A guide to navigating AI chemistry hype (cen.acs.org)

C&EN’s new guide walks researchers through the hype surrounding AI tools for chemistry, arguing for cautious, evidence-driven adoption. Experts warn of exaggerated claims, fragile reproducibility (LLMs often give different answers on repeat runs) and publication incentives that favor incremental new methods over reuse. The guide highlights both failures and successes: controversial preprints and an MIT study on AI-driven materials discovery that was later flagged for integrity concerns, contrasted with unambiguously transformative systems such as AlphaFold and practical wins like machine-learning potentials (MLPs) that match density functional theory (DFT) accuracy far faster and cut compute/energy costs. Technically, the guide explains which methods suit which problems: supervised learning and graph neural networks (GNNs) excel when thousands of labeled structures exist; fine-tuning can make models useful with only hundreds of examples. Transformers and chemistry-tuned LLMs (IBM’s MoLFormer-XL/RXN, ChemLLM, ChemCrow) use SMILES or language-style training to generate molecules, but general LLMs (GPT models) often underperform on structural/equation tasks and must be benchmarked. Practical checks include dataset size and similarity to training data, transferability limits of MLPs, and standardized benchmarks (SciBench, Tox21, MatBench, AMPL) — plus experimental validation for discovery claims. The bottom line: prefer specialized, benchmarked models, validate experimentally, and treat broad LLM outputs with skepticism.