CGU's MCP Bridge for Brazil's Open Data — Testing Whether LLMs Can Be Trusted with Government Datasets

Brazil's Controladoria-Geral da União (CGU), the federal transparency and anti-corruption body, partnered with the Open Knowledge Foundation (OKFN) and Uruguay's e-government agency AGESIC to test whether large language models can be trusted to answer citizens' questions about official open data.

Test whether LLMs, connected via the Model Context Protocol (MCP) to CKAN (the open-source software powering most public data portals, including Brazil's federal transparency portal), can reliably answer plain-language citizen questions about open government data, and produce a reproducible blueprint other CKAN-based governments could adopt.

In Brazil, the pilot targeted congressional parliamentary amendments data. In a parallel pilot, Uruguay connected the same architecture to its National Energy Balance dataset (energy imports, generation and installed capacity).

The technical connection worked; the harder problem was trust. Answer quality depended less on the pipeline than on how well the underlying data was described (units, field definitions, valid assumptions). In testing on the Uruguayan dataset, the model fabricated "climate factors" as part of its explanation even though no climate data existed anywhere in the source — a concrete, publicly documented hallucination.

The team argues for mandatory domain-expert review of dataset descriptions before any citizen-facing deployment. As of publication the work remains a two-dataset prototype; the next planned phase is user testing with real-world citizen questions rather than production rollout.