🧑💻 AI Slows Down Senior Coders
The AI research institute METR found that when using AI, experienced developers spend more time on tasks, even though they feel more productive.
In the experiment, 16 programmers (with an average of over 10 years of experience) completed 246 tasks in large, open-source repositories they were familiar with. Some tasks were done using Cursor, while others were solved "the old-fashioned way" without AI. Whether or not to use AI was assigned randomly.
The coders expected AI to make them 24% faster, but working with AI actually took 19% longer. Extra time was spent writing prompts, waiting for responses, and reviewing generated code.
Even after finishing the tasks, developers still believed AI had made them 20% faster. Meanwhile, external experts expected productivity to increase by nearly 40%.
Researchers believe the issue is overconfidence in AI's capabilities. In reality, AI's answers are often inaccurate, its narrow context window hinders continuous work, and in familiar repositories, AI suggestions are less valuable.
However, the authors emphasize that AI does not always impair productivity. Other research indicates that AI accelerates work on smaller projects and aids in the understanding of unknown topics, making it more useful for novices and less experienced programmers.
For experienced coders, AI may make the work easier and more pleasurable, and psychologically, that is more important than speed.
