With the rapid development of AI (Artificial Intelligence) and robotics technology, people are beginning to worry that machines will take away human jobs, despite the convenience they bring. Many believe that in the future, engineers will be able to easily write perfect program codes by just “talking to AI.” However, a recent study in 2025 has shed new light on this perception.
The study carried out by Cornell University in the United States involved 16 developers with moderate experience in using AI to participate in experiments. They tested the performance of the latest AI tools in early 2025, such as Cursor Pro and Claude 3.5/3.7, in real programming tasks.
In the study, 246 tasks were randomly divided into two groups: one allowing the use of AI and the other not allowing it.
Surprisingly, developers using AI took 19% longer to complete tasks than those not using AI!
This result was completely opposite to the expectations of the participants and experts, who originally predicted that AI could reduce task completion time by 20% to 39%. This raises the question: have we overestimated the capabilities of AI?
The study found that AI’s “assistance” became a burden for experienced developers. These engineers, who were well-versed in project details, discovered that the recommendations made by AI often did not meet the requirements, and even deviated from the project’s “unwritten rules” – norms, conventions, or specific practices that were not explicitly documented in the project but were understood by the participants.
This kind of “tacit knowledge” accumulated through experience by developers involved in projects long-term, such as specific code styles, design logics not documented, project priority rankings, or team working methods, is challenging for AI to learn.
Therefore, AI’s suggestions not only did not save time but also required developers to spend more effort correcting or rewriting code.
The research team analyzed 140 hours of video data and found that developers using AI spent a significant amount of time waiting for AI responses, checking outputs, and modifying suggestions.
Statistics showed that less than 44% of the code generated by AI was directly adopted, with the majority being rewritten or discarded. It was like having a well-meaning but unfamiliar newcomer who ends up slowing down progress.
Does this mean that AI is useless? Not at all! The study indicated that in areas where developers are unfamiliar, such as new data formats or writing test programs for the first time, AI can indeed be a helpful assistant when knowledge is lacking.
However, in familiar projects, AI’s “wishful thinking” suggestions can be more of a hindrance than a help. The real value of AI, as shown in the study, lies in filling knowledge gaps rather than replacing human intuition and experience.
This research provides insights into the true application of AI. While AI may shine in the laboratory setting, its efficiency in the real world may not meet expectations.
The research team believes that this does not mean AI is of no use but rather highlights areas for improvement in current tools, such as faster response times, better project understanding, and even customized training methods.
Collaborating with AI is akin to having a new team member: it may be chaotic at first, but once one learns how to divide tasks and provide guidance, the potential for the future is limitless.
The study serves as a reminder that AI is ultimately a tool, and the key lies in how to wield it. The top developers of the future will not be the fastest at writing programs but those who can accurately determine when to rely on AI and when to rely on their own instincts. In the future, AI will not replace you, but those who understand how to use AI may potentially replace others.