Large Language Models in Chemistry: Opportunities, Limitations, and Hallucination Risks

Large language models (LLMs) have moved quickly from general-purpose text tools to components of chemical research workflows. They are now used to extract data from the literature, propose synthesis routes, write analysis code, predict properties from few examples, and control laboratory instruments through tool-augmented agents. These capabilities create real opportunities, particularly for researchers and institutions with limited access to specialist software and training. They also carry risks that are specific to chemistry. Chemical information is dense with symbols, units, stoichiometry, and safety-critical numbers, and LLMs generate fluent text whether or not it is correct. This perspective reviews the main opportunities for LLMs in chemistry, the technical limitations that constrain them, and a taxonomy of hallucination types with examples of how each can cause harm in a laboratory or research setting. It then proposes a practical verification framework for everyday users, a set of reporting practices for authors, and a short research agenda. The central argument is that LLMs are most valuable in chemistry when they are treated as fallible assistants wrapped in verification tools, such as cheminformatics libraries, databases, and physical measurement, and not as sources of chemical truth.