NPTEL-NOC IITM · Transcript · · Updated
1. Role of AI in Research - 1 – Transcript
Video by NPTEL-NOC IITM · Watch the original on YouTube
Summary
AI-generated from the transcript. Verify details against the source video.
- Professor Karthik explains that AI functions like a tool that can decompose problems and automate tasks, requiring researchers to understand how it works rather than just using it blindly.
- Researchers must build expertise before augmentation and always verify AI outputs because models can still produce fabricated references and errors.
- Human authors retain complete responsibility for their published papers, figures, and manuscripts since AI is not recognized as a co-author.
Use this video with Pro
Reading, copying and transcript downloads are free. Create saved AI results for this video with a signed-in Pro account ($9/month).
Transcript
Search ignores accents. All transcript text stays visible.
Hi everyone, welcome to another episode of research methodology. Today uh we are going to have a uh very important question to address which probably uh will be interested not only for young researchers like you but also for established researchers, professors across the globe. Yes, we are going to talk about the role of artificial
intelligence in research and we have an expert here uh professor Karthik from department of data science and artificial intelligence IIT Madras. Welcome Karthik. >> Hi.
>> So um see when I was just checking uh random videos in YouTube about AI. So I'm getting two different kinds of videos. One set is saying that no as humans where you are done AI is going to take over. That is one set of videos and they are giving numbers also 2030 2027.
So another set is saying that just you know orthogonal. They say nothing is going to happen. It's just a tool. So my first question to you is is AI is ES AI actually a tool or it is beyond a tool? Yeah. So this is a very interesting question and there are lots of debates around it and probably nobody captured it as well as I think Arvin Naran in his ICML keynote a couple
of days back where he was talking about you know what AI what will it replace what may it not replace and uh things like that but um so clearly AI is able to do a lot more than we originally thought it could do >> okay >> right I think originally the idea was it was um at least in the initial stages of even in the LLM era right it was quite a bit of glorified autocomplete >> okay >> right you you have something it'll it'll predict the next token and things like that
>> but it looks like what AI is doing very well now is how it is decomposing problems right which is also a sort of human thing to do um of course uh it comes with a lot of training right how do you break down a problem break a large problem into smaller pieces and solve the smaller problems and put them back together and so on and it is something that you do routinely in research >> right and looks like AI is uh this well I'll probably come back to some examples later on >> okay >> but um because you said tool right I
think there are two aspects that we need to really uh clearly figure out at the set uh at the very beginning which is uh the use of AI should not be just the use of AI tools >> okay >> right I think especially as researchers who is I think the major audience for this uh uh discussion uh people need to understand how AI works >> okay So, and it's it's useful to think
of this in a few uh uh steps. So, there's first AI literacy >> and then AI skills. >> Then maybe you have knowledge and higher level understanding of AI and ultimately you build AI tools and you are an AI developer. >> Right.
>> So, in fact AI literacy and skills we are really focusing that it should come in at the school level itself. >> Okay. >> Right. So and I was recently part of the the CBAC committee that uh proposed the AI curriculum see computational thinking and AI curriculum for classes 3 to 8.
Okay. >> Right. So 3 4 5 it's mostly puzzle solving and computational thinking and 6 78 it's basics of AI >> right at but at no point here have we talked about the tools. >> Okay. >> Right. So because anybody can go to chat GPT put in something and get something and start playing around with it. But the thing is do you start understanding what is actually going into it >> only then you can know what to trust and what not to trust in some sense. So then then this question I'm having is okay
whatever your the the um road map that you're saying is that first build an awareness right and then try to see what AI is all about then use it as a so but then somebody for example from a chemistry background or a sociology background do you recommend the same way I mean like uh >> yeah yeah so so there are these so in fact we have developed this AI literacy framework again for schools >> okay
>> right giving a very very simple introduction to AI including you know issues about AI ethics biases and things like that >> so people should not blindly use AI right and I think for the longest time since AI came in people have been talking about hallucination and things like that it's less of a problem now the models have really improved you see less and less hallucination today and of course we are also very cognizant of these challenges and we are verifying the A output quite actively but this is exactly what the students need to be
taught >> never trust always verify. >> Okay. Okay. >> Right. So use AI. Use AI to automate. >> Right. And this is again a very important thing to think about. Right. So what what is how is AI fundamentally different? And I'm going to share one of these articles written by one of my collaborators Arjun Krishna. It advocates a lot of common sense. But it it draws some very interesting parallels and so on. AI is not like using a calculator.
>> Mhm. Yeah. Those are the differences. >> A calculator is your uh you automating a mechanical task, >> right? Suppose you have to add 100 numbers, >> a calculator will do it uh with fewer errors or like no errors. A computer would do it with no errors.
>> On the other hand, in AI, if you're trying to use A to write an essay, >> A is going to make a judge be the judge of what is important to include, what is what is not important to be excluded by >> so the cognition part is now outsourced to the AI.
>> And there are two issues here. One is as a student >> you have lost your opportunity to build that muscle, >> right? which is very very critical as a student and two bigger fear is that you're you're now going to lazily lean on that >> and you're not going to ever cultivate these muscles and worse still you might be passing off the essay as your own >> these are the challenges that teachers have had to deal with right so I get essays in the class >> and some of them scream AI some of them
don't scream AI but could still be AI but there is no meaningful way to trust take-h home assignments any longer in a course >> so these are some of the challenges but the the key aspect effect is as Arjun says it is um build expertise before augmentation >> you can augment if you are an expert if you're a professor you're somebody who's already and again it doesn't apply to you on me trying to pick up a new area suppose if I try to pick computational chemistry >> I I cannot start using AI to learn it
right I need to go through the motion I need to go through the grind >> because that is what builds skill I think we've had this argument for a long time right so should people uh you know go to a library or not Uh-huh.
>> Right. And in certain cases, some of the arguments still hold, right? So the reason to go to a library is that you'll see 10 books next to the book that you're looking for and you will learn a lot more and so on. >> And similarly, when you don't when you don't get a packaged answer, >> the pro the the process of arriving at the answer is actually so important in research.
>> Yes. >> Right. Research is contrary to I don't know what people think contrary to popular opinion or perception or whatever research is not about solving problems. It is about learning how to solve problems. Especially for students that's exactly what we want to teach them. You're confronted with a difficult new problem in a new domain. You have some expertise. You need to pick out, you need to figure out whom to collaborate with, what tools to use,
what experiments to learn, what uh ideas to outsource from other domains to solve this, right? Ultimately, your goal is to solve a problem. But solving the problem is is not everything. >> It is the the path the complicated contorted path to that problem that really counts.
>> Exactly. So in that context then suppose what will you tell to a young researcher who is just you know beginning the research career. So you said expert become expert first and then augment it. So now the person is probably working in a college or working maybe in that area.
He may have developed some expertise for now he's entering to research or she's entering to research in that particular scenario. How about start with this uh AI interaction? So uh is there any guideline that you can suggest to them?
>> Yeah. So um I think if you are doing literature review for the first time >> Mhm. >> don't use AI, right? uh a is very good at doing literature review but you should also learn some of the tricks of the trade otherwise those tricks will be forever lost right so you should talk to your advisor talk to your seniors and understand how to do literature review what are the things to watch out for what are the you know other terms that I should look for what are the errors I might uh do right how do I organize my
literature search >> okay >> you need to do a bit of that but if you've done that a few times you've done it you've done it for two projects >> in the third project you do it >> but now open this superpower right you use AI augmented right you're able to see how AI can pull from so many more databases it can walk through some pay walls maybe it can give you a lot more information you can quickly ask am I justified in assuming that this hypothesis is correct based on the information that you have put it could
answer something you can maybe fetch another set of papers based on that >> you can do a lot of those things but if you haven't figured this out and I think we also need to train for some sort of a weird future, right? What if one day everybody agrees that A is bad for the environment and shuts AI down?
>> It's it's it's quite possible, right? Forget anything else. It's definitely A is bad for the environment, >> whether it is bad for humans or not. >> So, how how do you react? How do you work with in a in a in a resource car situation, right? So, you cannot become dip and a you can a cannot become your crutch, >> right? It can be like you know something that helps you do better but it cannot be a crutch without which you can't even work.
>> So I mean so now that is clear. Now next point is uh probably in a research framework a student will do so many things like literature surveys beginning then he has to or she has to develop a hypothesis and then data collection data analysis conclusion. So in all these process do you see a a gradation of application of AI like you can use it here for more therefore less which will be uh in that sense
>> very good question and definitely the answer is yes it can be used everywhere and eventually it will be used everywhere uh whether it should be used everywhere I think is again a training question >> okay >> right so um and uh it's it's always going to be a challenge because there is this notion that you can do it faster using AI so why not use it. But in any case, I think what I said earlier just always holds uh always trust uh always verify never trust.
>> Yeah. >> Right. You have to because uh I see some astounding numbers, right? So people have studied uh literature over the last 2 three years and they found something like um one in every 450 references was fabricated or something like that.