On 10 November 2026, researchers from the Skolkovo Institute of Science and Technology (Skoltech), alongside colleagues from the Zelinsky Institute and ITMO University, published a comprehensive evaluation detailing the transformative impact—and the critical limitations—of artificial intelligence in chemical research. Their findings serve as a necessary reality check for a scientific community increasingly eager to hand the reins of discovery over to algorithmic models, clearly delineating where machine learning provides genuine breakthroughs and where its reliability fundamentally breaks down.

In recent years, the promise of generative artificial intelligence has captivated the physical sciences. Driven by the urgent need to discover novel materials for energy storage, advanced manufacturing, and new therapeutic compounds, laboratories have increasingly turned to computational models to filter through vast chemical spaces. The well-supported benefits of this approach are evident: algorithms can process multidimensional data sets, optimise existing chemical reactions, and propose molecular combinations at a speed and scale that human researchers could never match manually.

However, the transition from simulated environments to the physical test tube is precisely where the friction occurs. The joint assessment highlights that current AI systems often lack a grounded, rule-based understanding of physical and thermodynamic constraints. When models are trained on historical, sometimes incomplete or biased datasets, their predictive accuracy drops significantly when they are forced to navigate unexplored chemical territory. This limitation creates a profound vulnerability. A model may confidently recommend a chemical synthesis pathway that appears mathematically sound on screen but relies on highly unstable or physically impossible intermediate states.

The implications for scientific stewardship and laboratory safety are substantial. If researchers blindly trust AI-generated pathways without employing rigorous physical validation and containment tests, the consequences span from expensive, wasted laboratory resources to genuine physical hazards. For research institutions in Aotearoa and the broader Asia-Pacific region—many of which are currently investing heavily in automated biotechnology and agricultural tech—these specific failure modes require careful operational attention. The scientific community cannot afford to embed algorithmic hallucinations deep into foundational literature.

To harness this technology safely, the researchers advocate for a fundamental recalibration of how AI is positioned within the scientific pipeline. Rather than viewing machine learning models as autonomous decision-makers capable of directing end-to-end research, they must be treated as advanced, yet fallible, analytical instruments. Moving forward, establishing robust, independent testing frameworks to catch model drift and verify chemical predictions will be just as important as the initial AI-driven discoveries themselves.

Sources


AI disclosure: Automated systems assisted with research and drafting. Gemini independently peer-reviewed this article before publication. Sources are linked for verification.