AI Meets Human Expertise to Map How Mixed Forests Deliver Ecosystem Services

Forestry has long been a late adopter of information technology, but a new study from researchers at the University of Valladolid and the technology company SNGULAR suggests that the field may now leap to the front of the generative artificial intelligence wave. In a paper published in Discover Artificial Intelligence, Felipe Bravo, Irene Ruano, Diego Martín and Cristóbal Ordóñez describe a hybrid human–AI methodology for conducting scoping reviews, the labor-intensive literature syntheses that scientists use to map what is known about an emerging topic. Their test case is one of the hottest questions in forest ecology: how the functional traits of trees in mixed-species forests influence the delivery of ecosystem services such as biomass production, carbon storage and biodiversity support.

The stakes are considerable. Scoping reviews allow researchers to identify key concepts, theoretical frameworks and knowledge gaps by systematically charting the available literature, and they often determine whether a full systematic review is worth undertaking. Yet the sheer volume of published science makes comprehensive coverage increasingly difficult, particularly in interdisciplinary fields like forest ecology where relevant evidence is scattered across agronomy, ecology and silviculture. The Spanish team set out to determine whether a carefully adapted large language model could shoulder part of that burden without sacrificing the rigor that scientific synthesis demands—a question they frame as urgent, since using generative AI in literature reviews has been categorized as a high-risk practice for research integrity.

Their approach rests on a technical architecture known as Retrieval-Augmented Generation, or RAG. Rather than letting a general-purpose model such as GPT-4o answer from its pretraining alone, the researchers constrained the model to draw exclusively on a curated knowledge base of 33 peer-reviewed papers, stored in a SharePoint library and retrieved through semantic search. The RAG process unfolds in three phases: retrieval, in which the system identifies the most relevant text fragments for a given query; augmentation, in which those fragments are injected into the model’s prompt as supplementary context; and generation, in which the model composes its answer using only the retrieved material. This design dramatically reduces the risk of hallucinated citations and invented findings, because every claim is anchored to a verified source document.

Building the knowledge base was itself a deliberately human enterprise. The team searched Scopus and Web of Science using a structured query that required abstracts to mention trees or forests, to include the terms mixed, mixture or mixing, and to contain the word overyielding, while excluding tropical forests and restricting results to English-language articles published between 2020 and 2024. After removing duplicates and screening out an off-topic genetics paper, 33 documents remained. For each one, the researchers manually extracted the functional trait analyzed, the biome and location, the methods, the ecosystem services addressed and the observed outcome, creating a structured map of the literature that temperate European forests dominate, with Asian studies concentrated in Chinese and Japanese plantations and American work spanning Canada, Colombia and the United States.

The AI agent itself was built with a low-code approach using Microsoft Copilot Studio, a choice the authors emphasize for its accessibility. Constructing a RAG system from scratch demands data scientists, engineers and significant funding—resources rarely available in forest science labs. The low-code route required minimal customization and can be operated by the researchers themselves. The agent’s behavior was defined through a meta-prompt following the PERFECT framework, which specifies the prompt’s Purpose, Elements, Role, Format, Examples, Conditions and Timeframe. This meta-prompt established the agent’s role as a scientific reviewer, its professional tone, its obligation to stay within the curated corpus, and its required output structure based on the PICOT(S) scheme borrowed from clinical trial design, which characterizes Population, Intervention, Comparison, Outcome, Time and Study type.

A second, subtler layer of the design involved grounding. Beyond the RAG knowledge base, the researchers supplied a separate library of three authoritative reference works on mixed-species forests. These grounding documents are never cited in the agent’s responses, but they are available to the model in every iteration, allowing it to acquire the vocabulary, processes and conceptual relationships of the domain—to, in effect, think like an expert. The authors compare this to traditional fine-tuning, but at lower cost and with greater control, and they argue it improves the agent’s ability to disambiguate, prioritize and synthesize information in line with the project’s technical context.

To test the system, the team posed twelve instructions to the agent, covering six research questions about how mixing species with different functional traits affects ecosystem service provision. Half the prompts used a zero-shot approach—concise questions with no added detail—while half used refined-shot prompts tailored with specifics about the species identities in the analyzed forests. The results were revealing. Zero-shot prompts achieved high precision, with a mean success rate of 0.89, meaning the agent usually selected the correct articles, but suffered from low sensitivity, with a mean of 0.36, missing many relevant papers. Refined-shot prompts reduced omissions across all objectives but tended to increase the failure rate, sometimes pulling in related but non-germane articles.

The most striking finding concerned the economics question. For the topic of economical return, the refined prompt produced a complete retrieval failure, with a success rate and sensitivity of zero, apparently because the wording conflated biomass production with monetary terms and generated a query that matched nothing in the corpus. This demonstrates a crucial technical insight about RAG systems: prompt refinement affects not only how information is summarized but which documents are retrieved in the first place. A more specific prompt can fail to surface relevant papers that use different terminology. The authors conclude that prompt quality matters, but that the meta-prompt and retrieval configuration may matter more, and that each research question requires its own careful evaluation of prompt wording.

Validation came from a three-way comparison. The agent’s outputs were checked against the authors’ expert assessment and independently against the commercial model Claude, which was queried about each cited article without access to the agent’s knowledge base or grounding documents. All three evaluations agreed completely, supporting the team’s first hypothesis that a RAG-based language model can produce outputs valid for scientific scoping reviews. The main limitation, they found, was article retrieval rather than the correctness of retrieved information—the agent knew what to say about the papers it found but sometimes failed to find all the papers it should have.

The broader implications extend well beyond forestry. The methodology is explicitly transferable to other scientific domains, and it arrives as commercial tools such as Scopus AI and LeapSpace begin to mediate researchers’ access to literature—tools that, unlike the customizable agent described here, offer limited opportunities to inspect or modify the retrieval layer and meta-prompt. The authors call for interdisciplinary collaboration among computer scientists, Earth domain scientists and AI specialists to build accessible, domain-specific RAG systems at reasonable cost. They also acknowledge open questions: the environmental footprint of generative AI, the variability of model outputs across sessions, the exclusion of foundational older papers imposed by the curated time window, and the need for systematic experiments on how different levels of human expertise affect reliability. What the study demonstrates, however, is a practical middle path—neither full automation nor wholesale rejection—in which machines do the heavy lifting of retrieval and drafting while human scientists retain control over corpus selection, system configuration and final judgment. For a discipline that manages ecosystems on century timescales, that partnership may prove the most consequential mixing effect of all.

Subject of Research: Human–AI hybrid methodology using retrieval-augmented large language models for scoping reviews of ecosystem services in mixed forests

Article Title: Integrating human and artificial intelligence for conducting scoping reviews on ecosystem services in mixed forests

Article References: Bravo, F., Ruano, I., Martín, D., & Ordóñez, C. (2026). Integrating human and artificial intelligence for conducting scoping reviews on ecosystem services in mixed forests. Discover Artificial Intelligence, 6(1), Article 1438. https://doi.org/10.1007/s44163-026-02464-w

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02464-w

Keywords: generative artificial intelligence, large language models, retrieval-augmented generation, scoping review, mixed forests, plant functional traits, ecosystem services, overyielding, forest ecology, prompt engineering, research methodology, GPT-4o