Generative AI in Rijksoverheid: direction, opportunities and frameworks

First, Yannick God, AI and Algorithms coordinator at CIO Rijk, explained the government-wide position on genAI: "Since 2023, the Rijksoverheid has taken a preliminary position. In early 2024, we launched a government-wide vision. In that year, we also conducted many pilots and experiments in governments. Based on these, we developed our position further. This includes room for innovation and attention to the opportunities offered by technology. Since the beginning of this year there has been intensive coordination on this government-wide position."

Government-wide position: what is it?

The question, of course, is what this government-wide position is, God continued, "You have a first, because the position has not yet been officially released." God then listed the key conditions that apply to all government agencies procuring, developing and/or using genAI. "The deployment of genAI complies with existing laws and regulations, the purpose of the deployment is clearly and transparently formulated and a risk analysis has been conducted. Furthermore, necessity and proportionality have been substantiated. For example, what problem is being solved that cannot be solved in any other way? Are the preconditions, such as AI literacy and governance in place? Is an exit strategy available to be able to stop using genAI in case of problems, so that security is ensured? Furthermore, the deployment of genAI must fit within and align with relevant policy context. Finally, genAI is not allowed under consumer conditions, such as ChatGPT through a browser."

Per pilot proper consideration

One attendee asked how hard the government-wide position is, for example, on a pilot. "Anyway, a pilot must comply with laws and regulations and a risk analysis must be conducted," God said. "But otherwise it depends on the context and the deployment. We have to look at a pilot not generically but pragmatically and make a good trade-off between the business value on the one hand and the risks on the other."

DefGPT: the AI chatbot of Defense

Next to speak was Sofie Berns, DATA department head at the Ministry of Defense: "Our department makes tools, standards and a secure and managed platform available and translates promising innovations into concrete results. Our AI chatbot DefGPT is one such result: an internal alternative to popular AI services such as ChatGPT, which has allowed our employees to perform complex language tasks based on massive amounts of data since November 2024. DefGPT is specifically trained on defense intelligence and tasks."

Sensitive information

Previously, ChatGPT was used at Defense thousands of times a day, Berns estimates: "That carries great risks. Sensitive information can end up on the street. This gives your opponent an information advantage. In these times of geopolitical tension, our business operations take into account threats and weigh risks differently. A closed, secure environment like DefGPT is crucial to that."

Man and machine reinforce each other

The Ministry of Justice and Security (JenV) has a similar internal "AI assistant" with the call sign Robin. "With this alternative to ChatGPT, we want to be completely secure, reliable and in control," says project leader Joost de Haan. He emphasizes that man and machine do not replace each other, but rather reinforce each other. In developing Robin, JenV is working with partners such as the Immigration and Naturalization Service (IND) and the Central Judicial Collection Agency (CJIB). The beauty of Robin, according to De Haan, is that you can upload files and query the machine based on those files. And there is more. "We capture specific knowledge in special tools, so-called chatbot plugins. Because Robin can use these plugins we add extra functionality to the machine."

Critical success factors

For the development of Robin, De Haan mentions 5 critical success factors: "1st, we have our own ICT infrastructure and the JenV Trusted Cloud. These support our ministry in the flexible, secure and efficient deployment of ICT resources and fully meet our strict security and compliance requirements. 2nd, we work in multidisciplinary teams with both technicians and non-technicians. 3rd, we have a common goal of a multi-deployable application. The 4th success factor is focus: we require an availability of three days per week per person. Finally, at production we immediately set up our DevOps processes and ensure that our builders and administrators work together through one streamlined process."

Read the entire article on the website of Digitale Overheid.