Department Name Change
UWM's Center for Excellence in Teaching & Learning (CETL) has changed its name to the Center for Advancing Student Learning (CASL), learn more about it in our official statement. Alongside the name change, our website's url has changed to uwm.edu/advancing-learning/. Please update any bookmarks or links to point to our new url!

Custom AI Agents for Learning and Scholarship 

In November’s Active Teaching Lab, Stuart Moulthrop shared his experience using a custom GPT “agent” to support reading-based discussion and scholarly experimentation in a graduate-level theory course. What began for Stuart as a practical test – “What could AI do with a semester’s worth of dense theoretical readings?” – quickly evolved into a broader reflection on what it means to be a scholar, teacher, and creator alongside AI. His experiment showed that building a course-specific chatbot is surprisingly feasible and opens genuinely interesting pedagogical possibilities, from supporting discussion to modeling scholarly dialogue.

At the same time, the session underscored a crucial constraint: student attitudes toward AI ultimately determine what uses of AI are pedagogically appropriate. Stuart’s students raised thoughtful objections to the use of AI, even while engaging critically with it. The Lab discussion highlighted that while custom AI agents can support learning, their effectiveness depends on careful framing, transparency, and voluntary use that respects student concerns.

To continue exploring this month’s topic, start by watching the November 5th Active Teaching Lab recording below. Then, take a look at the key takeaways from our discussion and a few classroom experiments you can adapt right away. 

Lab Takeaways 

  1. Building a course-specific chatbot is surprisingly feasible 
    • Creating a custom AI agent for your course requires far less technical friction than is generally expected. Stuart Moulthrop built SemBot in about a week, and it successfully processed more than a thousand pages of course readings – including scanned PDFs – to answer questions and support weekly discussions. These results were possible only because Stuart iteratively tested, refined, and expanded the agent based on how it behaved with real course materials.
    • AI’s usefulness depends on thoughtful instructional design. AI can supplement your teaching, but it cannot replace an instructor’s judgment, presence, or engagement, especially in classrooms where student trust and comfort with AI vary widely.
    • See more on creating your own AI Agent below!
  2. Don’t assume students want AI in the classroom 
    • Stuart’s graduate students were broadly skeptical of AI for substantial ethical and political reasons. His students raised thoughtful concerns about: training data taken without consent, the hidden labor behind AI systems, AI’s high environmental costs, and the concentration of power in a few tech companies. Instead of shutting down the conversation, these concerns fueled strong, critical dialogue regarding AI use.
  3. Students’ attitudes toward AI strongly shape what’s pedagogically appropriate in your classroom 
    • Because of this antipathy toward AI, Stuart found that what “good teaching with AI” looks like is very much determined by student buy-in. He therefore kept AI use optional, and spent time in class examining and contextualizing its use in literature and education more broadly.
    • Ultimately, even students who refuse to use AI can still learn from it and engage with its implications for their field. For Stuart, he leveraged AI to help students see how humans write, imitate, and construct meaning.

Experiments Worth Trying 

Build your own AI agent

The SemBot instructions and information included below can be used to help you build your own AI agent. Additional step-by-step instructions for building an AI agent are available via this link. 

SemBot Resources

  • SemBot b05 – Stuart Mouthrop’s AI Agent
  • SemBot’s training data is available here – Stuart added precise dates and detailed instructions only after extensive trial-and-error, highlighting how crucial iterative testing is for accurate AI responses.
  • SemBot’s Mood Rules V5.2 – SemBot Mood Rules outline thirty distinct conversational “Moods,” each associated with a unique set of stylistic and structural constraints for generating responses. These rules dictate the bot’s persona and tone, ranging from “Scholarly” to “Salty” to “Dada,” and often include mandatory elements like specific greetings, closing phrases, or grammatical peculiarities.
  • Course Syllabus – Stuart reformatted this syllabus to improve SemBot’s ability to access it for answers