AI Reliability Concerns Raised for Chemistry Research
A Chemical Society Reviews paper from Skoltech, ITMO and the Zelinsky Institute examines where AI tools can mislead chemists.
Updated on Nov. 11, 2026 in Chemistry
Researchers from the Skoltech AI Center, the Zelinsky Institute of Organic Chemistry and ITMO University published a review in Chemical Society Reviews on the reliability of artificial intelligence in chemistry. The paper organizes chemical research into stages that generate, transform and validate information.
Why it matters
The review identifies information leakage, overfitting and extrapolation errors as risks that can inflate apparent model performance in chemistry. For laboratories worldwide adopting AI tools, the authors advise checking whether a model's training data match their own compounds and instruments before use.
The evidence is a peer-reviewed review rather than new experimental data, and it provides no quantitative benchmark figures in the extraction. It maps classical machine learning, knowledge-incorporating models and foundation models onto each stage of chemical research, but does not report how widely these failure modes occur in practice.
The players
Skoltech AI Center
The Skoltech AI Center is the artificial intelligence research center of the Skolkovo Institute of Science and Technology in Moscow, whose researchers led the review.
Chemical Society Reviews
Chemical Society Reviews is a peer-reviewed journal published by the Royal Society of Chemistry that features authoritative review articles across chemistry.
ITMO University
ITMO University is a research university in Saint Petersburg, Russia, that co-authored the review.
Zelinsky Institute of Organic Chemistry
The Zelinsky Institute of Organic Chemistry is a research institute of the Russian Academy of Sciences in Moscow that co-authored the review.
The details
The review covers knowledge extraction, spectral and microscopy data interpretation, property prediction, synthesis planning and closed-loop experimentation. The authors frame the paper around the sequence of decisions a chemist makes, rather than around benchmark scores alone.
Timeline
- Article about the review published on phys.org on November 11, 2026.
The Big Picture
The review extends the rapid adoption of foundation models and machine learning in computational chemistry by shifting focus from benchmark scores toward validation and recurring failure modes. Its framing suggests the field's next stage will pair large language models with physics-informed models that respect chemical constraints.
Laboratories adopting AI tools for synthesis planning or property prediction can reduce wasted experiments by first confirming that a model's training data match their own compounds and instruments. The authors' checklist targets information leakage, overfitting and extrapolation errors that can make a model look more accurate than it is in practice.
The takeaway
Before relying on an AI model for chemistry work, check that its training data resemble the compounds and instruments you actually use, since mismatched data can inflate apparent performance. Treat benchmark scores as a starting point rather than proof of reliability.
Further reading
For related coverage, see Chemistry news on AI tools and research methods.
Source note: This article includes information reported by Phys.
