‘Traceability is the answer to AI and open government data’

Lucas Pretti: Firstly, how did you get involved in open data, and what does your work in government involve today?

Otávio Neves (Brazil): I have been working on open data in the public sector since 2012. I started out in open government, but it was in 2019 that we moved the agenda to the Comptroller General’s Office (CGU), where it now coexists with the transparency agenda. Our approach is to ‘publish with purpose’: to make available what society actually demands, not just for the sake of compliance. That is why we have such close ties with civil society.

Gustavo Suarez (Uruguay): I also started in 2012, coincidentally, just as we launched the National Open Data Catalogue. I am a systems analyst and today, as well as coordinating open data, I coordinate the support and development teams for solutions and the unified portal (gub.uy). With regard to data openness processes, we are looking for opportunities to enhance them with AI without compromising on quality. We continue to drive the ecosystem forward through activities to promote usage, hackathons and competitions. We are also continuing to work on participatory processes with civil society and academia in Uruguay, and this ensures that these processes run smoothly and are meaningful.

Lucas Pretti: Explain in your own words: what exactly was the traceable AI pilot project we carried out together?

Otávio Neves: In Brazil, we worked with budget amendments – a part of the budget that members of parliament allocate with almost complete freedom (they can, for example, allocate funds to a football club or a concert). Years ago, this was known as the ‘secret budget’ because of its lack of transparency. We already have the Transparency Portal, but navigating it is complicated. We wanted to experiment with presenting that information in natural language, but we knew that standard AI wouldn’t be reliable. The solution offered to us by Open Knowledge through the MCP pilot allowed us to work with models that don’t ‘hallucinate’, and for us, it was the perfect idea.

Gustavo Suarez: In Uruguay, we chose the National Energy Balance (BEN). It has high-quality data and a time series dating back to 1965, as well as a technical team that is very open to innovation. Our top priority was reliability. If the government provides a tool, the answer must come from official sources, not from a generic AI interpretation. That was paramount. So the pilot was a first step towards evaluating a tool that would give us the certainty that it was based on data free from hallucinations.

Lucas Pretti: Do you think traceability is the key to AI tools when it comes to public data? What is still missing?

Otávio Neves: It’s a very special moment. AI enables any citizen to ask complex questions to the data without knowing how to code. Civil society is now more mature when it comes to working with data: businesses are also using more data, and academics are better equipped to use data because they have more tools at their disposal, including AI to assist with analysis. And there are new challenges too: for example, what does ‘open data’ mean when AI can interpret images or videos? The concept of structured data is no longer enough.

Gustavo Suarez: I agree. AI now acts as a translator and intermediary for most people. As for open data, we need to work on improving the quality of publications, datasets and metadata so that they can be interpreted correctly. In the pilot, if the abbreviation for an indicator wasn’t described, the AI became ambiguous. We had to refine the metadata so that the response would be accurate. The lack of high-quality metadata affects the quality of the responses.

Lucas Pretti: What did you learn internally from the pilot? Are there plans to roll it out or replicate it?

Otávio Neves: Firstly, the speed at which the pilot chatbot was rolled out was impressive. For the Comptroller’s Office, the issue of ‘hallucination’ was very, very important to us. Brazil’s Transparency Portal has a very close relationship with the press; any new developments on the portal appear in the newspapers, so the information must be an official government response. We cannot provide made-up answers. Not only will we continue to develop this solution, but we will certainly expand it to other contexts as well.

Gustavo Suarez: We are currently setting it up in a local environment to measure token consumption and feasibility. In addition, we are developing parallel pilots using knowledge graphs to link datasets by region or date. But the most revealing aspect was the cost: a traditional visualisation dashboard takes us a month to develop; with the MCP, within a week we had something functional and much cheaper. It’s a tool that’s here to stay. The next steps will be to test them in a controlled environment, assess their transition to production, and ensure that these tools genuinely serve to improve public access to open data – which is what matters most to us.

Lucas Pretti: There is pressure from Big Tech companies for governments to adopt their AI systems. The same thing happened previously with other areas of the public stack (email, documents, etc.). How do you view the situation where commercial AI providers are putting pressure on governments? And how will you try to prevent that from happening in your departments?

Otávio Neves: It’s very complex. In Brazil, we’ve had both good and bad experiences with open-source software and proprietary solutions. The idea of working with a ‘default’ approach – where everything has to be proprietary or everything has to be open-source – wasn’t the best. Realistically, we need to plan for the long term. Ideally, we would have robust, high-quality open-source solutions that stand the test of time, but realistically, we’ve seen that they require a great deal of maintenance. We must maintain a critical perspective. The most important points are autonomy and sovereignty. In the meantime, we must ensure that our official data also feeds into these commercial AIs so that, at the very least, the information reaching the public is accurate.

Gustavo Suarez: Technology changes so quickly that you always feel as though you’re playing catch-up. At AGESIC, we have an AI strategy in place, a data strategy and a digital citizenship strategy that aims to strengthen people’s capabilities so that they are at least able to understand that not everything AI says is valid. It’s a huge challenge. You need to be flexible to adapt the tools to the constant changes taking place. The advantage of the MCP is its flexibility: you can change the underlying language model without having to rebuild everything from scratch. In Uruguay, we’re collaborating on regional projects such as LatamGPT to develop alternatives that are closely aligned with the regional context.

Lucas Pretti: This joint pilot was designed with replicability in mind: the traceable AI solution is ready for adoption at different levels. How do you envision regional collaboration on AI and the possibility of continuing to exchange lessons learnt with other countries?

Otávio Neves: It is highly scalable and essential. Issues such as climate change, public safety and trade have no borders. Brazil already integrates data from states and municipalities; the next logical step is the international integration of portals. My dream is that by 2027 we will have a regional initiative in Latin America focusing on climate data.

Gustavo Suarez: When we show the pilot to other organisations and teams, the response is ‘I want that’. Because it gives them another quick way to access the data, where users can ask questions and get answers in natural language. Public bodies already see it as something that will really be of use to them. Presenting it to Abrelatam/Condatos will be key to assessing whether it is a solution that other countries would want to replicate. It is a cross-cutting solution, applicable to energy, transparency, climate, etc., providing the certainty that the answer comes from an official source. Collaboration with other countries in the region could be a next step that should follow naturally.