I’ve been thinking more deeply about how AI can help improve student learning. This is not a new subject for me. My previous blog, “Technology’s Misunderstood Potential,” offered a glimpse of the field as it was 30 years ago when, pursuant to Ernie Boyer’s Scholarship Reconsidered: Priorities for the Professorship (1991), I began studying technology in teaching. The blog described the then-controversial idea that so-called “course objects” could leverage teaching efforts and improve learning. These software tools stored key course information in the form of text, formulas, data, etc.; provided interfaces so students could access the information conveniently; and presented problems, practice quizzes, etc. for creating student engagement. The algorithms were necessarily simple, but the advent of AI has morphed them into sophisticated generative tools such as Gray’s CoCo Course Companion.
Student engagement is a key element in student success. Learning is not simply a matter of listening to lectures, memorizing information, and completing assignments. It is an active process in which students think critically, ask questions, collaborate with others, apply ideas, and reflect on their own progress. When students are truly engaged, they invest time, effort, curiosity, and personal meaning in their education. When engagement is superficial, learning is superficial or mostly absent. Unfortunately, students differ in “what turns them on”: what triggers their investment and facilitates their learning.
Achieving Student Engagement
Professors have known how to engage students with course material for many years. The method is simple: talk with them, stimulate their questions, and constructively challenge them to find answers. U. S. President James A. Garfield once famously stated that Mark Hopkins on one end of a log bench and himself on the other would be all the higher education he needed. (Mark Hopkins, long-time president of Williams College, was renowned for his ability to engage students in dynamic, individual debate rather than lecturing to them.) Institutions try to approximate this by using Socratic dialogue with small groups of students wherever possible. Unfortunately, it’s difficult or impossible to achieve student engagement at scale using this method.
Economics presents the first scaling problem. Small-group teaching predominates only at well-endowed liberal arts colleges, and many large classes can be found even there. Most students at other institutions experience mainly larger courses. Socratic dialogue is difficult but not impossible in large courses. I’ve used it myself. But that leads to the second scaling problem: one for which (pre-AI) I never found a truly satisfactory solution.
Achieving engagement for all students is the second scaling problem. It involves two sub-problems: what I call the responsiveness problem and the customization problem. Responsiveness means “being there at teachable moments” in the student’s educational journey: e.g., to answer questions and clear up key points and to open new areas of inquiry and understanding. Customization means tailoring interventions to students’ particular preparedness and learning styles: to sense why students get stuck and respond accordingly.
Responsiveness and customization can be viewed as the quantity and quality dimensions of faculty-student engagement. Both place heavy demands on faculty time, attention, and cognition.
Achieving faculty-student engagement in a small seminar usually is not a problem. However, it becomes exponentially more difficult as class size grows. Take Socratic dialogue, for example. The better and more prepared students wave their hands and clamor for attention, while the others are slower if they seek to enter the competition at all. Skilled teachers will recognize this problem and draw out the less active participants. But that is easier said than done, particularly as the class becomes larger. The best students usually get the most attention, and thus become the most engaged. The average and less-than-average students are more like spectators, perhaps with a twinge of angst that may inhibit learning.
The result is that student performance gradients get amplified, which leaves a great deal of untapped learning potential. This should not be viewed as professorial failure: the problem of engaging with most or all students is inherent in the actual classroom situation.
This problem in not unique to Socratic dialogue. Students may struggle in any kind of learning situation. Instructors should identify these problems and seek to intervene. But how? Socratic questions targeted to difficulties implied by the students’ in-class performance represent one approach, but of course there are many others. Finding the right interventions in particular cases requires the professor to diagnose and then prescribe a remedy for the student’s particular difficulty. This is a heavy cognitive load for faculty, especially if it must be “on the fly” in a dynamic situation. As with basic responsiveness, the customization load increases dramatically with student numbers.
Course Companions Leverage Faculty Effort
The algorithmic “course objects” I mentioned above were developed to address the responsiveness and customization problems as they were understood 30 years ago. So were widely used teaching tactics like “think-pair-share” (aka, “turn and talk”), where students are presented a problem in class and turn to their neighbor for discussion. Unfortunately, neither algorithmic models nor student peers have the knowledge base and reasoning power for truly responsive customization.
But today’s AI-based course companions have the knowledge base and cognitive capacity for such customization: for all students, all the time, at scale, with low cost.
One needs to experience a course companion tool in order to appreciate it, but here are a few facts in support of my statement. Remember, in all cases, that the professor of record remains in charge of the course. The AI’s job is to amplify faculty time and cognitive capacity, to mitigate the responsiveness and customization problems for all students.
Course Companion Features and Functionality
A course companion applies the logical and conversational power of AI in the service of tutoring students. Like all AI, the companion is:
- Always available. A companion addresses the responsiveness problem by allowing students to receive its full suite of services 24/7—thus catering for exam preparation, writing deadlines, and “teachable moments” generally, without always requiring human tutoring. Such access is invaluable for all learners, but especially for non-traditional students, commuters, people with family responsibilities, or late-night habits.
- Capable of retrieving and summarizing information as needed by students, in support of content provided by instructors and textbooks. Communication may be written or oral, through images and equations as well as text, and foreign-language translation is available. Students customize the queries to suit their particular learning problems. Having answers available 24/7 is hugely valuable when students are grappling with new material.
These two functionalities are available on standard AI platforms. What is it, then, that differentiates a course companion? Here are the most important examples. Course companions:
- Harness the AI to a grounded knowledge base: The AI is strictly bound to the instructor’s resources (syllabi, slides, lecture transcripts, readings, and assignment rubrics). The AI does not search the open Internet without instructor permission. This prevents referencing inaccurate or otherwise inappropriate material and reduces the possibility of hallucination.
- Provide Verifiable Source Citations: When answering student queries, the companion provides page-, slide-, or document-level citations referencing the professor’s uploaded materials. The student is encouraged to dig more deeply into the syllabus material.
- Offer Configurable Teaching Modes: Instructors can set the system to operate in:
- Socratic Mode (asking guiding questions, tailored to individual students);
- Full answer mode (conversation);
- Provision of hints and step-by-step verification of student work (among other things, prevents usage as a shortcut generator).
One can expect more options to be added with more usage and instructor feedback. These are important examples of learning customization.
- Offer practice quizzes, case studies, flashcards, and other study tools on demand to help students with their learning efforts. They are based on the syllabus and prior student interaction with it.
Course companions are faculty- as well as student- facing. They actively leverage the professor’s ability to identify the need for personal intervention and provide clues about the kind of intervention needed. Course Companions:
- Provide professors with information about student usage, and whether their interactions with it suggest they are stuck on some issue or concept. Student breakout experiences, which indicate insight creation and readiness to move ahead, may also be identified.
Professors have more visibility into their students’ learning processes than is usually available from conventional teaching methods. The AI’s ability to identify struggling students will improve over time. The same is true for instructors’ ability to use the companion to improve student learning without increasing their own workloads.
Finally, course companions include special guardrails to protect students and keep the learning process on track. They will:
- Intercept student efforts to shortcut the learning process (i.e., to get the answers directly from the AI) by redirecting them into Socratic or other appropriate questioning. Queries that are too broad or outside the syllabus will also be redirected.
- Report indicators of student distress to the instructor. Also report toxic behavior.
- Ensure that course materials remain private and are not used to train public AI models. This is very important for some institutions and faculty.
Thirty years ago, I wouldn’t have dreamed that a software package ever could deliver these benefits. But AI does that now, and its performance will only get better.
My next blog will discuss the probable adoption trajectory for course companions. The adoption of major innovations is never seamless, but I began studying them in my PhD thesis (“Innovation and Market Penetration,” 1960) and am confident that, in this case, the benefits will overcome the barriers in short order.
