SpheriCity: Conversational AI for trustworthy sustainability report analysis

SpheriCity is a new conversational agent designed to help sustainability experts and practitioners make sense of complex city-level circularity assessment reports. Built by researchers from multiple institutions, the system addresses a key pain point: these reports contain rich data on materials, infrastructure, and policies, but their length and inconsistent structure make cross-document synthesis nearly impossible manually. While large language models (LLMs) could accelerate this process, their opaque reasoning, tendency to hallucinate, and lack of source transparency make them untrustworthy for high-stakes sustainability decisions.

To solve this, SpheriCity takes a "provenance-first" approach — every claim is linked to specific sources in the original reports. The system supports exploratory querying, cross-city comparisons, policy summarization, and recommendation-oriented tasks. In a formative evaluation with six sustainability experts, researchers tested representative queries and gathered qualitative feedback. Results showed that transparent sourcing, contextual explanations, interpretability, and alignment with expert workflow were the strongest drivers of trust and perceived usefulness. The paper, accepted at ACM COMPASS '26, contributes the prototype, an evaluation framework, and design insights for building trustworthy AI in knowledge-intensive domains.