem; generate repetitive boilerplate; create initial test cases; help me understand documentation; review or refactor code; and challenge my existing solution. But I should still be responsible for understanding, testing and validating the final code. 更好的方法是将 AI 作为开发伙伴，而不是开发知识的替代品。我可以利用它来：解释陌生的错误、建议解决问题的不同方法、生成重复的样板代码、创建初始测试用例、帮助我理解文档、审查或重构代码，以及挑战我现有的解决方案。但我仍然必须负责理解、测试和验证最终的代码。

The Real Productivity Skill

真正的生产力技能

The future of software development may therefore require a different kind of productivity skill. It will not simply be about who can generate the most code with AI. It will be about who knows when AI should be used, how it should be used, and when its output should be questioned. The research does not give us one universal answer to whether AI makes developers faster. Instead, it gives us evidence that the… 因此，软件开发的未来可能需要一种不同类型的生产力技能。它不再仅仅关乎谁能用 AI 生成最多的代码，而在于谁知道何时该使用 AI、如何使用 AI，以及何时该质疑 AI 的输出。研究并没有给我们一个关于“AI 是否让开发者更快”的统一答案，相反，它为我们提供了证据，表明……