AI as Teacher Not as Cheater

Cheating with AI is easy. But how can modern LLM-AIs be used for teaching?

The key to using AIs as teaching assistant is understanding that an AI is more A than I. There is no true intelligence behind the words they print on screen after a prompt.

The nature of AIs

A Parrot, much like an AI, does not understand the words it mimics, only that when it says certain words it gets a cracker. The parrot is rewarded, so it says more of those words. If the bird is not rewarded, or even punished, it will say less of these words.

An AI is a Stochastic Parrot¹. But instead of being limited to the words they hear from their owner and from the TV, they are trained on significant portions of the entire internet. That means that they have quite a few more words they can try out to see if they will be rewarded or punished given a certain prompt.

Effects on prompts and replies

Because of their nature, an AI will respond to a generic prompt with a generic answer, and to a specific and detailed prompt with a specific and detailed answer. Lets use a practical example: What do you believe the response to this prompt will be?

Create 10 exercises for maths, curriculum: IB/MYP grade 10, subject: Quadratic equations. Append a solution section after the exercises. Produce the output in LaTeX.

Ignore the last line for a second, LaTeX simply allows me to generate good looking PDFs

This is the reply:

Those exercises may be too hard or too easy, depending on the student’s current level. They may be very complex or rather simple. Only the roll of the dice will determine that.

How could we be get more specific answers? Well, part of the training set on which AIs are trained includes not only most school subjects and curricula of this world, but also an idea called Blooms revised taxonomy . So lets be more specific:

Create 10 exercises for maths, curriculum: IB/MYP grade 10, subject: Quadratic equations.

Using bloom's revised taxonomy ensure a mix of exercises at level 3-4 for the most part with only a few at level 5. Briefly explain why each exercise is at the respective level in Bloom's revised taxonomy.

Append a solution section after the exercises. Generate the output in LaTeX.

With this prompt, I received this reply (Note how the AI tries to explain why some exercises are more complex than others.) :

With this tool in hand, we can now generate exercises at a much more precise complexity level, all the while keeping the difficulty and grade level the same:

Repeat your last answer but produce exercises at level 4-5 of bloom's revised taxonomy only. Keep everything else the same.

This leads to:

By the way, all prompts and replies can be found here: DeepSeek Quadratic Exercises Examples

Conclusion

AIs are not intelligent, only very fast and very knowledgeable. To use them to learn, take this into account. Prompt precisely and with many details to receive precise and detailed answers.