By Melanie Cooper
The use of generative AI (genAI) systems in education is inevitable. Sure, there are drawbacks and caveats (some AI systems hallucinate, student data may be used to train the LLM model, there are fears of students cheating), but our students (or at least those “in the know”) are already using genAI tools to help study, prepare for tests, and complete course work. All of which means we need to think carefully not only about how to provide equitable access for all students, but also about just what we want teaching and learning to look like as we move forward into this new era. The central theme of this short piece is that we should not squander this opportunity by merely automating business as usual; the affordances of genAI can be used support students and instructors as they engage in deeper more meaningful learning activities.
Indeed, commercial publishers are already releasing courseware that integrates AI for both instructors and students. AI can reliably summarize lectures and produce multiple choice questions from PowerPoint decks (probably also supplied by the publisher). We also see some examples of AI systems as tutors, typically helping students answer multiple choice questions, or directly providing answers to queries. The challenges with these use cases are that, if we are not careful, the intellectual struggles required for deep learning and understanding of principles will be short-circuited. Though genAI systems can certainly be used to develop questions and analyze (and grade) student answers, no-one will be surprised if students use such genAI systems to answer them. Misapplied or ignored, genAI could cut out the tiresome work of learning
I strongly urge that we seize the opportunity offered by genAI to rethink what we want students to know and do, and how we can support them in these endeavors through the judicious use of genAI. One potential approach uses so-called Retrieval Augmented Generative AI (RAG AI) systems. These systems do rely on the large LLM models (current examples GPT 4.o or Claude 3.5) but rather than using content from the whole internet to produce responses, A RAG chatbot formulates responses to queries from user supplied materials (such as texts, course materials, research papers, course web resources, etc.), which substantially lowers (and may eliminate) hallucination. Additionally, student responses are firewalled from the model itself, so are secure and confidential. RAG-based chatbots can be engineered to take on different personas. For example, a Socratic Tutor can provide supportive feedback, while guiding students in constructing meaning for themselves rather than simply providing answers. This is especially important if we want students to learn to connect and use ideas, rather than simply memorize and regurgitate. A supportive coach can make all the difference. An analyst bot can score and provide feedback to complex written responses from many students (e.g. in large lecture classes) thus negating the use of trivial multiple-choice questions, which is often the current necessity given the resources and time available. A different type of bot can rapidly provide an instructor with a summary of a whole class of student responses to open-ended prompts, giving a much richer indication of how students are responding to instruction. In these ways genAI can support students and instructors as they grapple with teaching and learning more complex material. We have made examples of all these bots and will be testing them with students this fall.
To reiterate, genAI systems have affordances that enable us to move beyond the traditional “plug and chug” formats of instruction and assessment, and give us the opportunity to critically analyze what we are teaching while enriching and deepen students’ learning experience, but only if we are careful and deliberate in how we use these new technologies.
Melanie Cooper is the Lappan-Phillips Professor in the Department of Chemistry at Michigan State University.