Kids who use ChatGPT as a study assistant do worse on tests

However, I don't always agree with the fact that we don't have the answer in front of us in many situations. There are a lot of situations beyond grade school and undergrad where you do have the answer in front of you. Sometimes it's in for form of Numpy or Matlab or Simulink. It might be someone's research publication. Replicating these by working both forwards and backwards from their libraries or results can be much more effective.

How spacing is utilized can be helpful too. Struggle with the problem for 20-30 minutes. Ask for a nudge. Struggle some more. Repeat many times.

Some concepts also just have to be thought about differently to get to the aha moment, especially in math. AI may have an opportunity to present many instances of “think about it this way instead”.

this sounds a lot like the "when you're an adult, you won't always have a calculator in your pocket" line we all heard in elementary school. and of course, we all know now how wrong that was.

in life, chatGPT exists. and tools like it are only going to become more widespread. we are going to have the answers in front of us. knowing the things an LLM knows is not a useful skill anymore.

I mean, any pre-prompt telling the LLM to be a good tutor and not just give kids the answer can be trivially bypassed by 8 year olds; but if the session logs were available for (potentially LLM-based) review to check whether they stayed in tutor mode...

Shit, I should wrap a UI around this and sell it.

Very few of us have the luxury or good fortune to achieve anything of substance without effort, as the old saying goes 'no pain, no gain'.

Good teachers help and are essential for encouragement. If or when AI becomes a truly encouraging mentor then it will be useful.

When a kid I recall hearing a recording of Sparky's Magic Piano on the radio which drove home that there's no shortcut and that effort and hard work eventually pay off. For anyone that hasn't heard it there's a copy on the Internet Archive: https://archive.org/details/78_sparkys-magic-piano_henry-bla...

Here's a Wiki synopsis: https://en.m.wikipedia.org/wiki/Sparky%27s_Magic_Piano

It’s a tool that can be used in a variety of ways. For example, if you have it act as Socrates working through a dialogue where you still have to use your logic to get to the answer, I doubt that is akin to looking up the answer.

The takeaway should be (besides the research still needing reproduction) to encourage the control of the type of AI agent that is given to students, ones that don't just give answers to copy but provide tutoring. OpenAI should be forced to develop such a "student mode" immediately and parents and educators need to be made aware of it to make sure students are using it, otherwise students are going to get much worse in tests, as they just ask it for answers to copy in assignments.

This effect likely comes from novelty being more interesting so kids gets more alert, but when they are used to it then it is the same old boring thing and education results go back to normal. Of course things can improve or get worse, but in general it is really hard to say, you need to have a massive advantage over the standard during testing to actually get any real improvements, most of the time you just make things worse.

In what fields? I’ve tried this with some simple finance and aerospace problems; it’s me, sophomore year undergrad, except someone laced everything I drank with LSD.

Are you validating this in a rigorous way like this study, or are you just "feeling" like it's useful for your learning?

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4895486

> However, we additionally find that when access is subsequently taken away, students actually perform worse than those who never had access (17% reduction for GPT Base).

The abstract continues:

> That is, access to GPT-4 can harm educational outcomes. These negative learning effects are largely mitigated by the safeguards included in GPT Tutor. Our results suggest that students attempt to use GPT-4 as a "crutch" during practice problem sessions, and when successful, perform worse on their own. Thus, to maintain long-term productivity, we must be cautious when deploying generative AI to ensure humans continue to learn critical skills.

Someone educated by the News Feed algorithm (which is what RLHF amounts to: reward for getting human to click) is going to be the worst kind of wrong: /r/ConfidentlyIncorrect.

If this is your takeaway you misread the paper. Students have access to GPT (if they have access, the control didn't) while working through practice problems. Not for the exam itself. From the paper in the experimental design section:

> Each session has three parts:

> 1. In the first part, teachers review a topic (e.g., combinatorics) previously covered in the course, and solve one or more examples on the board. This part is identical to a standard high school one-to-many (i.e., teacher-to-students) lecture.

> 2. The second part is an assisted practice period, where students solve a sequence of exercises designed by teachers to reinforce the covered concept. Our randomized intervention (described in more detail below) only affects this second, self-study part.

> 3. The third part is an unassisted evaluation, where students take a closed-book, closed laptop exam. Importantly, each problem in the exam corresponds to a conceptually very similar practice problem from the previous part—this design was chosen to help students practice the key concepts needed to perform well on the exam.

Students with GPT (either form) did better during the practice problem portion and then worse during the actual exam (without GPT access) than students in the control.

No, ChatGPT does significantly better than the kids who don't have access to ChatGPT.

Copy-pasting answers from ChatGPT isn't some amazing skill.

https://onlineteaching.umich.edu/articles/the-myth-of-learni... https://www.youtube.com/watch?v=rhgwIhB58PA&t=2s

There isn't any evidence that individuals learn best to a single style and generally approaching learning from multiple facets is the best way for everyone to learn!

Many users of ChatGPT clearly know that it doesn't do math.

Now imagine your teacher is wrong at solving 40+% of problems it’s teaching you? Or that your measuring tape is wrong at nearly every second measurement?

Yeah it’s a tool - but you must nail your fundamentals right. Where I grew up - calculators were not allowed in elementary school - so every student must nail down basic arithmetic.

I remember when I started there the old guys would point at stuff and say the length of it, and I was just amazed that they were correct, even if if it was a long 4.8 meter long plank of wood that they had never seen before. I gained that ability after like 3 months there and I still have a strong ability for it now. Weird random ability.

I bet all the wealthy and middle class parents with STEM background will get tutoring for their kids in the “old ways” knowing full well that the people who aren’t reliant on AI and can spot mistakes in output will now have a huge advantage in the workforce.

> Student performance in the GPT Tutor cohort was statistically indistinguishable from that of the control cohort, and the point estimate was smaller by an order of magnitude (-0.004), suggesting minimal impact to performance in the unassisted exam.

"GPT Base" would provide complete answers, but "GPT Tutor" was prompted to provide only hints. So the result is perhaps that:

1. Given the option to let a machine ("GPT Base") do their homework, many kids will lazily take it. These kids won't learn as much.

2. A machine that refuses to do their homework ("GPT Tutor") doesn't cause that problem. It doesn't seem to help either, though.

I'd guess that laziness would explain most of the harm rather than mistakes made by "GPT Base", though I have no particular evidence for that. Maybe someone will repeat this study in a domain where "GPT Base" makes fewer mistakes, allowing those two effects to be distinguished. (Though would that pass ethics review, now that "GPT Base" is known to impair learning in at least some cases?)

At best they do the same, at the worst they do worse. There was no scenario where ChatGPT actually helped (unless you count practice problems, which are apparently a meaningless stat, since it doesn’t translate to test scores).

Most kids in the general public would use normal ChatGPT, not one specially optimized to act like a tutor. So in the real world, kids who use it will do worse, and it’s not with seeking out the special one, as there is no advantage.

The title seems like a pretty fair 1 line summary, and pours some much needed cold water on a space that has been full of hype and promises, many of which have fallen flat.

They fared worse once that access was taken away during the actual exam.

Working with a customer/client? I don’t think you’ll take a minute to say “give me one second let me ask ChatGPT”

Someday I noticed that I would automatically hit up the search engine within seconds when I hit a road block. It's kinda the same as with GenAI. You just have to know how to query search engines efficiently. I got stuff done, but I often did not learn much and sometimes didn't even understand what I was doing. So I started to force myself to use data- or reference sheets and try to come up with a solution myself. And if that wasn't good enough or I had real trouble, then I tried using some different resource. And it improved my programming experience tremendously. It's hard, but it's worth it.

This is different fromt students relying on ChatGPT to pass tests.

Your use is much more adulty and you're trying to understand before proceeding. You use ChatGPT like a tutor rather than a calculator, which is improving what you know rather than taking from it

LLMs are great for finding key words in a domain you don't know which you can then use to search.

As a mostly self taught programmer the advice I'll give it to read docs and learn to read code. This is when my skills really increased. It's easy to get caught up in trying to just find the answer and doing these things definitely takes longer. But it has a multiplicative effect. You'll not only learn to be able to do it faster but you'll need to do it less often. So never pass up the opportunity to learn. There are times you need to rush but they're far less often than you think.

The other big advice I have is two strategies when coding. First write it quick and fast. Sloppy is okay, you're learning to solve the problem (move fast and break things). Then while you're cleaning up the mess you made, document. While you're doing that it'll stress the lessons into your memory and you'll almost always discovery things you missed. It is literally the rubber ducky method. Plus, you get the benefit of docs (others will read your code. And months or years later you'll come back, asking what idiot wrote this garbage to only find it's you. But also, that's a good sign, because you improved!). The second point is that the best code is flexible code. Remember that you suck at code (the secret is we all do) and you're going to have to come back, edit, and debug. So by writing flexible code you're making life easier for future you. It's really easy to forget these lessons because we think in the moment and our egos don't want us to admit we're bumbling idiots, but like in a video game, when it gets harder it means you're progressing. Don't misinterpret that signal.

> One of the co-authors, Özge Kabakcı, a high school math teacher and former department chair of the math department at our partner Turkish high school, led the development of all session materials.

But why Turkish not British or any other place is going to be a question no matter the location. But do you really think the results will be significantly different if it is done lets say on Vietnamese students?

I’ve noticed this even using Copilot in VS Code. I rarely use it, but if I start pulling it out, I notice at the first hint of actually thinking about how to do something, my brain seeks to ask Copilot instead. It’s like there is an off switch that gets flicked when there is an easy button available. If I were to figure it out on my own, I’d know what to do next time I run into a problem like that… if I use Copilot, I’ve learned nothing, other than next time I run into this, use Copilot.

It’s a crutch when it comes to learning.

ChatGPT worked correctly for them but they learned nothing. It should be pretty obvious that you don’t learn by copying answers.

Your characterization of "do their homework" seems wrong as the paper says of the GPT Tutor branch that "is hard for students to use it as a crutch since its prompt asks it to avoid giving them the answer and instead guide them in a step-by-step fashion."

I’m thinking of science education, for example. It’s not just a matter of teaching kids a bunch of disconnected facts about science like “the mitochondria is the powerhouse of the cell” but teaching kids to use the scientific method. To teach the scientific method well, you have to get the kids to design experiments and be comfortable with the prospect of failure. This is, sadly, a very modern approach and requires training the teachers how to do this. Some schools don’t but it’s not something you would have likely gotten much of in the 20th century.

Instead, people are still fighting over dumb shit like which facts are taught in class. “Critical thinking skills” are not on the menu most of the time.

I think new tools are great, but it seems to me that the people who can utilize them the best are the ones who learned without them. Those are the people who understand the concepts behind them, so they can ask the right questions, with the right terminology, and they can recognize if an answer makes sense.

When learning is done with these tools, even just calculators, little is being learned. Memorize the button sequence to press leading up to the test. Once the test is over, forget the sequence… and even if the buttons are relented, it won’t matter, because while we all walk around with a calculator in our pockets now, it’s not a TI-83 (or whatever people use now), so without the conceptual knowledge, what good is the button sequence. That doesn’t even prepare a person to leverage ChatGPT well.

As a ML researcher myself I'm often baffled with the direction we go with these tools. It feels like we try to use then to do the things humans like and are best at and then have humans do the things ML models are better at and humans hate doing.

The tech is no doubt exciting (it's why I research it!) but that neither justifies unbounded hype nor blind ourselves. But I guess that's the classic engineering problem: it's easy to get lost in the good and exciting parts of what you're building and lose sight of the harm it can do. It's hard to not fall victim to this and I'm sure we're all guilty of it to some extent, I know I am

No LLM processing required, just old fashioned critical reading and comprehension (my emphasis added):

>> The students who used it did spectacularly better on the practice problems, solving 127 percent more of them correctly compared with students who did their practice work without any high-tech aids. But on a test afterwards, these AI-tutored students did no better.

In my experience, usually where GPT and other LLMs get things wrong is in the steps. I see people are frequently over confident about the results and I've long thought this was a big part of it that the answer may be right but steps to get there are wrong. I see this a lot with river crossing problems. People will often show a prompt that solves it (half the time there's information leakage) and a good portion of those will have the logic errors. But people just look at the answer (btw, this is the reason I claim LLMs don't reason. It's not about getting the answer wrong, it's about the logical steps. That's the real evaluation!)

Personally I think this is good evidence of over fitting (memorization).

I also think this is why you should be careful when using it to code. It's the details that matter a lot. Since LLMs are aggregators I'd like to remind everyone that the average coder and average code is terrible. Hell, even good programmers often suck. Code is hard! Whenever I try using LLMs to code I find that I write lines faster but I end up spending more time debugging and prompt engineering than if I just read the docs and did it myself (this also has the added benefit of the struggle making me remember more). So I'll use it to write things I don't really care about but otherwise I'm not getting how people are finding it so helpful (yes, I read my comment). But I suspect many people are similar to the students in the study.