Ask HN: Have students started questioning their curriculums due to ChatGPT?
In particular, what is the point of learning something that is answerable by an AI such as ChatGPT?
How do teachers reply to such questions?
Are teachers starting to change what they teach?
68 comments
[ 3.0 ms ] story [ 134 ms ] threadIn the long term, assuming AI takes over a significant share of knowledge work, curriculums will need to adapt, just as they adapted to computers and calculators.
ChatGPT will often make up an answer which is absolutely incorrect. It will even insist that it’s correct.
It appears that the only responsible way to use ChatGPT is for questions where you already know the answer or can quickly validate the correctness somehow.
You're assuming this will be the case forever... check back in a year :)
...just like [some] humans
Basically everything it says is a lie, some of those lies happen to randomly be true.
Edit: This also might be helpful: https://learn.microsoft.com/en-us/azure/cognitive-services/q...
Yes and no, the goal wasn’t focused on just being a source of truth. They wanted it to be able to write fairy tales, poetry, and song lyrics so creative answers where rewarded even if they weren’t factually accurate.
There was an attempt to reduce harmful and deceitful responses, but the team acknowledge the tendency for “plausible-sounding but incorrect or nonsensical answers” calling them hallucinations. Which becomes common when prompts are outside of it’s training set.
So yes there is definitely a link between the internal “Correctness” score and “Truth” but they are measuring very different things.
People usually do act as if they know what they're saying is incorrect, especially if it's a case of lack of recollection rather than a conscious and strategic lie. ChatGPT defaults to offering a continuation with the correct format and plausible text strings, without regard to truth values it doesn't know because it's indifferent to them.
And even if human learning or human motivation did lead a to person to internalise nothing about, say, the titles of scientific papers but the general format and authors well enough to make up similar sounding ones on demand in future, they probably wouldn't be as good as it...
It's the age old question, if you'll wear diapers when you're 90, why bother with toilet training?
Critical evidence-based thinking and the ability to properly apply understanding to solve a problem, and validate or disprove a proposed solution is something that a language model struggles to do, as it can't apply the reasoning process on facts and circumstances and context.
GPT output can be very believable and confident, but I would expect a good student to be able to question and interrogate a position and establish whether it is accurate or not, based on an evaluation of the available evidence.
Education moves slowly as a sector, and I doubt many have changed anything in the last month or two, but the pandemic and a shift to remote exams or "open book" exams has (for good courses and educators) resulted in more of a focus on explanations, justification, application of knowledge to new areas, and critical evaluation. If you could copy it down in a book and repeat it in an open book exam, chances are a language model poses the same threat.
It's a great idea generator & tool to bounce things off of to use as part of your own execution in producing your own work. It's hard to find people with the right skills, who have the time & same schedule to do this.
I do wonder the extent to which GPT-type tools could become teaching/learning aids - where a student asks for 5 arguments both for and against, and then fact checks them, figures out provenance of arguments, and validates them. The understanding that a student will get from taht process will, as you said, drive you to look closer at things, and ultimately if you are critically evaluating it, you'll probably learn about the subject pretty effectively.
In the same way that it's often said the best way to learn something really well is to teach/explain it to others, I think the same could be said for carefully and accurately rebutting something and explaining why it's wrong.
Yes, exactly this. It can be hard to find this though. I'm unable to get this out of Google or social media search engines. GPT is doing the best for me so far at helping me find them.
> Indeed - but you still require that "subject matter expert" ability to actually evaluate the output and figure out if it's coherent or correct.
Yes, you still have to do the work but knowing what to look into & where to put the work is much more than half the battle. It's giving you an opportunity you may never have had before which I think is as big as the Internet or written books or programming.
It dismays me seeing such poor takes everywhere about how they think education will falter and people ask questions like this, it's almost comical.
Perhaps people should start to learn that education is about learning how to learn and applying that to new concepts.
This last point means it doesn't particularly serve people around or above the median, which is why, in fact, it feels to many of us that the education we experienced was about learning rote facts.
Now maybe that's not the education you received, but it is the one I received. Asking if students have started questioning it is in fact asking if students are using it to question the value of rote facts (as students have rightly been doing since forever anyway.)
I guess GPT-3 is more akin to a calculator. And yeah, tons of kids ask “why should I learn [this math equation] when I can type it into a calculator?”
The difference is the calculator is guaranteed to be correct if the inputs are done properly.
It’s still worth learning the “why” of a non-solvable topic for intellectual curiosity and because there’s no bottom or border to these topics.
But yeah, for stuff you aren’t interested in, I can see lots of kids wanting to use GPT-3 to just “get through it”.
That said, qualitatively it feels different because it can produce an answer specific to your question. Like, a lot of 'Googling for the answer' involves putting puzzle pieces together and combining that with what you already know to get an answer. In contrast, it feels like if you wanted a better program you can copy and paste your code into ChatGPT, ask it "How do I improve this?" and it spits out an improved version with a short explanation. No more poring over web pages with advice about code quality, no more trying to figure out how to apply that advice to your specific program, etc.
So at this point it's not clear where things are headed, but it certainly seems reasonable that this may be qualitatively different.
Moreover, while ChatGPT is really good, it can make mistakes, and if said mistake isn't an obvious one, then you'd need to have a good general understanding of the thing you're asking a question about to spot that mistake.
We should also consider that ChatGPT won't be free forever, so it might be a bad idea to change teaching in a way that assumes that these kinds of things will always be accessible to students and wherever you're going to work.
If recent trends are any indication, we should consider that ChatGPT will probably be supplemented by a number of open source GPT models, likely within the next 12 months, and a rich ecosystem of developer activity will continue to evolve and flourish.
Also, ChatGPT in particular is a terrible "source of truth". It gives plausible sounding answers, but they may not be factually correct.
I just asked ChatGPT to write a press release for a new scientific discovery, which would be impossible to actually happen. It happily did so, making up a name for a new element along with creating a fake quote from a made-up researcher at a made-up National Laboratory. You could do the same for historical events, current institutions, anything.
As most (all?) researchers will tell you, asking good questions is the often times the hard part.
A good curriculum, sadly rarely found, will teach you how to think and how to ask good questions. Chatgpt doesn’t really impact this, imho.
A simple answer would be to have students take handwritten exams in class, with no electronics allowed. This is not always possible (ie online classes). Homework has always been subject to a number of methods that relieve the student from the burden of actually learning the material (ie parents helping, copying other students, searching online, etc).
Online classes are a tougher problem to solve. Maybe the answer for that is requiring that they be paired with proctored exams.
Teachers aren't all that original in the topics they ask students to write essays on. Often, a teacher or professor will use the same questions for years, and they usually fit common themes, books, literature, that other teachers and professors frequently cover. So it's usually trivial to find resources online for this. I've even seen university level exam questions shared in various forums so that students need not do anything but rote memorize the responses.
seems like the wrong question
I've largely given up on trying to "catch cheaters." At this point, either a student wants to really learn the information or they don't.
What are you gonna do? Cheat your way to a programming degree and then... bomb at some programming job because you don't know how to actually code or think?
I try to set students up for success, w/clear & obvious requirements (e.g. test suites for project grading) and then appeal to their sense of reciprocity: I'm trying to be cool to you, please be cool w/me & don't cheat. If you feel like you need to cheat to pass, come talk to me so we can address the material instead.
Regurgitating was never a great way to measure understanding anyway. In an ideal world, I'd have the time to sit down with each student and have a 10 minute talk, and get a sense of how well they understand the material, like the old english system.
Aren't in-person exams a decent way to gauge that?
I don’t really understand the uptick of “Rate how much the idea of a confident-sounding chat bot has shaken your world view to the core on a scale from 1-10” posts. It’s a language model, not a New Internet.
In most cases, homework / basic questions are to strengthen your knowledge. If you are letting the AI do this for you, are you able to have a conversation about said topic?
Changing what teachers teach will not change the problem. Students will always use this bot now. The cat is out of the bag.
People have been having this same discussion about math for many years. A copy of Mathematica can answer just about every question on every math test you will ever take. So what? We didn't stop having to learn math in order to understand anything that's based on it. Mindlessly quoting answers from ChatGPT will not teach you new ways of thinking.
As for the knowledge itself, the key thing to understand is that knowledge inside your brain is more useful in more circumstances than knowledge outside your brain. ChatGPT won't notice connections between unrelated subjects for you. It won't notify you in advance when a question ought to be asked.
Because understanding and looking it up are not the same. We've been able to look up things since we've had libraries and experts.
One one hand, you may end up doing something really stupid/dangerous just because you don't know what the answer means or what are the implications and blindly do what the statistical model tells you.
On the other hand learning gives you "HD vision" for the world.
Anything new you learn will let you see much more detail of the things around you, and you don't really know what you're missing, you really can't know, because until you learn it, you're blind.