Artificial Intelligence may ace basic lab tasks, but it still can’t reason like a scientist. A new study by researchers at IIT Delhi and Friedrich Schiller University (FSU) Jena has found that while leading AI models perform impressively in perception-based and routine scientific functions, they lack the deeper reasoning required for complex research and discovery.

Published in Nature Computational Science, the study underscores a crucial limitation: today’s AI understands patterns, not principles. Researchers found that although current systems can accurately recognize laboratory instruments and handle simple visual data, they struggle with spatial reasoning, cross-modal analysis, and multi-step logic — the very skills that underpin scientific insight.

“Our findings are a reality check for the scientific community,” said Dr. N.M. Anoop Krishnan, associate professor at IIT Delhi and co-lead author of the study. “These AI models perform well on data-heavy tasks but rely heavily on pattern-matching from internet data rather than true understanding.”

The research team, also led by Prof. Kevin Maik Jablonka from FSU Jena, developed MaCBench, the first comprehensive benchmark for testing how vision-language AI systems perform in real-world chemistry and materials science contexts.

When recognition fails reasoning

The results revealed a troubling imbalance: models achieved 77% accuracy in identifying lab equipment, but only 46% accuracy in assessing safety hazards — a gap researchers called “alarming.”
“This disparity shows that AI can recognize objects but cannot apply contextual judgment — a vital aspect of lab safety and scientific reasoning,” Jablonka noted.

Humans still indispensable in AI-driven science

The team also found that AI systems fared better when given textual rather than visual data, highlighting weak multimodal integration — a major obstacle to autonomous scientific reasoning.
“AI can support, but not replace, human scientists,” the researchers stressed, urging caution in deploying unsupervised AI in laboratories or research settings.

The findings reaffirm that the future of science with AI must remain human-guided, with machine intelligence serving as an assistant, not an authority.