Researchers from the Universities of Yale and Chicago conducted a large-scale scientific experiment. They compared nine popular AI models with real live scientists. The main goal was to evaluate the ability of neural networks to generate truly new ideas. For an honest test, we took more than eleven thousand real scientific publications. These articles covered seventy-one disciplines, including physics, chemistry, and biology. The scientists carefully disassembled each work down to its basic initial conditions. The models were then given only brief descriptions of several previous studies. After that, they were asked to come up with a completely new concept on their own. The results were extremely disappointing and strikingly monotonous. Neural networks almost always suggested simply mechanically linking two existing jobs. Real people use such a primitive approach only in twelve percent of cases. Different models chose it in almost half of all situations. This directly indicates the deep pattern of thinking of modern artificial intelligence.

The variety of proposed ideas was evaluated using a special biaxial taxonomy. The indicators of human creativity significantly and confidently exceeded the results of all tested models. The inclusion of a special reflection mode only significantly worsened the overall picture. For example, the share of template ideas in the Qwen3-8B model has increased dramatically and predictably. The mode of thinking did not broaden the horizons of the algorithm, but only sharpened its favorite template. An absolutely similar negative trend was clearly manifested in the DeepSeek-V4-Flash system. Even providing full texts of articles did not help the algorithms to become more creative at all. Models have become even more strongly and noticeably removed from the human distribution of ideas. A thorough analysis of the vocabulary revealed a specific and boring recipe for all neural networks. They very often used the words integrate, combine, or adapt. People prefer more specific and bold actions, for example, to replace or formalize. Interestingly, the ideas of the different models turned out to be strikingly similar to each other.

Most of the tested models were noticeably inferior to humans in the accuracy of problem formulation. They often gave out general and vague phrases instead of specific and useful solutions. The only pleasant exception was the Claude Sonnet 4.6 model from a well-known company. She even confidently surpassed the human level in some indicators of overall quality. However, her final distribution of ideas still remained unnaturally skewed. The researchers themselves honestly acknowledge certain methodological limitations of their interesting experiment. The task was to reduce the idea to one dry query without a live dialogue. A real scientist always relies on implicit experience and past failed attempts. Lively collaboration and critical feedback from reviewers also play an important role. Interactive agents and highly specialized systems can narrow this annoying gap a bit. But if you need really original and breakthrough ideas, you should only contact people for now. Artificial intelligence is not yet ready for full-fledged scientific creativity.