How (and Why) the University of Michigan Built Its Own Closed Generative AI Tools

Case Study

min read

To address privacy, accessibility, and affordability issues, the University of Michigan decided to develop closed generative AI tools for use by its campus community.

Case Study
Credit: Muslianshah Masrie / © 2024

Institutional Profile

The University of Michigan (U-M) is a public research university with campuses in Ann Arbor, Dearborn, and Flint, as well as a medical center in Ann Arbor. Founded in 1817, the university is the oldest in Michigan and a founding member of the Association of American Universities. U-M offers more than 250 degree programs through 19 schools and colleges. In fall 2023, the enrollment was 52,065, including 33,730 undergraduates.

The Challenge/Opportunity

Informed by the recent findings of a campus committee studying generative artificial intelligence (AI), the IT leadership at University of Michigan confronted an interesting tension in July 2023.

On the one hand, the exciting tool of the moment — ChatGPT from the San Francisco startup OpenAI — raised a host of privacy, accessibility, and affordability questions that a public institution of higher education has to account for. For example, data from queries on ChatGPT is shared back with OpenAI. Additionally, the product wasn't fully compliant with the web accessibility technical guidelines U-M uses, which meant it wasn't ideal for the screen reader technology that many students rely on.

On the other hand, everyone in the IT leadership agreed that a major research university needed to make the latest AI tools available to all those who will inevitably want to explore it: researchers, instructors, student support professionals, operational staff, and students.

"Our vision at the University of Michigan is to be a leader when it comes to generative AI," said Ravi Pendse, U-M's vice president for information technology and chief information officer. "Not only for our peers, but also for the rest of the world. We want to show how you can provide fair, responsible, and forward-thinking access to this kind of technology for an entire community."

The U-M Generative Artificial Intelligence Advisory Committee — convened by the U-M provost and Pendse and comprising 20 faculty, students, and staff members from various disciplines — also outlined how ChatGPT is offered in more and less powerful versions at different costs, which unfortunately had the potential to further inequality among students.

The committee released its report on June 30, 2023, detailing the concerns about privacy, accessibility, and affordability, among other issues.Footnote1 One of the recommendations was to "create, accommodate & deliver" – "as soon as feasible" – basic consumer and advanced research generative AI tools that are secure and equitably accessible.

Figure 1. Summary of Overarching Findings
(I. Strategic Benefits & Opportunities | A. Advancing mission and accelerating vision 2034. B. Discovery: Accelerating, redefining disciplines and defining new ones. C. Efficiency / Effectiveness: Doing what we now do better and faster. D. Leadership: Being the North Star for peer institutions. ) come with  (II. Weaknesses, Threats, Risks | A. Technical: Temporal ignorance, hallucination, mis-attribution, etc. B. Social: Plagiarism, need to re-evaluate assignments, learning outcomes, etc. ) which may be mitigated by  (Ill. Policies, Guidance, Regulation, Compliance | A. Policy: Modification of new SPG, course design, syllabus language, etc. B. Guidance: Efficacy of assignments, learning goals, assessments, etc. C. Regulation: Truthful documentation of use, HIPAA, FERPA, etc. D. Compliance: PEERS training. certification, etc.)  which will require  (IV. Resources & Costs | A. GenAI system costs: Procurement & in-house development.  B. Cloud hosting costs: recurring and likely externally provided.  C. Data pipeline and infrastructure costs.  D. Other: Resource and environmental ( emissions, etc.) )
Credit: Generative Artificial Intelligence Advisory Committee Report. Reprinted with permission.

As Pendse said: "The report was a great start, but we did not want to end our efforts there. That is not the Michigan way. Michigan leads. That is when we decided that we needed a sandbox of our own to experiment with these AI tools."

Pendse and U-M's Information and Technology Services (ITS) set out to develop custom AI tools for the U-M community to use in lieu of emerging public products from OpenAI and others. "There are a lot of people playing in this space, but we didn't see anyone marching forward with inclusion, equitability, and accessibility at the forefront," said Robert Jones, ITS executive director of emerging technology and support services. "That's how we thought we could make a difference."

They envisioned a suite of tools to serve the broad and in-depth needs of a large research university:

  • A Q&A tool similar to ChatGPT
  • A no-code platform that allows different units around campus to develop their own AI-based tools for specific contexts
  • A toolkit with which developers and researchers can build, train, and host their own AI models and environments

They knew a key part of their approach would be "closed AI." Instead of queries being used to continually train and tune the AI, information would be kept private, and the language model the U-M tools relied on would be updated and tuned by ITS staff.

As the advisory committee's report also noted, the challenges that generative AI presented were imminent, since the community was already using OpenAI products. U-M leaders decided that they wanted students, faculty, and researchers who would be arriving on campus in fall 2023 to find a set of generative AI tools waiting for them with three advantages: utilized closed AI methods that protect privacy, were free to use, and had accessible design. They gave themselves a deadline of August 21 to build the tools.

The Process

The project was unique, said Jones, since it brought together both vendor resources and internal software development to build custom AI services that met the specific needs of the U-M community.

Members of the ITS Emerging Technology team worked through the summer to create and test U-M's AI tools. They achieved this in part by licensing from Microsoft to use its large language model within a walled-off Michigan environment so that queries made by their users are not shared with OpenAI.

This allowed them to deploy other language models, including any developed at Michigan. "You'll notice I didn't say 'large language,'" Pendse said. "If you have verified data, you can get equally accurate results with a much smaller model."

The team was made up of about a dozen people, including two front-end developers, a full-stack developer, and several other supporting roles. "The other executive directors in my organization were incredibly supportive and generous," Jones said of the ambitious timeline. "The senior leadership established an outstanding team. There were a lot of unknowns about whether we could even do what we wanted to do. It was a big idea with no precedent, but there was commitment. If there was a question without an answer, we were going to figure it out."

Outcomes and Lessons Learned

On August 21, U-M did have three unique generative AI tools ready for returning students and employees.

  1. U-M GPT is the tool that most resembles ChatGPT. It is able to answer questions, produce written content, and make recommendations. Additionally, U-M GPT supports multiple commercial and open-source language models and AI art generators, broadening its utility and applications. U-M GPT was also designed to work seamlessly with screen readers, making it more accessible than ChatGPT. U-M GPT is free for all members of the U-M community.
  2. U-M Maizey is a no-code platform that allows users to build unique and customized chat programs by using their own datasets in combination with U-M's AI language models. Users design context-specific AI experiences that they then share with their audience in the U-M community. Individual departments, units, or projects pay for U-M Maizey through the university's departmental charge system.
  3. U-M GPT Toolkit is designed for AI developers who require full control over the AI model and environment that they are building, training, and hosting. Researchers and developers who want to use the U-M GPT Toolkit must contact the ITS AI team for access.

All three tools are currently approved for "moderate sensitive" data. This means they can be used with information covered under the Family Educational Rights and Privacy Act, but they cannot yet be used with highly sensitive data, such as protected health information.

The Response from the U-M Community

U-M's AI tools currently have an average of 15,000 users a day, including those who are using it as part of coursework. For example, English instructors have asked students to use U-M GPT to create the first draft of an essay and then to revise the draft in class.

U-M Maizey—the no-code platform for creating unique chat tools—has already been used in a wide range of projects:

  • Student advising in the College of Literature, Science, and the Arts is creating an AI academic advisor to handle frequently asked questions.
  • Procurement is creating an AI-enabled platform that will help users easily complete requests for proposals.
  • ITS built a chatbot for Student Life to help students explore more than 1,500 student organizations at U-M.
  • The U-M Library is creating a natural-language book-recommendation service. U-M students will be able to ask questions such as "What are some of the best books in the library about early Netherlandish paintings?" or "What books will teach me about social epidemiology?"
  • Two Ross School of Business faculty members, Andrew Wu and Jun Li, built an AI tutor for their courses. Informed by past course data, it is currently serving more than 1,000 students. The AI tutor was tested on qualitative and quantitative questions and was found to have answered questions 94 percent as effectively as, or better than, those teaching the course. During the testing, the AI tutor was also found to have outperformed ChatGPT 24 percent of the time.

Meanwhile, U-M GPT Toolkit is being used for experiments in clinical settings to optimize patient notes and to translate prescription notes to ensure that patients are offered generic drug options when appropriate.

Advice for Other Organizations

Since the launch of U-M's suite of closed AI tools, more than 49 institutions have gotten in touch with U-M to ask for advice about creating their own. U-M is sharing its experience with those institutions and in the future may even share its IT talent with smaller institutions, such as community colleges, to help them develop their own AI.

In the meantime. U-M is continuing to build and refine its tools. Next year, U-M students may find themselves with a personal AI assistant, for example. The university is also planning to improve the AI so that it can be used with sensitive medical information in the future, allowing the university's academic medical center to benefit more from the tools.

Moving into the Future with AI

In some senses, AI has been around for decades for developers and researchers. But generative AI is a new phase, and public tools like ChatGPT have democratized access to it. Like the web browser, which made the internet available to non-coders, generative AI has opened the possibilities of AI to everyone, no matter their technological skill level. This has come with its share of fears about automation and job loss.

"I empathize with those concerns," said Pendse. "We have to approach AI thoughtfully and carefully. But I firmly believe that the true value of GenAI is in its potential to enhance and augment humanity, not replace it. AI will not take jobs away from you. But people who know how to use AI might." He believes U-M's new tools will give its students not only the ability to move into a world that has been changed by ChatGPT but also the skills to use that technology responsibly and well.


  1. Karthik Duraisamy, Generative Artificial Intelligence Advisory Committee Report, University of Michigan, June 30, 2023. Jump back to footnote 1 in the text.

Case Study Contributors: Ravi Pendse, Vice President for Information Technology and Chief Information Officer, University of Michigan; Robert Jones, Executive Director of Emerging Technology and Support Services for Information and Technology Services (ITS), University of Michigan

This case study is one in a series being produced by EDUCAUSE Communities and Research. A. J. O'Connell is a writer with McGuire Editorial & Consulting specializing in education technology.

© 2024 EDUCAUSE. The content of this work is licensed under a Creative Commons BY-NC-ND 4.0 International License.