Researchers assessed the effectiveness of AI in 27 scientific disciplines

A group of scientists published a large-scale study, comparing the capabilities of neural networks with traditional methods of scientific analysis. The authors analyzed 2507 direct comparisons, covering 27 scientific disciplines, over the period from 2000 to the beginning of 2025.

The results of the work demonstrate a non-uniform picture of AI effectiveness. In comparison with classical statistics, neural networks often show higher results, however, they require significantly greater computational costs. At the same time, in 25% of cases, AI turns out to be not only more expensive, but also less accurate than traditional statistical tools, and this indicator remains stable over the last decade.

Regarding methods of scientific modeling and calculations, the situation is changing more dynamically. For a long time, neural networks lagged behind specialized algorithms, but since 2020, the situation has begun to transform. At the moment, AI surpasses traditional methods of scientific modeling in more than 50% of cases, demonstrating lower requirements for computational resources in these tasks.

The authors emphasize that neural networks are not a universal replacement for existing scientific approaches. Instead, they form a new segment of the "AI-oriented scientific frontier", where effectiveness depends on the specific discipline and type of problem being solved. The work is published under the Creative Commons Attribution 4.0 license.