Where AI Helps Chemistry Research, and Where Its Reliability Falls Short

A review published in Chemical Society Reviews starts from a plain question: when an AI tool performs well on a test, does it actually help a chemist decide what to do next? The paper comes from the Skoltech AI Center, working with the Zelinsky Institute of Organic Chemistry of the Russian Academy of Sciences, ITMO University and other institutions. It examines where AI produces measurable value in a chemical research workflow, how to judge its reliability, and why a strong benchmark result does not necessarily carry over to laboratory performance.

The authors build the review around the decisions chemists make. Chemical research generates many kinds of information: publications, images, spectra, measurements and reaction records. Understanding how molecular structure relates to properties remains essential. Researchers also make a sequence of practical choices, including which evidence to examine, which experiment to run next and how to change the plan once a new result comes in.

Using that sequence as a frame, the authors divide chemical research into stages that generate, transform and validate information, then map AI methods onto each one. They cover classical machine learning, models that build in chemical knowledge, and the emerging foundation models for chemistry. Their emphasis falls on measurable outcomes and recurring failure modes rather than benchmark scores alone.

The areas surveyed include extracting knowledge from the scientific literature and accumulated data sets; interpreting experimental data from spectroscopy, mass spectrometry and microscopy; predicting properties and reactivity; planning syntheses; and systems in which one experiment's outcome guides the next. Representative applications show what can transfer to a new problem and where a method's limits matter.

In their conclusion, the authors name three ways AI can change a chemist's work. It can draw mechanistic insight from complex spectra, images and scientific texts. It can pick out promising compounds, synthesis routes and experimental conditions from among many options. And it can link predictions to automated experiments whose results steer the next step. Throughout, they hold that AI's practical value depends on whether it improves decisions under real research constraints.

Alexander Ryabov, a senior research scientist at the Skoltech AI Center, describes the perspective the team took. They wanted to view AI as a chemist who must make research decisions would, he says, and high accuracy on test data is only the tip of the iceberg. What matters more is whether a model helps interpret measurements, plan informative experiments and keep its predictions reliable on new compounds and reactions.

The review also lays out risks that can make a model look more successful than it is: information leakage between training and evaluation, overfitting to a limited set of examples, and extrapolation errors beyond familiar data. These matter especially in chemistry, the authors note, because substances, reaction conditions and instruments vary, while experimental validation takes time and resources. Before adopting a model, they advise checking three things: whether its training data match a particular laboratory's compounds and instruments, how reliably it estimates its own prediction uncertainty, and whether it has been tested on new experiments whose results were unknown during training.

Looking ahead, Michael G. Medvedev, head of the Theoretical Chemistry Group at the Zelinsky Institute and first author of the review, expects the next stage to combine advanced large language models with physics-informed models that respect chemical constraints. An AI agent would need to judge when a physics-informed model's prediction is well founded, when it goes beyond its training data, and which experiment would best reduce uncertainty. The researcher, he says, remains responsible for defining the scientific question and verifying what the agent produces.