The Role of AI in IPBES Assessments and Biodiversity Literature Analysis - hype or Salvation?

The rapidly growing volume of scientific publications, grey literature, policy documents, research datasets, and materials reflecting Indigenous and Local Knowledge (ILK) and diverse worldviews has outpaced the capacity of expert teams to screen, assess, and synthesise the global evidence base. This challenge is particularly acute for IPBES assessments, which depend on comprehensive, transparent, and globally representative evidence syntheses undertaken by large international, multidisciplinary author teams. As the available literature and knowledge sources expand, so does the risk of overlooking emerging developments, underrepresenting regional, cultural, or knowledge-system perspectives, and placing increasing demands on already stretched experts.

Artificial intelligence (AI)—especially large language models (LLMs)—is often presented as a powerful response to these pressures. AI tools promise to help analyse, filter, and structure vast multilingual bodies of evidence, potentially making it feasible to work with corpora comprising several million references. They appear to offer opportunities for topic clustering, literature triage, metadata enrichment, semantic search, summarisation, and support for drafting specific components of assessments. However, it remains uncertain to what extent these promises can be realised reliably and responsibly in science–policy contexts. Many capabilities are evolving rapidly, but so too are concerns about accuracy, reproducibility, opacity, and the risk of over-reliance on automated outputs.

This talk outlines the current IPBES principles - but also best practices from other communities, e.g., DOME or ARR recommendations - and guidelines governing the use of AI in assessment processes. It highlights the central requirements of transparency, traceability, expert oversight, and methodological rigour when integrating AI tools into evidence-synthesis workflows. Particular attention is given to issues such as confidentiality, risks of bias and hallucination, validation of automated outputs, and the need for reproducible documentation that allows independent verification.

Rather than framing AI as either a universal solution or an emerging threat, the presentation provides a subjective and pragmatic perspective on where AI tools may offer genuine value—and where their limitations warrant caution or exclusion. The aim is to help situate AI within a balanced, responsible evidence-synthesis framework that strengthens, rather than replaces, the scientific expertise at the core of IPBES assessments.