University of Michigan Offers Several AI Services to Staff and Students

Overview

In August 2023 the University of Michigan launched UM-GPT, becoming (per the source) the first higher education institution to offer AI services at scale, providing secure access to commercially available large language models through the university's private Microsoft Azure environment. Michigan has since launched U-M Maizey, a codeless platform for building custom AI experiences from uploaded data sets using retrieval-augmented generation, with more than 3,500 Maizey instances in production across academic and operational departments, and Go Blue, a mobile AI assistant for campus life. To avoid vendor lock-in, Michigan runs AI workloads across multiple environments including on-premises infrastructure, letting users pick their LLM via a dropdown.

The challenge

Before OpenAI had introduced enterprise offerings, universities lacked a way to give students and faculty secure, equitable access to commercial large language models without exposing institutional and personal data to public LLM training, or leaving access limited to those who could afford subscription fees.

The solution

In August 2023 the University of Michigan launched UM-GPT, providing secure access to commercially available large language models through the university's private Microsoft Azure environment, with users able to select their LLM via a dropdown menu. The university has since launched U-M Maizey, a codeless retrieval-augmented-generation platform letting faculty build custom AI experiences from their own uploaded data sets (including course-specific AI tutoring integrated with the university's learning management system), and Go Blue, a mobile AI assistant for campus life. To avoid vendor lock-in, Michigan runs its AI workloads across multiple environments, including on-premises infrastructure.

Reported business value

There are now more than 3,500 Maizey instances in production across academic departments and operational offices such as procurement and HR. The approach keeps student and faculty data out of public LLM training models, supports screen readers for visually impaired users, and provides equitable access for students who could not otherwise afford AI subscription fees.