Can AI Language Models Improve How Scientists Assess Water Pollution Risks?

Environmental risk assessment for water pollution faces an information problem that has nothing to do with a shortage of data. The opposite is closer to the truth: decades of monitoring studies, toxicology experiments, ecotoxicological surveys, and regulatory assessments have generated enormous quantities of information about pollutants and their effects on aquatic life. The challenge is that this information exists in thousands of scientific papers, agency reports, and policy documents with no consistent structure, no unified database, and no automated way to draw connections across sources.

A review published in Environmental and Biogeochemical Processes asks whether large language models - the artificial intelligence systems underlying modern text-generation tools - could help solve this problem. The answer the authors offer is a qualified yes, accompanied by a candid inventory of what remains unresolved.

What LLMs Bring to the Problem

Traditional natural language processing tools for extracting information from scientific literature require substantial manual effort. Researchers must define features, train specialized models for each type of document, and adapt them when moving between domains. Large language models trained on vast text corpora operate differently: they can interpret long, complex passages, recognize domain-specific terminology, and identify relationships between concepts across documents without requiring the same degree of manual customization.

For aquatic risk assessment, the relevant capabilities include named entity recognition (identifying the names of chemicals, species, and locations), relation extraction (linking compounds to their toxic effects), and semantic reasoning across documents. These are exactly the tasks that currently require a human expert to read through literature and synthesize findings manually - a process that is slow and inevitably incomplete given the volume of available material.

The review notes that LLMs have demonstrated strong performance in adjacent scientific domains including chemistry, materials science, and biomedical research. Several published systems already use language model components to extract chemical-disease associations from biomedical literature, a task structurally similar to linking pollutants to ecotoxicological outcomes in environmental data.

Aquatic Risk Assessment Has Special Requirements

Moving from biomedical to environmental applications introduces additional complexity. Aquatic risk assessment must account for pollutant transformation in water - compounds break down, react with other substances, and change form as they move through watersheds. It must also consider mixtures: real-world water bodies contain dozens or hundreds of compounds simultaneously, and their combined effects may differ substantially from any single compound's individual toxicity profile.

Environmental monitoring data, unlike clinical trial data, often lacks standardization. Sampling protocols, detection limits, and reporting formats vary across jurisdictions, decades, and research groups. An LLM integrating information from this heterogeneous corpus must either deal with these inconsistencies or risk drawing spurious connections from incomparable data.

The Limitations Are Real

The authors are explicit that LLM applications in aquatic risk assessment remain at an early stage. Three challenges stand out.

First, high-quality environmental corpora for training or fine-tuning domain-specific models are scarce. LLMs trained primarily on general text will underperform on specialized environmental science terminology and reasoning.

Second, language models can generate incorrect associations with confidence - a phenomenon often called hallucination. In a risk assessment context, a false connection between a compound and a toxic outcome could lead to misallocation of regulatory attention or, more seriously, to failure to flag an actual hazard. Expert validation of model outputs is not optional.

Third, the computational resources required to train and run frontier-scale LLMs are substantial. For research institutions and regulatory agencies in lower-income countries - precisely the regions with some of the most serious aquatic pollution problems - this barrier may be significant.

The review argues that hybrid architectures combining specialized environmental datasets, retrieval-augmented generation techniques (where the model is given access to a curated database to consult rather than relying purely on training), and structured expert review workflows could address many of these concerns. Whether that optimism is warranted will depend on the development of better domain-specific resources and more rigorous validation frameworks.

For now, the prospect of LLM-assisted environmental risk assessment represents a genuinely promising direction rather than an imminent capability. The technology's trajectory in adjacent scientific domains suggests the underlying tools will improve. The work of adapting them specifically to the needs of aquatic pollution science is, by the authors' own assessment, only beginning.