Utilizing AI Study Techniques

While finishing my final undergraduate semester as a math major at the University of North Texas, I decided to experiment with artificial intelligence (AI) to enhance my study techniques. This article walks through my journey of what inspired me to start using AI (specifically, large language models (LLMs)), and some of the lessons I learned. I have divided it into four main sections: 1) motivation; 2) custom GPTs vs. individual conversations; 3) prompt engineering; and 4) the AI complexity trade-off.

MOTIVATION

Between school, extracurriculars and applying for jobs, my final semester was the busiest I had experienced up to this point. For Linear Algebra in particular, I was going about my regular study routine: go to class, do homework, review notes before the test—rinse and repeat. Then the first test of the semester came up, and I got a 50.

My initial response to this bad grade was that I hadn’t tried hard enough, that I needed to study more. However, I didn’t have time to study more! I had to come up with a more creative way to learn the material. That’s when I began thinking: perhaps I could use AI to make my study methods more efficient and further engage with the material.

I was hesitant at first, asking questions like “What if AI gives me false outputs?” or “What if I don’t understand the material as well as I should?” But the system I was using in the past semesters wasn’t working, and I needed to try something new. I decided to build interactive study guides using AI. I hoped these guides would act as a private tutor to prepare me for exams.

I saw immediate success on the next test, answering all but one question correctly. I began to build more complicated and detailed models, which had both frustrating and useful results. The main AI tool I used was Gemini, but this article will apply to most AI tools.

CUSTOM GPTs VS. INDIVIDUAL CHATS

When I initially decided to use AI to help me prepare for my second and third tests, I opened Gemini and typed in a prompt to set up how I wanted the interaction to work. I did this in an individual chat—typically the first chat option you see when opening an AI tool, which offers minimal customization. I gathered course materials, which I was allowed to upload to AI, into three sections: the first to prepare for the second test, the second to prepare for the third test, and all three to prepare for the final course evaluation. Each section was also divided into six subsections.

Prompt step 1—telling Gemini what I was giving it:

I am uploading study materials to prepare for an upcoming test tomorrow and want you to prepare an interactive study guide that prepares me to get a perfect score on the test. The study materials contain three sections, and the test will be related to the first section. Each section is also divided into six subsections.


Prompt step 2—giving Gemini some instructions on what to cover:

In each subsection, you will review a portion of the content in the study materials and provide a summary of the content; these reviews should be comprehensive and enable me to have mastery and full knowledge of the material.


Prompt step 3—putting in some behaviors I wanted to see:

Also in each subsection, you will ask me one question at a time until I get 4 questions correct. These questions should encompass all of the material covered in the section. If I answer a question incorrectly, you will provide an explanation of the correct answer and ask me a follow-up question that covers the same material as the original.

This worked very well for the second test and yielded good results for the third test, but I began to notice some of its limitations.

• First, anytime I wanted to do something new, I had to start completely from scratch. When switching from the second to the third test, I had to repeat activities within tests and topics.
• Second, it skipped information and bled over into the sections that I didn’t specify. It started at section 1, as I requested, but went into section 2 towards the end of the review.

For my final exam, I decided to build a custom GPT with the goal of making it easier to expedite the process. With Gemini, this involves two main sections: an instructions section and a knowledge section. The instruction section is where you input a singular prompt into your GPT. The knowledge section is where you upload files. This enables you to create conversations that use your custom GPT as the backdrop for every interaction.

Instruction section prompt:

Attached are study materials that cover the content I need to prepare for my final. These files are divided into three sections, and each section is divided into six subsections. The first section is for test two, the second section is for test three, and the third section encompasses all remaining material needed for the final. I am going to create individual conversations and ask you to review individual subsections. In each conversation, you will only review one subsection and prepare an interactive study guide that helps me understand the material in a way that prepares me for the test. At the start, you will provide an overview of the relevant material you will cover. This material should effectively explain the material as opposed to just stating what we will do. After that, there should be 3 different questions, each covering different aspects of the material to be covered, and these three questions should cover all material that will be represented on the test of the relevant material. If I get a question wrong, you will ask a different question that covers the exact same format. Keep asking questions in each conversation until I get 3 correct. Do not use material covered in the next subsection for the current review. Only cover material in the subsection asked for.


In populating the knowledge section, I, by not having to recreate prompts every time, was able to dive into more specific topics whenever I created conversations. I made five individual conversations, going into different subsections of the material. Instead of creating a complicated prompt each time I wanted to chat, this was already handled by the GPT, and I used simpler prompts like “Review subsection 1.1” or “Review subsection 3.2.”

The second issue, pulling extra information beyond the specified subsection, was still there but was significantly minimized due to the way I used the GPT. Instead of asking for a full review of six subsections, I only asked for one subsection at a time. Using less data in each conversation resulted in less overlap from other portions of the data.

Overall, in my experience, individual conversations are easy to create and are good for quick or one-time tasks. However, when you need to handle a more complicated or repetitive task, that is when I think you should create a custom GPT.

PROMPT ENGINEERING

The first thing I learned from the prompt engineering I did was that I needed to be more specific. In the first prompt I made, I said I wanted to prepare for an exam with an “interactive” study guide. It immediately output the entire review with answers and nothing else.

Using the word “interactive” did not yield good results. Describing what I meant by “interactive” worked better. I clarified that I wanted it to ask me individual questions and get responses before moving on to the next question, which it did successfully.

As another example, I wanted the engine to give me a high-level review of the subject before proceeding to questions. However, it did not explain the concepts and only told me what we would review. I adjusted the prompt to include effective explanations of the material; my changes were applied in the next conversation. I will also note that, by using my custom GPT approach, adjusting the prompt was much simpler than before.

The next step was fixing the behavior of responses. I noticed that when it asked me a question, if I asked for clarity or more information, it would answer my question, but it would also mark the problem as incorrect and then tell me the answer. I changed the prompt so that it wouldn’t count questions as incorrect if I asked a follow-up, but it then counted them as correct and moved on to other parts of the review. I gave the prompt more explicit instructions to allow me to ask clarifying questions without counting my answer as correct or wrong, but to allow space for follow-ups.

Lastly, I gave instructions to ask about a single subject until I reached a certain number of correct answers. However, I found it would move on without the specific number of correct questions. Unfortunately, I was unable to resolve this issue before my test. But I believe it is caused by what AI educator Dan Chuparkoff outlines: “Modern LLMs read and write tokens (common chunks of text), not single characters.” Thus, a word like ‘blueberry’” might be split into pieces like “blue” and “berry.” “It could also be “ ‘blueb’” + “erry” (it’s not always at the syllable break).” As Chuparkoff explains, “To count letters, the model would need to ‘look inside’ each token and do character-level bookkeeping. That’s doable in principle, but the model isn’t explicitly trained to do it, so it’s fragile.” In a similar manner to counting the number of r’s in “blueberry,” the AI seems to be unable to count the number of questions I answer correctly.

Overall, I found prompt engineering to be a critical aspect of building this study guide. When building something this complex, I came across many issues. Adjusting my prompt as I got different outputs helped the model yield more accurate results, enabling me to build an effective study tool.

THE AI COMPLEXITY TRADE-OFF

My prompt engineering led me to an interesting paradox. Whenever I tried to make changes to this model, it didn’t implement the new features the way I wanted and got worse at executing the already existing commands. But the simpler I could make the model, the better it behaved. This is what I call the AI complexity trade-off. If you make your model too simple, it will be hard for it to interpret what you want to accomplish, and it might interpret what you are saying differently from what you intended. On the other hand, if you make your model more complicated, it might become overloaded and fail to effectively accomplish the tasks you want. When prompting, I believe reaching a happy medium is best: specific enough for the AI to understand, but simple enough to execute.

USAGE IN ACTUARIAL EXAMS

For my first three Society of Actuaries (SOA) exams: P (Probability); FM (Financial Mathematics); and SRM (Statistics for Risk Modeling), I used the learning service Coaching Actuaries to assist my studying. Their platform was effective for me and helped me pass these exams on my first try; but I think using AI will also help me expand my learning capabilities.

I plan to use AI to help generate practice questions and exams. In a way that is similar to how coaching companies provide practice problems and practice tests, I will use AI to create questions that reflect what I might encounter in the exams, helping me study more effectively. As previously mentioned, I will handle this process one subunit at a time and combine the results to create practice exams.

In the early exams, it is easy to grade practice exams because all the problems are multiple choice. However, as essay and other question formats appear in later exams, my access to graded practice exams will, I assume, decrease. I plan to complement the learning approach here with a custom GPT to grade my responses to the previously generated practice exams. To ensure accuracy of the AI-generated exams, I will review selected questions and answers and verify my understanding offline by comparing the underlying concepts and solutions to those in SOA sample exams and other resources, rather than relying solely on AI-generated output, thereby enhancing my learning.

I believe it’s important to treat AI-generated questions or feedback as supplemental rather than authoritative, and to verify accuracy against official materials and trusted study resources.

AI tools should be used only for personal study, in compliance with SOA exam policies and applicable copyright restrictions.

IN CLOSING

I believe AI is very good at creating solutions for problems. So good in fact, that it can come up with answers for your hard problems without much critical thinking. I do not recommend using AI as a tool to replace or circumvent the learning process, but rather to enhance it and creatively engage with material that would be otherwise difficult to interact with.

When using AI to assist in the learning process, checking the output is important. It is especially difficult to tell if you are getting a hallucination when you don’t fully understand the material. To ensure you get accurate results, use reliable, authoritative sources—such as official curriculum materials—to check your answers. This will not only serve as a tool to check the AI output but will also be another way for you to learn the material.

Lastly, a simple solution is sometimes all you need to solve a problem. I found that, often, when performing prompt engineering, problems had both simple and complex solutions. Figuring out simple ways to solve problems often yielded better results and created easier-to-understand prompts.

Use of AI tools must comply with all SOA exam policies and integrity requirements and is limited to personal study outside the exam environment. Users should not upload, reproduce, or distribute SOA exam content or other protected materials in ways that violate applicable copyright, licensing terms or SOA policies. The SOA does not endorse or validate AI-generated content, and candidates remain responsible for ensuring the accuracy and appropriate use of any study materials.

Statements of fact and opinions expressed herein are those of the individual authors and are not necessarily those of the Society of Actuaries or the respective authors’ employers.

Copyright © 2026 by the Society of Actuaries, Chicago, Illinois.