I have always had this feeling of guilt when using LLMs to produce code for me. It has always felt like cheating to me. You just stop thinking about the problem and jump to the solution with a prompt. Over time, I have abused LLMs for things like university projects and assignments where you just offload them to an LLM, and get a reasonable output. It is virtually not different than submitting someone else’s code as your own.
This has lead to me losing track of how using AI responsibly looks like. I constantly question myself:
Am I even learning anything? Am I being lazy? Am I tricking myself into thinking that I’m progressing?
This has led to me thinking that we should have some sort of rulebook for using AIs, or maybe some sort of optimized AI for learning that refuses to answer questions when your learning might be bypassed. This article from Stanford has done some research on how using LLMs can harm learning and, they clearly show that using LLMs for substituting tasks that require mental effort result in less understanding.
Some also claim that the studies don’t really apply to their case as an adult who is trying to learn new topics. The context matters a lot here, whether you are trying to use LLMs for academia or your own knowledge.
Here are some of my thoughts:
Jumping to solutions is bad
There is no doubt that this is wrong. If your goal is to learn what’s happening behind the scenes, simply using AI to just get the solutions for specific problems on the first try is wrong. It might be justified if you have been trying to solve the problem for a while, and you need a little bit of hint to get it done. It is always mentioned that it is the process of constantly struggling to find solutions that facilitates learning.
Struggle or not?
Now, many use AI to understand concepts better. It is very convenient, you can ask it as many follow-up questions as you want. It is like a personal tutor, and it can address any hiccups that you have.
But here is something that popped up in my mind when browsing online:
If we skip the “struggling to understand” part of learning and address it by using LLMs, will we harm our learning process?
Maybe I’m being too sensitive and overthinking this matter as well. We have to get the new information from somewhere somehow. This again, leads to understanding the importance of a rulebook of some sort to follow to make sure that the learning process is not damaged. Of course, this requires proper understanding of how we learn.
AI’s opinion
It’s quite ironic that I just asked Claude’s opinion on this blog post. It actually offered some good points regarding this blog post:
- An alternative to having a rulebook would be to have a routine while learning or addressing problems:
- Before getting answers from a model, attempt to think about it beforehand and write your own thoughts on the problem. It doesn’t matter how bad it is.
- After getting the response from the model, try to explain it to yourself afterwards or attempt to reproduce the result on your own without any help from the model.
This Approach really reminds me of integration testing. Instead of limiting the model or constraining yourself to a set of rules, just see the end result. Can you recall the material on your own?
Conclusion
I would like to point out that I am not Anti-AI per se. I believe that AI can actually be a helpful tool. You just have to be very concious when using it and have a specific goal in mind when using it. I have heard this statement from a stranger online that I think is very relevant here:
There is a very thin line between learning and cheating when using AI.