Undergraduate beginnings
Initial group projects at Cambridge explore whether AI can help screen the vast conservation literature.
Conventional systematic reviews are slow and costly, and they start going out of date the moment they are published. Yet policymakers often need current, context-specific evidence within narrow windows. Evidence TAP replaces one-off reviews with living evidence databases that keep updating as new research appears.
At its core is a traceable AI pipeline. It ingests the literature across many sources and languages, screens it for relevance, appraises study design, and extracts structured data. Every output remains traceable to its original source. The pipeline runs on local, self-hosted models, pairing keyword and semantic retrieval with a statistically principled stopping rule.
97% recall against a large-scale manual review, in our flagship study.
The pipeline adapts to the decision at hand. For urgent questions it can synthesise all of the available evidence rapidly with minimal human checking, flagging gaps for follow-up; where the stakes are higher, experts verify each stage. Every verification is retained and feeds back into the models, so the system keeps improving. The trade-off between speed and accuracy is transparent and quantifiable.
It began in conservation, through Cambridge's Conservation Evidence collaboration, and is now being applied to education. Health, climate and other fields will follow. The longer-term aim is a global mesh of self-hosted nodes that shares evidence equitably across and within countries.
Evidence TAP grew out of the Conservation Evidence Copilots project and has broadened, discipline by discipline, into a general living evidence pipeline.
Initial group projects at Cambridge explore whether AI can help screen the vast conservation literature.
Early preprints show that carefully designed pipelines can reach expert-level retrieval, while off-the-shelf LLMs fall short. The project is selected as an ai@cam flagship challenge.
A working paper sets out a self-hosted, end-to-end pipeline, which in an initial evaluation reached 97% recall against a large-scale manual review.
The project broadens beyond conservation into education, with health and climate to follow, and becomes Evidence TAP.