LISA first wants to have the security of data and other information fully in order, before the first UT employees will be allowed to 'officially' use a large language model that the UT itself facilitates, explains Chief Information Security Officer Henk Swaters. He expects the pilot to start later this month. 'Many people were already using artificial intelligence. But not as a product that we facilitate ourselves.'
Access apps and data
The choice initially falls on Microsoft's Copilot365. This software is more advanced than the current free chat variant that employees already have available on their PCs. Not a simple AI chatbot, but software that is integrated with other apps and has extensive access to all kinds of (potentially sensitive) user data. 'That is precisely why it is so important that we create the right preconditions for safe and responsible use,' says Swaters.
Last year, SURF, the ICT cooperative of Dutch education and research institutions, warned of privacy risks when using Copilot. Despite improvements by provider Microsoft, the organisation still sees some privacy risks. No reason for the UT to keep the door shut, according to Swaters. 'There is a thorough data protection assessment in place. In my role as Chief Information Security Officer, I also look very critically at all AI developments, because there are a lot of new vulnerabilities for our cyber security.'
'Better to facilitate than to stop'
Swaters points to the 'great demand from the UT community'. 'People see that developments are going fast. As a university, we cannot close ourselves off from that. That's why we say as a service department: better to facilitate it safely and responsibly than to forbid people from using it.'
The choice for Copilot from the American Microsoft is not a definitive one, according to Swaters. 'No, we have licenses for several providers and keep an eye on developments.' For example, SURF is now committed to EduGenAI and Utrecht University today launched its own chatbot based on ChatGPT. According to Swaters, the UT first wants to gain experience with an existing platform. 'Also in our case, inputs and prompts are not used to train the AI models.'
Own AI model too expensive
And then there is a proprietary AI model, developed by the Digital Society Institute and LISA. According to Swaters, this is not a realistic option for now. 'The current environment has sufficient capacity for research purposes, but is too limited in terms of computing capacity to support a generative AI service for the entire university. Scaling up comes at a significant cost. That's why we chose Microsoft's software,’ says Swaters. 'Given the arguments for digital sovereignty, it may not be the perfect choice, but we don't want to fall behind.'
By the end of this year, the LISA service department hopes to have a good understanding of how UT staff use the AI tool and what the findings are.
