17 February 2025 in Environment International
Concerns about the negative impacts of microplastics on human health are increasing in society, while exposure and risk assessments require high-quality, reliable data. Although quality assurance and -control (QA/QC) frameworks exist to evaluate the reliability of data for these purposes, manually assessing studies is too time-consuming and prone to inconsistencies due to semantic ambiguities and evaluator bias. The rapid growth of microplastic studies makes manually screening relevant data practically unfeasible. This study explores the potential of artificial intelligence (AI), specifically large language models (LLMs) such as OpenAI's ChatGPT and Google's Gemini, to streamline and standardize the QA/QC screening of data in microplastics research. We developed specific prompts based on previously published QA/QC criteria for the analysis of microplastics in drinking water and its sources, and used these to instruct AI tools to evaluate 73 studies published between 2011 and 2024. Our approach demonstrated the effectiveness of AI in extracting relevant information, interpreting the reliability of studies, and replicating human assessments. The findings indicate that AI-assisted assessments show promise in improving speed, consistency and applicability in QA/QC tasks, as well as in ranking studies or datasets based on their suitability for exposure and risk assessments. This groundbreaking application of LLMs in the environmental sciences suggests that AI can play a vital role in harmonizing microplastics risk assessments within regulatory frameworks and demonstrates how to meet the demands of an increasingly data-intensive application domain.
8 August 2022 in Restoration Ecology
Active restoration is frequently implemented to restore degraded drylands globally, yet predicting the overall success of these restoration projects remains challenging. Here, we aim to explore if vegetation spatial patterns can be used to monitor ecosystem recovery and anticipate the success of restoration practices. We combined field surveys and high‐resolution drone images to analyze how vegetation spatial patterns change after the large‐scale straw checkerboards restoration projects in the sand dune systems of the Tengger Desert in northern China. We found that vegetation cover, plant species richness, and diversity increased rapidly, approaching the level of the naturally vegetated sand dunes, after 5 years of restoration. Soil fertility remained low despite the positive vegetation change. Along with the recovery process, we found a larger diversity of patch sizes (i.e. spatial variance) with an increasing proportion of large‐size patches. These patterns, combined with a constant positive recovery rate in vegetation cover, are consistent with theoretical predictions used to anticipate transitions to alternative ecosystem states. Although some indicators may be masked by the artificial planting scheme and community succession process during recovery, our results show that vegetation spatial patterns can be used to forecast and monitor the possible recovery of drylands.
25 August 2026 in SSRN Electronic Journal
Reliable exposure data is critical for determining the potential human health risks of microplastics. To date, however, exposure data are often difficult to compare and do not always meet quality criteria, limiting their suitability for risk assessment. This study aimed to provide a large-scale quality assessment (QA/QC) of studies on microplastics in the main human exposure routes: Air, Food, and Beverages. A hybrid method was used, in which an AI model (Gemini 1.5-pro) automatically extracted information from 950 studies, followed by logical score assignment using Python scripts. Building on earlier QA/QC screenings, each study was evaluated against 13 criteria and assigned a score of 2 (fully reliable), 1 (reliable with restrictions), or 0 (unreliable). AI performance was verified by human assessment in a statistically determined sample of studies (n= 87). The average total score across all datasets was 12 out of 26, with an average of 5.3 inadequate scores (zero scores) per study. The highest scoring studies achieved 20 points. The lowest scores were associated with criteria involving positive and negative controls, blank correction, and the representativeness of sampling and usage conditions. The results show that most current studies have significant shortcomings for application in risk assessment. In particular, data on Food and Air may not adequately represent actual human exposure. At the same time, this QA/QC analysis is the largest to date and the first to cover Air, Food, and Beverages simultaneously; it provides a valuable ranking of usable studies and makes large-scale risk assessment more feasible. The dataset can assist policymakers in selecting the most reliable data for health evaluations and underscores the need for standardization and improved analytical methods in future studies.
18 October 2026 in Environment International
Determining human microplastic exposure from external intake estimates combined with biodistribution modelling remains highly uncertain because errors propagate across multiple pathways and assumptions. We therefore present a framework for estimating internal human microplastic exposure directly from published measurements in organs, tissues, and bodily fluids. The framework integrates dataset selection, QA/QC screening, and an updated mathematical alignment method incorporating statistical fitting, goodness-of-fit testing, and comparison of alternative particle-size distributions to support cross-compartment comparison and whole-body inference. As a proof of principle, we applied this approach to published human microplastic particle data under two scenarios: an 'as measured' scenario and a 'biologically plausible' scenario in which particle size distributions were truncated at literature-based accessibility limits. The scenarios produced strongly divergent mean estimates: 159.4 mg (80.9-824; parametric 95 % Monte Carlo interval) under the 'as measured' scenario and 0.41 mg (0.22-1.56; parametric 95 % Monte Carlo interval) for the truncated 'biologically plausible' scenario. This 388-fold divergence shows that current quantitative inference is dominated by assumptions about the treatment of unusually large particles and data quality. The values should therefore be interpreted as scenario-bounded framework outputs conditional on the currently available, clinically biased dataset, and not as estimates representative of healthy populations or the general population. The framework also yielded apparent, size-dependent organ-to-blood ratios that may serve as provisional benchmarks or hypothesis-generating boundary conditions for future physiologically based kinetic (PBK) model development, but not yet as validated tissue-partitioning constants. Although current data quality limits regulatory-grade quantification, the framework provides a transparent basis for exposure assessment using human monitoring data, for identifying key uncertainties, and for illustrating potential future regulatory application.
8 July 2026 in SSRN Electronic Journal
Mismanaged urban waste is a persistent challenge for dense cities. Effective intervention depends on knowing where and when waste accumulates across the urban landscape. However, existing monitoring approaches remain labor-intensive, costly, and uneven in spatiotemporal coverage, often limiting consistent city-wide assessment. Street-view imagery (SVI) offers a scalable complementary data source by providing repeated, georeferenced, street-level observations across entire cities over multiple years. Nevertheless, most large-scale image-analysis pipelines rely on annotated datasets and task-specific model training, which can be costly and difficult to transfer across cities or applications. Here we show that off-the-shelf multimodal large language models (MLLMs) can support scalable, categorical urban waste monitoring directly from SVI without task-specific training. Using the Municipality of Amsterdam’s street-view panorama archive from 2016 to 2023, we compared GPT and Gemini variants using a 2,000-image manually annotated reference set and applied the most cost-effective model to 364,676 image-level observations across the historical inner city (8.04 km²). We further linked waste distributions to urban-environment indicators to explore potential correlates. Results revealed persistent but category-specific hotspot structures: litter concentrated in visitor-oriented central areas, whereas garbage bags and cardboard aligned more strongly with residential structures. We also identified a distinct redistribution signal in 2020, when central litter hotspots weakened while residentially associated waste signals persisted. These results show that general-purpose MLLMs can support city-scale waste monitoring. More broadly, MLLM-assisted SVI analysis could help cities develop more adaptive, data-driven approaches to street cleaning, waste management, and urban environmental governance.