Is Your University Building a Custom AI Platform?
In June 2023, the University of Michigan’s Generative AI Committee released a report with a list of recommendations for how to adapt to increased use of generative AI tools like ChatGPT. One of the committee’s key recommendations was that the university provide “secure and equitable access to GenAI platforms and tools for the entire U-M community.” To meet this need, the University of Michigan launched a custom AI service platform in time for fall semester. In September, two other universities–Harvard and the University of Tennessee Knoxville–announced the release of customized platforms with AI tool integration for their respective communities.
Whether and how a university can provide its communities with generative AI tools has the potential to become the next technological arms race in higher education, and there are many strategic issues for universities to navigate as they enter the fray. Today we detail the approaches the earliest entrants in the race have taken, and what that could portend as the race widens over the next year and beyond.
The Current Entrants
The University of Michigan
On August 21, the University of Michigan became the first major US university to offer a custom AI platform for its faculty, staff, and students. These tools were trained by the service providers and are hosted by the university or controlled in a university cloud environment. The platform features three tools:
- U-M GPT, which allows users to access GPT 3.5, GPT 4, and Llama 2, with a 25 prompt per hour limit.
- U-M Maizey, which allows users to query their own datasets.
- U-M GPT Toolkit, which allows users to build, train, and host their own models. It is only available upon request.
Michigan’s ITS website specifies that U-M GPT is initially free and that the latter two tools will have usage-based fees starting in November 2023. They highlight that their tools were tested for accessibility and have an advantage over ChatGPT in their capacity to work with screen readers. Users can enter “moderately sensitive” data, and their data will not be used to train commercial generative AI models. With three tools already available to the entire community, Michigan’s platform is the most expansive and ambitious so far. Their fast rollout may have the most immediate impact, but also comes with the highest degree of risk given the complexity of rolling out services university-wide. For example, University Michigan timed its campus-wide roll out of the new technology with the start of the fall semester. This widely touted roll out was ultimately diminished by an extreme IT service disruption that left the university without access to email, on-campus wifi, and zoom, among other core offerings, during the first week of classes.
Harvard University’s office of Information Technology (IT) announced its AI sandbox pilot on September 4. This limited release offers access to GPT-3.5 and GPT 4 through Microsoft’s Azure OpenAI, as well as to Anthropic’s Claude 2 and Google’s PaLM 2 Bison. Like Michigan’s platform, Harvard’s sandbox is approved for “medium risk confidential data” and will not be used to train public models. Detailed information on the platform–such as the anticipated pricing model for the future–are not currently available, but Harvard claims it plans to “expand the pilot later this fall.”
The University of Tennessee Knoxville
On September 14, UT Knoxville announced internal AI resources now available for its community. The first, U-T Verse, relies on GPT-3.5 through Microsoft’s Azure OpenAI. It cannot yet query user datasets, but their Office of Innovative Technologies notes that this feature will be available soon. Microsoft’s Bing Chat Enterprise is the second available tool. Unlike the other two universities, UT Knoxville distinguishes their two tools from each other based on who can use them and for what purposes. U-T Verse is available to faculty, staff, and students for “UT-specific AI conversations, including conversations about your research and proprietary university data.” Bing Chat Enterprise, on the other hand, is available only to faculty and staff with a Microsoft A5 license, and can be used for conversations that do not discuss their “research or proprietary university data.” These services are available at no additional cost.
Comparing Services: Data Security, Access, and Pricing
The three platforms described here share several key features designed to address core challenges with adopting generative AI into higher education. One of these relates to universities’ need for generative AI environments that provide data security necessary to work with confidential and sensitive research data, protect university IP, and shelter students and other community members from having their queries incorporated into commercial platforms. Providing access to powerful tools like GPT-4 in a sandbox environment is an important step towards meeting those needs.
There is also the challenge of ensuring that the platforms are optimized to support research workflows. Maximizing the value of generative AI for researchers will involve allowing users to query specific datasets or easily build custom AI models. The inclusion of “Maizey” and the “GPT Toolkit” are examples of how universities can respond to this need when building out their offerings.
Another challenge that these platforms address relates to scope of access, given that generative AI technology has potential uses for so many of the universities’ stakeholders concurrently. Harvard, Michigan, and UT Knoxville’s platforms are noteworthy because they are attempting to make at least some of their offerings available for the broadest possible use, which is similar to how universities approach access to mainstream enterprise software such as the Windows Office Suite. For the same reason, it is notable that in all three cases these new platforms are being developed under the auspices of enterprise IT, as opposed to research computing, educational technology, or the library—although it is also worth recognizing in the case of Harvard that the initiative is “in collaboration with” the vice provost for advances in learning and Faculty of Arts and Sciences,Division of Science.
With Microsoft organically positioned to place its offerings at the forefront through its relationship to the university’s enterprise IT unit, there is a question of how more specialized information and educational technology vendors will be able to position their own interests and offerings. There is also a question of how universities will generally fare financially if they allow their interest in generative AI to further solidify their reliance on big tech.
Whether institutions will be able to—or even want to—offer free access to these tools remains to be seen. What we know so far about Michigan’s proposed fee-based model, including for students, suggests that schools may take some very different strategic approaches to offering these tools. There is a scenario where some schools who can afford to will approach providing free access to these kinds of tools a perk of affiliation for its communities to provide a competitive advantage. Other institutions may charge for the use of these resources, either to cover the costs or generate new sources of income. One model some universities may look to is the cost recovery model of research cores, where “charge backs” to grants help to cover the costs of maintaining the university’s underlying research infrastructure. In the context of cost recovery for student or teaching-geared services or resources, universities may look at their formula for how funds are drawn from the student technology fee. In all of these scenarios is the reality that it will be difficult for universities to prioritize access for its entire community given the considerable costs
We anticipate that other research universities will create similar platforms in the coming months to secure AI access for their communities. When doing so, these universities will be making decisions about generative AI services in a volatile environment in which available technologies are quickly evolving. The ethical dilemmas around generative AI are also unlikely to disappear anytime soon. The choices institutions make with regard to their custom AI service platforms is a key facet of their response to this complex environment that generative AI has engendered.
At Ithaka S+R we are continuing to monitor this space closely as part of our multi-year research project in collaboration with 19 schools on how to make AI generative for higher education. If your institution is planning to follow in the footsteps of Michigan, Harvard, and UT Knoxville in offering custom AI platforms, we’d love to hear more about what you are working on. If you’d like to be in touch, please send an email to firstname.lastname@example.org.