Human-in-the-loop AI in analytical chemistry: a critical review from large language models to agentic systems

Abstract

Artificial intelligence (AI) has moved from a research topic to an operating tool across the chemical sciences, where large language models (LLMs), generative deep learning and agentic systems are now embedded in drug discovery, materials design and mass spectrometry (MS)-based analysis at a pace that has outstripped the peer-reviewed literature. Based on studies published between 2021 and 2026, this critical review argues that AI extends a recent lineage of disruptive tools in chemistry. It further proposes the human-in-the-loop agentic system, in which the chemist orchestrates and validates while the agent executes, as a candidate operative architecture for the coming decade, in preference to autonomous AI. Three converging streams (LLMs as scientific instruments, generative models for chemical design, and agentic systems with tool use and planning) are mapped onto a four-level spectrum of autonomy, from the single-LLM copilot to the autonomous robotic chemist. Practical implementation is illustrated through compound annotation in non-targeted MS, LLM-driven spectroscopic copilots, retrieval-augmented knowledge graphs, and AI coding assistants in routine methods development. On the debates over AI-assisted writing, publisher policies and peer review, it has been argued that transparent disclosure is a more robust path than detection or prohibition. It concludes by examining a possible productivity bifurcation between AI-adopting and AI-resistant groups by 2030 and proposing AI literacy as a core competency for analytical chemists in training.