Answers are grounded in this paper and cite their sources.
Large language models (LLMs) can provide plausible synthesis methods and reaction conditions for catalysis experiments in response to simple prompts, but the validity and accuracy of the responses are not necessarily reliable. We evaluated a zero-shot GPT-5 proposal for a trimetallic, precious metal-free reverse water–gas shift catalyst. The output initially appeared chemically plausible, but we identified errors by tasking students to verify the model’s claims against the primary literature, thermodynamic constraints, and control experiments. We then fed our observations and experimental results back into the model with a second prompt to determine whether the model would converge on a more scientifically grounded and experimentally testable catalyst candidate. Although the revised catalyst did not improve performance, the process revealed unexpected findings about catalyst stability and generated new hypotheses. Ultimately, our results show that iterative model verification can transform plausible but unreliable LLM outputs into a useful framework for student-driven catalyst preparation, hypothesis generation, and scientific training.
No takes yet. Share an insight, caveat, or question.
Moran et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: