AI tools designed to identify unknown chemicals are actually just guessing which molecules are popular in databases.

Machine learning models for identifying small molecules often achieve high scores by exploiting statistical shortcuts instead of analyzing chemical signatures. These algorithms look for patterns that suggest a molecule is common enough to be in a reference library rather than deciphering its actual structure. Traditional benchmarks fail to catch this behavior, making many top-performing tools look significantly more capable than they really are. A model that simply memorizes the most likely molecules in a dataset can outperform sophisticated systems that try to learn the underlying physics of mass spectrometry. This flaw means that researchers might be relying on false positives when searching for new medicines or environmental toxins. Fixing these evaluation methods is the only way to ensure that AI actually understands chemistry instead of just gaming the system.