Aicacia Lab

Support global reforestation goals by building a trustworthy, transparent AI system that helps practitioners access, interpret, and apply ecological knowledge responsibly.

Effective reforestation depends on practitioners having timely access to proven best practices, contextual ecological knowledge, and region-specific guidance. While a large body of high-quality restoration literature exists, it is often fragmented across multitudes of documents, highly technical, and difficult to navigate, creating a significant barrier to real-world decision-making.

AI offers the promise of synthesizing large knowledge bases into accessible answers. However, applying AI to ecological decision support introduces epistemic risks: incomplete or biased knowledge bases, loss of context during summarization, over-confident or misleading outputs, and evaluation metrics that prioritize surface-level relevance over decision quality.

Example of use with a current state of the tool after adding in a few more in scope sources from WRI.

Aicacia addresses both the access challenge and the epistemic risk challenge. Rather than treating AI outputs as authoritative, Aicacia is designed to support human judgment by emphasizing evidence provenance, uncertainty, and failure awareness.

Aicacia is a domain-specific search and question-answering system built on curated reforestation literature and expert-vetted sources. Its core goals are to:
‍
1. Catalog and organize reforestation knowledge in a transparent, structured, and user-friendly way, preserving links to original sources and context.Enable evidence-grounded search and Q&A, helping practitioners surface information relevant to specific geographies, ecosystems, and restoration interventions.
‍
2. Enable evidence-grounded search and Q&A, helping practitioners surface information relevant to specific geographies, ecosystems, and restoration interventions.

While off-the-shelf language models can support generic search, they fall short in ecological domains where nuance, uncertainty, and local context matter. Aicacia develops and evaluates domain-specific models and datasets, trained on expert-vetted, curated sources, with an emphasis on interpretability and responsible use.

Embedded in our work on this project is research into epistemic risk in AI-assisted decision-making. We study where common evaluation metrics (such as retrieval relevance or RAG-based scores) fail to capture real-world usefulness, and we develop qualitative failure taxonomies to better understand how AI systems can mislead, oversimplify, or omit critical ecological considerations.

Wherever possible, Aicacia’s datasets, models, and research findings are released openly, with the goal of contributing not just a tool, but a framework for building safer, more reliable AI systems for environmental decision support.