AI Teaching Assistants Provide Extra Support for Faculty and Students

“AI TAs can be extremely helpful when they’re providing feedback on a student’s work before it is submitted, enabling the student to engage actively and more deeply on a topic,” says Steven Butschi, a director with Google for Education. He has seen examples of students uploading a position paper into AI “and then debating with it to find weaknesses in their arguments. The goal of these tools is to help students learn better by enabling them to better understand the material for themselves.”

AI Teaching Assistant Models Are Trained on Syllabi

Before AI TAs can help students learn, the models need to be trained on course-specific materials.

Using a commercial generative AI platform, Smith uploaded her course syllabus in a Microsoft Word document, along with course assignments. “If a student went in there and used a prompt that says, ‘When are the article reviews due?’ it can read that information that I have fed the AI and provide that to the students,” she says.

To train the model, “I told it, ‘This is your role as a course-specific teaching assistant. You coach students to learn through hands-on experience. You never write graded work or reveal any test answers,’” she says. “The more specific you are, the more likely the TA will provide the best feedback to your students.”

The University of Michigan’s Stephen M. Ross School of Business, meanwhile, has launched a virtual TA pilot program powered by Google Gemini. To train the model, it first needed ingest the textbook.

“The technical term is retrieval augmented generation,” says Andrew Wu, associate professor of technology and operations. “That converts text language into math, almost like a coordinate. This is a custom pipeline built on Google’s Vertex AI platform.” From there, some fine-tuning is needed.

“I have been teaching a course for 12 or 13 years, and over that time, I have accumulated so much material — reading materials, practice problems, slides,” says Jun Li, professor of technology and operations. All of that goes into the model too.

From there, Li must classify information based on how it’s used. Students are assigned both practice problems and business cases, “so I have created different agents: one for the case study discussion and the other for the practice problems,” she says. “The agents get a different knowledge base; I curate it, knowing the learning modes of my students.”

READ MORE: Modern classrooms are driven by modern technology.

Institutions Ensure Data Security When Using Chatbots

To keep all that data safe and ensure student privacy, schools are relying on vendors as well as on their own safeguards.

At the University of Pennsylvania’s Graduate School of Education, for example, an AI TA answers student questions and gives formative feedback on assignments. The project uses secure connections and vendors whose terms of service “have guarantees of data privacy and how data is used — and services with reasonably good track records of data security,” says Ryan Baker, director of the Penn Center for Learning Analytics.

Project leaders take care to secure data internally, “and make sure that any data that is transmitted to an external AI vendor is either carefully deidentified or that student permission has been properly obtained,” he says.

At Fort Hays, Smith focuses on training both the model and users to ensure security. “We do not provide any student identifiable information to the TA, only class content. I don’t upload my roster, and there is no individual identifiable information ever submitted,” she says.

On the user end, she offered students some basic instruction at the start.

AI Teaching Assistants Improve Student Performance

Those running AI TA projects say they are freeing up teacher time while driving improved learning outcomes.

DISCOVER: Managed services can streamline AI innovation.

Smith surveyed students and found they appreciated the convenience of a 24/7 TA. “If a student wants to email an instructor at midnight, chances are, they’re not going to get a reply back until maybe the next day,” she says. “They loved that they could use it at any time.”

This, in turn, eased her workload. “As a faculty member, it helped answer a lot of questions that would normally have come my way,” she says.

In Michigan, Li says the AI assistant allows faculty to be more creative. “On the teaching side, you have seen a lot of innovation, different ways of using AI,” she says. “Some professors have multiple AI agents that students can talk to, with different characters. We also have faculty members who let the students teach the AI — switching the roles. And we have faculty members coding the AI to deliberately make mistakes and then asking students to identify them.”

Baker sees a direct impact on learning. “We have seen better student performance on future assignments. Our AI TA gives feedback to students on assignment No. 1, and they don’t make the same mistake on assignment No. 2,” he says.

Improvements include “higher student grades, faster responses to student questions and less time by instructors answering student questions,” he says. Overall, “lots of benefits.”

“I went through examples of some prompts they could use, how to use it and so forth,” she says. “I also informed them that there’s no student data in here, and that they should not put any personal information in there.”

Wu relies heavily on the vendor’s capabilities. “The reason we work with Google is that Google Cloud as a technology provider is FERPA-compliant. Student data is housed in a safe fashion; it doesn’t leak out,” he says.

“Everything is deployed in a standard Google Cloud environment. Everything is stand-alone. And the model itself that interacts with students is frozen. It cannot actually access any data to be used as training data,” he says. “Everything is very isolated.”

UP NEXT: Four AI trends to watch in 2026.