You are wrong: AI does not improve your productivity if you exercise this job

We often praise AI tools as capable of boosting productivity, but according to this non -profit research organization, developers who participated in the study were slowed down by AI. Similarly, But have also developed illusions similar to the hallucinations of AI.

you wrong: ai does not

AI does not increase the productivity of programmers

Experience shows a gap between perception and reality. Nevertheless, The developers anticipated increased productivity of 24 %. However, even after the study, continued to believe that the AI had helped them perform their tasks more quickly by 20 %.

But in reality, the use of AI increased their 19 %working time. Moreover, After completing the study, the developers believe that the use of AI reduces the completion time by 20 %. Meanwhile, “Surprisingly, we find that AI increases the completion time by 19 %. However, IA tools have slowed down developers”specify the study.

The study was conducted with 16 experienced developers who work on major open source projects. The respondents provided a list of 246 real problems to be resolved (corrections. bugs, new features) and estimated the time necessary for each task.

The problems were assigned you wrong: ai does not randomly, with or without authorization to use AI. The developers then worked on their problems between February and June 2025, using Cursor Pro mainly with Claude 3.5 and Claude 3.7 Sonnet authorized. A rigorous methodology to establish reliable conclusions on the real impact of the AI.

The study identifies five factors which explain this slowdown. Suroptimism about the usefulness of AI pushes developers to unrealistic expectations. The great familiarity of developers in terms of code makes the help of the IA superfluous for these experts.

The size and complexity of the code (over a million lines) were also a challenge for AI, which is much less efficient under these conditions. Low reliability forces developers to accept only 44 % of the suggestions generated, then spend time cleaning you wrong: ai does not and revising the code.

The lack of context of the code also prevents AI from entering the environment in which it operates. which pushes it to offer unsuitable solutions.

Other factors such as generation latency. failure to provide an optimal context to the model have undoubtedly influenced the results, but their impact remains uncertain according to the researchers.

This study joins other research which questions the effectiveness of AI. Qodo. a company specializing in assistance IA for the code, noted that the profits are canceled by additional human verification work. A Danish survey has not found any impact of generative AI on employment or wages.

You wrong: ai does not

Programmers are too expert for AI

When AI is authorized. the developers also spend less time actively code and seek information, but much more to write prompt, wait for AI responses, revise outputs and remain you wrong: ai does not inactive.

New programmers recognize that AI helps to quickly test new scenarios and automate certain daily tasks without saving time. The validation of the generated code remains necessary and, unlike a trainee, the AI does not learn from its errors.

The authors Joel Becker, Nate Rush, Beth Barnes and David Rein insist on the context of their study. Additionally, The slowdown does not mean that IA tools never improve the productivity of developers.

Their expertise and the size of the codes contribute to the slowdown. But these are elements that are absent in software development contexts in general.

Research therefore does not claim that current AIs are useless or that future models will not improve. The researchers rather offer a nuanced look at the current limits of these technologies under real and complex conditions.

Want to save even more? Discover our selected promo codes for you.