The College is now offering students, staff and faculty access to premium models of five generative artificial intelligence (AI) companies, as of early August.

Bowdoin’s Office of Information Technology (IT) has made the models available through LibreChat, an open-source application that provides an interface for accessing multiple AI models on one website.

Last year, the College used a similar AI interface called Amplify, though it was only available to faculty who requested it for their classes. Senior Vice President and Chief Information Officer Michael Cato—who spoke about LibreChat at last week’s faculty meeting—said IT decided to switch to LibreChat because the interface has more users, more developers and is evolving faster than Amplify.

The 15 models included in Bowdoin’s LibreChat platform are products of the companies OpenAI, Anthropic, Meta, DeepSeek and Amazon Web Services.​​ The College is paying those companies on a per-use basis, according to Cato.

“It’s not a predictable bill,” Cato said. “It’s completely based on how much use there is. We’ve got safeguards in place…, so it’s not gonna miraculously switch from $200 a month to $10,000.”

Cato noted that one of the College’s main reasons for buying the models is to ensure all Bowdoin students have equal access to the same technology.

“This is part of the institutional commitment to access,” Cato said. “When it comes to generative AI, that was one of the questions IT brought to the [Committee On Teaching and Classroom Practices (CoTCP)] last year. If we didn’t do anything, then the students who could afford it were likely already buying their own licenses. And the students who couldn’t might be using free versions.”

IT and the AI subcommittee of the CoTCP—which is responsible for considering how the College should approach the use of AI in classrooms—were also concerned by the fact that most free AI models do not protect users’ data privacy. One of the most important advantages of the premium models Bowdoin is paying for, according to Cato, is that all of the information its users share with the models stays private and is not used to further train any of the models.

“If you’re using one of these accounts for free, then the terms of use mean that they can train the models off of not only what they already know, but what you are asking,” Cato said. “So when you pay, part of what you are paying for is that they don’t train the models off of what you are asking. And your data is held separate from the models themselves.”

Last year, the AI subcommittee of the CoTCP had several discussions about the ethical implications of using AI, according to Professor of Physics Dale Syphers, the chair of the CoTCP and the subcommittee.

“[Members of the subcommittee] were worried about the energy and water usage of AI,” Syphers said. “They were worried about the ethics of how the [AI model] training took place … and the way in which it was offloaded into third-world countries … that did not necessarily use their labor with appropriate compensation. There was the ethics of how AI used information and got information, so, for instance, the fact that they read in a tremendous amount of material and books and journals … and are making money off that but did not compensate anybody.”

To gain access to Bowdoin’s LibreChat website, students must complete a digital AI ethics course that, through articles and YouTube videos, addresses many of the topics the AI subcommittee discussed last year. The course, which was created mainly by librarians Beth Hoppe and Juli Haugen, has eight modules, entitled “Environmental Impact,” “Human/Labor Impact—Who Is Training AI for Accuracy?”, “Bias—The Value of Diversity and Inclusion,” “Data Privacy and Security,” “Copyright and Intellectual Property,” “Regulating Artificial Intelligence—Who Is Responsible?”, “Information, Misinformation, and Verifying Sources,” and “Academic Integrity—Ethical Use of AI in Education.”

After completing the modules, students must write down three “I believe” statements about AI before gaining access to LibreChat. For staff and faculty, the course is optional.

Cato explained that most AI companies do not share information about the environmental impact of their service, including those in Bowdoin’s LibreChat platform, but he said IT is calling on the companies to release that information and hopes the College eventually gains access to it.

“They know how much computation our queries use because each month we get a different bill based on how computationally intensive our requests have been of them,” Cato said. “But what they don’t provide right now is a version of that that says: ‘Here’s the environmental impact of that computational use.’… So we’re pushing on them to get that.”

The College is waiting to officially announce to students that LibreChat is available until later in the semester, as Cato and his colleagues want to give faculty time to learn and think about the service first.

“We wanted to intentionally follow the faculty’s lead, so the faculty get to talk to their students about expectations for the semester, any conversations they want to have about generative AI—we wanted those to happen before anything came from IT about LibreChat directly to students,” Cato said.