Rebecca A. Norton, Ph.D
Massachusetts Maritime Academy
This is the fourth in a series of four blog posts in response to the call: For many reasons, including AI, we had to adapt our teaching strategies and learn as we go this year. So let’s share what we found that was useful, promising, and/or positive in some way! Our audience is NE-COMMIT and anyone else curious about improving teaching mathematics.
The posts are ordered from those describing practices more resistant to AI to those describing practices more embracing of AI.
Here are links to the remaining posts: AI 1 Thinking Video Notes to my future forgetful self by Christine von Renesse, AI 2 Changing My Writing Assignments by Debbie Borkovitz, and AI 3 Teaching Calculus III in the Age of AI by Ileana Vasu.
Using AI to help students prepare for class
In recent years, as you might suspect, students began to use AI to respond to the questions without ever opening their eBook. I gave students’ zeros and started requiring page and paragraph citations in their answers, but I quickly grew weary of “policing” their work. I still liked the idea of students preparing for class, though, and I was curious to know how to use AI to help students learn. So, I designed an assignment that asked students to paste the “Getting Ready for Class” question into an AI tool of their choice, and then answer some questions about the response that AI provided them. I implemented the new assignment via Blackboard in the Spring 2026 semester, and throughout the semester I made tweaks to the questions to try to get students to read and reflect on the information that AI provided them. Students completed the questions on Blackboard, so it is difficult to share the assignments with you, but below is a fairly typical exchange that a student had with AI when completing one of the assignments. The questions are numbered with the students’ response below each question. I kept the formatting the same as how it appeared in the student’s submission.
Directions: The following question comes from a "Getting Ready for Class" (GRC) assignment in your eBook. Read the GRC question and then complete the questions on the form below. NOTE: It is okay to copy and paste the question into an AI platform, but make sure to also include the designated prompt that is given in question 2.
1. Which AI tool(s) are you using?
chaptgpt
2. Prompt your AI tool to give you an answer to the GRC question without a lot of calculus jargon. Copy and paste the response it gives you below (include the entire response):
\(f(x)=ln(g(x)) \)
and
\(f(x)=e^{g(x)} \)
are similar because both use the chain rule. In each function, there is an “inside
function” \(g(x)\), so after differentiating the outside function, you multiply
by the derivative of the inside function, \(g'(x)\).
For \(f(x)=ln(g(x))\)
\(f'(x)=g(x)g'(x)\). For \(f(x)=e^{g(x)}\):
\(f'(x)=e^{g(x)} \cdot g'(x)\) In both cases, the derivative includes \(g′(x)\), which shows
that both functions follow the same chain rule process.
3. After reading the AI tool's response, think of a good follow-up question that would help you clarify your understanding of the topic. Do not use AI's suggested follow-up questions; think of your own follow-up question based on the answer AI provided. The follow-up question must be directly related to AI's previous response. Paste the prompt you give AI here:
how come for the ln problem, you have to divide compared to multiply?
4. Paste the AI tool's response to your follow-up question here:
For logarithms:
\(d/dx[ln(x)]=1/x\)
So when the inside is \(g(x)\), the chain rule gives:
\(\frac{1}{g(x)} \cdot g'(x)\) which is the same as:
\(\frac{g'(x)}{g(x)}\). That is why you divide.
5. What is one thing that you learned in AI's initial and/or follow-up response(s)? Be specific, and use your own words. (Do not copy and paste from the AI tool's response(s).
I learned that the reason you divide is because it's actually 1 over g(x) times the
derivative of g(x) which is the same as dividing.
6. What is one thing that you think AI could have done better to help you understand the topic?
I feel like it could have answered the questions in a shorter, more simple way.
For my last questions when I asked why I divided, there was a long complex
explanation, but when I asked again it gave me it in a simple sentence.
7. OPTIONAL: Did you find any errors in any of the AI tool's responses that you pasted above? If so, explain what the mistake was. If Dr. Norton agrees with you, you'll earn an extra point on the next quiz. If not, type "N/A".
NA
Overall, I think using AI made for a better assignment than I had been giving in
previous semesters. There were still some students who were “gaming” the
assignment and not putting much effort into learning anything by completing it.
There are others who I think found it helpful. As a first adventure into using AI
in my classroom, I think it worked fairly well. Here are a few positives I observed:
A) More students were completing the AI assignments than they had been in my
previous classes where I used the original prep assignments. They seemed more
willing to read the output of an AI model than they were a section of an eBook.
B) AI’s responses do not include body language, tone of voice, or negative words
that may discourage students from asking questions. Plus, it never tires of
questions or runs out of time. In that way, AI can serve as a valuable and patient
tutor or study partner, and students in my class seemed to be comfortable
interacting with AI.
C) My students were exposed to different ways to prompt AI, witnessed the results
that followed based on different types of prompts, and reflected whether AI’s
response was helpful. “Prompt engineering” is something that I have personally
gotten much better at compared to when I first started using AI; I think my students
may have learned a little about that, too.
D) Finally, on a personal level, I learned more about which AI tools my students
were relying on, which was something I was curious about, and I also saw the
differences in how the different tools answered the same questions. ChatGPT
was the most popular model among my students, but there were a variety of
different models that students used. ChatGPT is the most verbose in its responses.
Here are a few negatives:
A) Though AI seems to be fairly reliable in explaining topics related to math, it still
gets things wrong, and sometimes hallucinates. It seems to have difficulty creating
graphics. I saw a lot of confusing and inaccurate graphs/diagrams in students’
submissions. Here’s an example:
B) Very few students took me up on the last question to earn extra credit if they
found something wrong with AI’s response. That worried me because it means
that they were not recognizing when AI was telling them something inappropriate.
C) Certain models, especially ChatGPT, provided “tips” for students about what
they should write for credit. For example, it would say something like, “If you want,
I can give you a super short one-sentence answer that you can write on a test." It
was as if the model had been trained on providing students shortcuts instead of
helping them to learn the concepts.
Despite what my students may or may not have learned by completing these
assignments, I learned a great deal by reading the responses, and I grew much
more comfortable with knowing what to expect from AI. I am debating on whether
I want to use this type of assignment again with my students. On one hand, it
would be easy to do, because all of the assignments are set up in Blackboard
ready to be used again. On the other hand, based on some additional experimentation
I did with AI over the summer, I think I have found a more useful way to use AI to
help students prepare. I write about it in the section below titled “Creating a ‘Study
Buddy’ bot for students to interact with.”
Trying to get AI to evaluate students’ homework assignments
At the start of the Spring 2026 semester, I decided not to use the online homework
system that was associated with the eBook I was using in my Applied Calculus
course. Based on my own observations and discussions with other professors in
my department, I realized that even though students were completing the
assignments, they were not getting the practice they needed. It was too easy to
cut and paste homework problems into AI and get the answer for the problems on
the assignment. I thought that if I asked students to turn in a written assignment on
Blackboard that even if they ended up using AI to complete the problems, they would
at least have to write the work down, and that would be more helpful than just cutting
and pasting answers online.
Fresh from the winter break, and not thinking carefully about my time commitments,
I thought that if I used a rubric to grade the homework assignments, I’d be able to
turn them around quickly enough to provide relevant feedback to my students. As
the semester progressed, however, I got behind in grading them, and found myself
at the bottom of a grading hole. Since grading is tedious, and it is difficult to motivate
myself to do it, I started experimenting with AI to see if it could help.
I began using an AI platform called BoodleBox because my school had bought a
license for the Spring semester for faculty to experiment with. The platform is
designed for educational settings. It claims that it somehow provides access to
different AI agents without allowing those agents to collect information about the
interactions that you have with it. I am skeptical about that part, but I took my
chances and experimented anyway. I wanted to see if I could teach it how to apply
a rubric to my students’ submissions and provide an appropriate grade.
I provided the AI agent with my rubric and an answer key for the relevant problems,
and without any other input, it did a decent grading job. It applied the rubric a little
differently than I did, though, so I created a flowchart to guide the AI agent to apply
the rubric in a way that it would arrive at the same score as me. That took some
effort. I tweaked the flowchart by doing several grading runs where I compared
the scores I assigned using the rubric versus the scores that AI assigned. I also
added in some steps based on the feedback that the AI agent gave me so it would provide me with more reliable results. Here is a link to the flowchart I ended up with.
In general, I was happy with the way AI implemented the flowchart; I was definitely pleased with how quickly it could apply it and format the output for me. It turned out, however, that the whole process was not as efficient as I had hoped. AI often had difficulty with students’ handwriting, or finding all of the problems within a student’s submission, or being unable to read the problems if the student submitted a photograph of poor quality. In those cases, it would grade the assignments based on what it could read, and then flag it for me to review. When I looked at those cases, it rarely provided the appropriate grade. In its defense, it gave me some advice to help improve its ability to score the submissions, but its main suggestion was that I ask students to format their homework in a specific way. Since the students had already turned in the assignments, I was not able to implement the advice it gave. Additionally, based on specifications with the BoodleBox platform, I had to upload the students’ work in batches and that really slowed the whole process down.
In the end, I don’t think I saved any time in grading my assignments for the Spring 2026 semester. In fact, I may have spent more time trying to tweak things with my flowchart than I would have if I had just graded the assignments. It wouldn’t have been as fun, though! Plus, I’ll be able to use what I designed in the future. Now that I know what the model needs in order to process the students’ submissions accurately, I can require that students complete their work in the required format. I will also consider using a homework folder on BoodleBox, which is another suggestion the platform gave to me, because then it could check to see if a student’s submission was acceptable and require the student to resubmit their assignment if the model was unable to read it. Plus, having the student upload it directly to BoodleBox would keep me from having to transfer the files between Blackboard and BoodleBox, which was the biggest time consumer in getting AI to grade the assignments.
Overall, I learned a lot about how to interact with an AI agent in a productive way. It felt a bit like having a teaching assistant. As I noted in the previous section, AI never gets tired and it is not possible to hurt its feelings. So it is perfectly “happy” to make adjustments until you are satisfied with the outcome. It is important to be explicit in your instructions, though, because it will assume things that may or may not align with your own philosophies. In coming semesters, I plan to implement the lessons I learned during the Spring 2026 semester and continue to experiment with asking AI to apply rubrics to help grade student assignments.
Creating a “Study Buddy” bot for students to interact with
At the last minute, I was asked to co-teach a summer course called Introduction to Python for Business Data Analysis. I did not have much time to prepare. It was not a course I had ever taught before, and it was not a course that my school had offered before, so I didn’t have anyone else’s materials to borrow from. Plus, it had been a while since I’d really done any programming myself. So, I knew I was going to have to leave myself lots of planning and thinking time to make sure I was prepared to teach the classes, especially since each summer class was the equivalent of a week during a regular semester.
I wanted to make assignments that provided students with some feedback, but that I would not have to spend much time looking at or commenting on myself. I had learned my lesson trying to collect and look over my students’ homework in my Applied Calculus class during the Spring 2026 semester, and I did not want to dig myself into any more grading holes! So, I tried to design a bot within BoodleBox that specialized in the material that we were focusing on. I named it “Python Study Buddy”. You may have to sign up for a BoodleBox account, but if you do, you can try it out here:
https://box.boodle.ai/a/@PythonStudyBuddy
Last I checked, BoodleBox was giving two-month free trials. If you do call up the bot, the first thing it will ask you is what chapter you want to work on. Tell it a chapter number between 7-14; those are the chapters in Part II of a book by Lee Vaughn titled Python Tools for Scientists. Students were required to buy a pdf version of the book.
In between class meetings, students were assigned to interact with the bot. At the end of the session, students would ask the bot to print out a transcript, and then upload the transcript to Blackboard. I then reviewed the transcript and assigned a grade based on a rubric. The rubric included four categories: Topic, Productive Interaction, Demonstrated Learning, and Time-on-Task. As an aside, I had AI create the initial draft of the rubric based on a description I gave it.
I considered this use of AI to be a success. The bot acted as an assistant that condensed the information that I needed to give students the proper grade and allowed me to apply my rubric quickly. Additionally, at the end of the course, several of the students mentioned how helpful they had found Study Buddy and that they used it as their main study tool because it helped them decipher the information within each chapter in a conversational way. I will definitely continue to experiment with similar tools in my upcoming classes this fall.
