The challenge
VITO, the Flemish Institute for Technological Research, a trailblazer in cleantech and sustainable development, is constantly looking for innovative solutions. Their latest challenge? Streamlining the hefty task of legal document updates — a routine that was not just time-consuming but also a drain on resources. Now, they're stepping up their game with GPT-4 to revamp how they update legal texts in a tailor-made application.
One of the services VITO offers is keeping the Navigator environment up-to-date. Thousands of people, from municipal employees to lawyers, use this website to consult the current legislation on the environment, nature, spatial planning, and energy in Flanders. Their challenge in keeping this environment up-to-date? Every day, the government makes minor changes to the legal texts, sometimes rewriting just a sentence or a single word. This requires VITO to review the Official Gazette (Staatsblad) every day, identify the changes, and then add adjustments to the database.
This process is not only expensive but also time-consuming, partly because there is no uniform structure in how legal texts are written. At VITO, they wondered if it was possible to optimize this process with the help of AI. We were brought in to examine the possibilities of generative AI in this context, thanks to Cronos Public Services.
“It seemed interesting to investigate how far we could go to automate this labor-intensive process with AI. This frees up more time for our employees for more complex, challenging tasks.” - Karel Styns, ICT Project Coordinator at VITO
OpenAI's GPT-4 was the tool we needed to tackle this project, as it's known for its ability to generate coherent texts and perform complex language tasks. In this case, it is suitable for interpreting legal texts and suggesting adjustments. The language model identifies the indicated changes in the legal texts, searches for the relevant paragraphs within the specific articles, and makes the necessary adjustments. By analyzing and interpreting the text, it can propose the correct changes in a few seconds. VITO can quickly review and accept or reject the adjustments suggested by GPT-4.
“A human error is quickly made. If we can improve both the speed and the quality of a process thanks to generative AI, that is a huge asset.” - Karel Styns, ICT Project Coordinator at VITO
Language models like GPT-4 are highly accessible to the public. Still, in this case, a skilled prompt engineer with the proper knowledge is needed to deploy the AI as effectively and efficiently as possible. Especially since a large language model (LLM) can make mistakes, our prompt engineer needed to test the flow through. They run examples through the application and quickly identify if there is a problem, which can, in turn, be immediately resolved. Since this case concerns legal texts, it is crucial that the application does exactly what is expected and does not adjust texts it shouldn't.
“If you know what's behind such a language model, you know better than anyone how to ensure that the model does exactly what you expect of it. For us, it's a long process of trial and error, for the AI experts at Raccoons, it's a piece of cake.” - Karel Styns, ICT Project Coordinator at VITO
Furthermore, our experts ensure the application does not 'hallucinate' — in other words, generate incorrect information, make wrong assumptions, or invent events, data, or statements not based on truth. Because of these hallucinations, an LLM could sometimes provide answers that may sound convincing but are factually incorrect. Moreover, we have to protect our application from prompt injection. Prompt injection is a technique where someone (un)intentionally inputs data to manipulate the AI model. We must protect the model from this misleading so it does not generate inappropriate, incorrect, or harmful content.
VITO will integrate our tool into its daily work processes in the coming months. This will enable the organization to respond quickly and accurately to changing legal requirements, significantly improving its service and strengthening its position as an innovative player. But that's not all. The organization is full of ideas for new AI applications, such as automatic translations or accessible legal texts. To be continued!
“The collaboration with Raccoons went very smoothly and pleasantly. Because we received regular updates and demos, we were able to switch quickly and completed the project within the planned time.”
