Skip to content
Griffon Prep202-643-4743
  1. We use AI in our student-facing drills only when its presence creates a valuable tool that we couldn’t reliably build with conventional programming.

    We pledge never to use AI for trivial tasks that a single well-built computer program could accomplish.

    At Griffon Prep, we use AI in student-facing applications when a drill requires immediate feedback that recognizes natural language and nuance. Some skills are so valuable that it’s important for a student to get instant, targeted feedback so that they develop in the right direction, and those drills are where we use AI. For drills where a video or print explanation can allow a student to easily course-correct, we keep out the AI.

  2. We keep an eye on our AI use.

    Because we’re a small company on a tight budget, we keep a tight rein on how much we use AI. The average AI call for one of our drills costs about two cents. Although that might seem insignificant to a huge company, we watch those costs closely and make sure each use offers a substantial boost to your learning.

  3. We never assume that AI can just “do the job.”

    For our few drills that run AI, we spend about four hours developing and testing the prompt for each individual question to make the feedback and guidance as accurate and helpful as we can. We check in on the AI feedback several times a week and refine the prompts when we find something that needs work.

    AI can still make mistakes. For all of our drills that include AI, we offer a place to flag responses and ask questions of your instructor. That way, you never have to worry about being left in the hands of an uncaring AI: there is always a human somewhere behind the scenes.

  4. We use our own proprietary explanations to craft our AI prompts.

    Our AI prompts are engineered by us, and the explanations we use to craft them are our own, not pulled from other companies or sources on the internet. That is part of using AI in an intellectually honest way: we do not copy other people’s copyrighted explanations into our prompts.

    Even more importantly, it gives us control over the quality of the teaching guidance we provide. We use our own high-quality explanations, crafted by human instructors who understand the myriad ways students can approach an LSAT question.

Bernie’s note on AI

Eight years ago, I tried to create a program where students could type in a response to an argument and the computer could use language matching to tell whether the student’s answer was correct.

It was a failure.

After creating about 30 different logic branches for possible responses, I realized there were over 80 more to go. Moreover, there was just no way to get the program to match the language in the response in a way that could reliably understand when students misspelled words or used synonyms. After about two days of work, I abandoned the project.

When ChatGPT came on the scene in 2022, my hopes were high. But the early ChatGPT models were bizarrely stupid. They could understand language, but couldn’t reliably follow the reasoning. And they were so stubborn—I spent an hour arguing with an early model that 2 cubed plus 3 cubed didn’t equal 5 cubed. They weren’t directly useful.

It wasn’t until early 2026 that I found models that could recognize reasoning and language well enough for me to resurrect my dream project of creating a program that could evaluate a student’s response to an argument prompt.

And that released a flood of creative energy. I’ve been non-stop busy ever since.

The trait that makes me an exceptional LSAT tutor is also the trait that has allowed me to use AI in a unique way. I listen carefully to what students are thinking, I evaluate where their thinking has gone wrong, and then I give that feedback so they can correct course. Where other tutors explain questions, I coach students through questions.

“Wow,” said a woman during one of our first hours of tutoring. “You are really in my head. That is exactly what I’m thinking.”

Yup. My philosophy has always been that anyone can find explanations to questions, so simply explaining questions and concepts is not the core job of a tutor. I have worked full time tutoring the LSAT for over 25 years and worked with hundreds of students, many of them neurodivergent and all of them unique in their thinking. I have a vast working knowledge of the many unique ways students approach and think about various questions and know how to coach those unique approaches through the LSAT.

So when the models I was using finally reached that pivotal point in 2026—the point where I could craft prompts based on the different ways students might be responding to an argument—I was ready to craft something unique and pioneering.

Our curriculum has always been unique in that it requires a student to explain their reasoning for most question types, rather than to simply answer large sets of questions. I based the curriculum on years of educational research on the value of having students explain their reasoning. The problem was that often I was getting the student’s reasoning in batches and late at night, so by the time the students got my feedback they had already done several examples incorrectly and created a habit that they had to undo.

And in the summer of 2026 we created our first AI feedback check. At first, it was a bit too helpful, giving feedback even when a student’s response was basically correct. Its overly helpful persona would give suggestions on how to improve responses for grammar, spelling, and clarity. But as the models improved and our prompts grew more refined, it became the evaluation tool that I had dreamed of creating so many years ago.

Meanwhile, I’m hyper-aware of the environmental impact that AI has, as well as the other issues that it raises. My two children are both Gen Z and are very happy to explain how terrible AI is.

Recently, one of my students had a competitor’s website open, and I was flabbergasted that it offered an AI tool to create a study set of questions. It said, “Ask AI to make you a set of eight hard weaken questions that use causality.”

Why would you do that when a simple non-AI computer program could accomplish the same functionality? Especially when you could instead use AI to build a program to pull questions by category, so that the AI use happens during development rather than every time a student puts together a practice set. The finished program still uses energy, but it doesn’t need AI to select the questions.

So my core philosophy for our student-facing tools is that AI should be used for projects that need AI in order to work. As a company that has always valued our human touch above everything else, AI will never replace the humans you interact with.

Back to top