The AIhub coffee corner captures the musings of AI experts over a short conversation. This month we ask whether AI tools are changing the way we think. Joining the conversation this time are: Joydeep Biswas (The University of Texas at Austin), Sanmay Das (Virginia Tech), Rina Dechter (University of California, Irvine), Sabine Hauert (University of Bristol), Michael Littman (Brown University), and Marija Slavkovik (University of Bergen).
Joydeep Biswas: There’s just so much to discuss there. I don’t know how to start. First of all, it is definitely changing how we write code, develop code, think about correctness, and work on projects. Once upon a time, implementation used to be bottlenecked by writing the code. We are no longer bottlenecked by writing the code, we are bottlenecked in testing and understanding for correctness. As a researcher, there’s a lot of what I would call throwaway tools. For example, if you have one data set and you need to create one custom visualization for that one thing, and then after that it’s going to be useless. These tools are amazing for doing these one-offs. Before that, you’d have to invest human time and effort and often it would just not be worth it. But this is one way in which I think it is actually helping us be better researchers.
The other thing that happens is that previously we used to be beholden to whatever tools we were given. I’m working on macOS, I have to restrict myself to whatever macOS allows me to do and what it does not do. Now I feel like I’m a lot more emboldened to actually improve the tools that I have at my disposal such that I can actually do better. For example, I really dislike the built-in dictation. It’s not very good, it doesn’t understand my accent. So I spent the weekend coding a dictation tool which uses OpenAI’s Whisper model and is so much better. And it actually has made me significantly more productive.
Sanmay Das: So I’ll go with the positive side before becoming a curmudgeon. I think these tools are useful in all kinds of ways that I hadn’t necessarily thought of earlier. I’m hearing from a few people that it helps getting started on things, which is interesting. Even if it’s something like needing to review a paper or these kinds of tasks. These are often things that I procrastinate a bunch or I don’t want to deal with immediately. And actually just spending five minutes looking at something and then having a conversation with Claude about it can actually be quite useful. Even if the things that Claude ends up telling me are not the most useful in the world, it can at least be a convenient substitute to having a conversation with a person to just get started, which is not something that I’d expected. And it can hold a relatively interesting conversation that at least gets you thinking about the issues. So that’s one surprising thing that I’ve found pretty useful. I’ve also found it quite useful for doing structured subtasks when I’m thinking about things, for example if I need an element of a proof. I’ve figured out workflows where I can get useful stuff out of that.
The other big area in which I’ve seen it being very useful is for people who deal with structured streams of data on a very regular basis and who are not necessarily in the business of running or setting up databases. Being able to work with AI tools has been super useful for people with these kinds of pipelines. Which is kind of funny because you would think that they would have thought how to automate this stuff earlier, but I think it’s lowered the barriers to entry enough. So those are the positive takes.
Sabine Hauert: We recently did a code hackathon with my lab, and one of my senior postdocs spent 30 minutes teaching us how he uses Claude as his research assistant, in a very deep way to do in-depth research with ideas that he’s had for a long time. And his main worry was that he would burn out from deep thinking, because instead of thinking and then coding, he thinks, and then it codes, and then he thinks, and then it codes. So he spends his whole day in deep thought. And so that made me think, this whole spiel about it replacing thinking, I think it does that for many things, but for some people, it does the opposite, where they end up thinking much more. And so that’s something I hadn’t thought of.
Joydeep: David Bau from Northeastern wrote this article and he really hit the nail on the head where he was talking about needing to be clear about what you’re looking for. And he called it the art of wanting: be precise on what you’re looking for. Think a little bit about what the measures of success are and be cognizant about the choices that you’re making. And these are even more important now with the advent of these tools.
Rina Dechter: As is usually the case, whenever we have an assistant that helps us do tasks we delegate to it, we lose knowledge about those tasks. We do that willingly, thinking it frees us to do other things. The same can be expected with an AI assistant. Since we’re using AI as an assistant now, the question is: to what extent is AI doing the thinking? We know it’s definitely doing the writing, so there’s a big concern about the deterioration of humans’ ability to write. Overall, there are pluses and minuses to using AI as an assistant. One plus is that it frees us from some chores. For example, with arithmetic, we no longer need to do calculations as we did in the past. Initially, there was a concern that people wouldn’t know how to do arithmetic, but we’re not concerned about that anymore. So the main question is one of balance: to what extent will AI assistance improve our ability to think? Obviously, it’s helpful in writing and editing, and may shift our activities toward supervising and editing rather than doing things from scratch. The worry is that by not doing things from scratch, we’re losing something in depth.
Sanmay: It touches on our prior conversation on AI and education in some ways as well. There’s now a ton of research out there. The one interesting piece of research was a math experiment where people using a GenAI tutor improved their practice scores. But once the AI was removed, those folks did worse in the actual tests. There’s this huge problem in any kind of apprenticeship learning model when you have used AI to convince yourself that you know what you’re doing when you actually don’t know what you’re doing, and therefore you won’t know when the AI gets it wrong. And that is one of the big worries – are we going to be producing a generation of people who actually do not have the basic skills? And at that point, what are we left with at the end of the day? We don’t really know what the trajectory of AI generation of stuff will look like. I actually worry about this homogenization of everything. The next time I have to read the word genuinely in some cold e-mail someone has sent me, I’m just going to throw something at my computer. Maybe there’s somebody out there who genuinely uses the word genuinely a lot but I’m suspicious. If everything shifts over to becoming AI and then everything gets homogenized and we are losing a generation of people with fundamental skills, it’s a worry. It’s the same worry of losing all these software engineering jobs, and then nobody’s going to know software engineering, and then we’re going to need to hire a whole bunch of software engineers two years from now to fix the messes that AI code generation is producing. I don’t know if that’s going to be the outcome, but it could be. I think similar concerns arise with academic literature. What happens if it gets entirely polluted by AI-generated slop?
Sabine: I wonder as well how we design these systems to avoid this non-thinking. I was recently at a workshop for a new project in manufacturing, and we’re thinking about keeping the human in the loop at all levels. And the way we keep the human in the loop in those discussions is that we ask for approval, or we ask the human to check something, or look at the text or the program that was generated. But all the thinking happens through the AI system, so there’s not a lot of thinking that happens in the check. In fact, probably people are just going to click on it and agree with everything. Whereas one of my colleagues was saying he likes Claude because it doesn’t do things for him unless he asks it to do so. So it’ll say “maybe you could write your code like this”, but it doesn’t write the code necessarily automatically. Maybe that’s just the way he’s prompted the system, so the human thinks more in that construct. I wonder how we create these systems so that the prompting is smart enough that you still do some of the thinking and not just the checking of what’s been generated.
Joydeep: To add to some of these thoughts, and channeling my inner curmudgeon mode. One of the concerns has to do with education. I really honestly feel like the people who know how to use these tools best are the people who have not used these tools for a long period of time. An example of that is having felt the pain of trying different things, it’s very easy to catch when these things go off the rails and rein them back. But I fear that if we surrender now to these coding agents and not think and experience these things, tomorrow when the systems are different, we won’t have the skills to be able to stop them the way I’m capable of doing today.
The other thing that comes to mind is that, historically, in computer science, before the current advent of coding agents, there’s been a transition of what skills people develop and nurture, and that’s changed over time. When I was in grad school, it would not be uncommon to have to understand how to invoke GCC from the command line, and how to write Makefiles. Nobody does that anymore. I was talking to some students and some researchers. They had this massive complicated job scheduling system in Python and I explained you could just write a simple Makefile and just do the same thing, and it’s much more efficient. So I feel there’s going to be an induced transition in the expertise. But what I wanted to say is that this is not something new. This transitioning in expertise has historically always been true in computer science.
Michael Littman: I wanted to elaborate a little bit on one of the comments that Sanmay made in terms of people using the tool as a crutch and, when it’s taken away, they actually are less capable than if they had never used it in the first place. We did, let’s call it, an uncontrolled experiment on campus this semester where an economics professor announced in advance that all his exams would be done remotely this year. And he got a lot more people taking the class than he had in years past. When the midterm happened, the average was about 15 points higher than it had ever been before. And that’s 15 points because it basically hit the top. It probably would have been more if you could go above 100. And so he announced that the final would be in person. In the final, the scores were 30 points less than in past years. And so not only did they do better on the midterm, presumably due to the use of chatbot tools, but they did worse on the final when the tool was taken away from them. It just left them completely unable to do the work of the class. So, yes, that’s real, it is really happening. The notion that we could be training a generation of students that don’t actually know anything is something we’re seeing in real time. I think it’s unclear exactly how it will play out, but we’re definitely seeing some of the early signs.
That said, I wanted to push back a little bit on what Sabine said about this notion that if you give people the AI coding tool, then they’ll spend more of their time thinking deeply. I think that there’s lessons that we can pull from some of the early days of computing when, for example, we went from batch programming to interactive and we went from slow compilers to fast compilers. I think what we found is that it absolutely changes the rhythm of when you’re thinking is happening and the depth to which it’s happening. But I don’t think it leaves people thinking more deeply, more consistently. Because what happened when we went from batch, for example, to interactive programming, instead of people spending a week staring at the punch cards to make sure that they were right, I think the depth of thought was actually reduced. With the punch card, you got this one shot and it cost you a week if you didn’t get it right. People really did put a lot of effort into that. Once it’s cheap, and you spend less time waiting for the system, you spend more time shooting from the hip – “I’ll fix that bug. Oh, that didn’t actually fix the bug. Wait, what was I trying to do again?” And so the depth of thought I think was actually reduced, even though you’re spending more time active. And so I do think that, for some people anyway, this notion that you can will programs into existence using chatbots is not leading to deeper thinking.
Sanmay: I think Michael’s trying to say that his generation were better and deeper programmers than my generation, because I definitely grew up in the generation of trying something, if it doesn’t work, try something again, rapid prototyping.
Joydeep: No, Michael is saying that his generation are deeper thinkers. This is an even deeper concept than programmers.
But I would say two things. Actually, even before the rise of these coding agents, in our introduction to computer science classes, we actually deliberately thought about whether we should put a timeout on compilers. Every time you hit a compiler or you run something, there should be a timeout where you can’t do anything else before then. I think that’s important. But I also think learning how to program on a whiteboard, which is the way I did it, has certain benefits. You actually learn how to think like a computer a lot better.
One other point is that there is actually a barrier and a gap between the availability of these tools. So the gap between the frontier model, the absolute largest model, (like Opus 4.8 or GPT 5.5) and the next tier is very big. When you get used to the frontier models, it’s very hard to go back, it’s like a drug. And secondly, the access to these tools is very disparate. In industry, people have access to an insane amount of compute for these tools, in research, we actually do not. And I worry a little bit about this accelerating gap between the haves and the have-nots.
Marija Slavkovik: So the disclaimer, I normally state that programming is against my religion, so I try to not do it. But I’m wondering whether we are worrying a little bit too much. There are plumbers and computer scientists. We have all done pluming, does not mean there is no need for computer science. The plumbers are going to download libraries and copy-paste code and stick things together. And you always have a whole bunch of students who are going to do it that way. I think maybe the biggest issue here is that I am dissatisfied by the lack of understanding in the general public about the difference between software engineering and computer science and AI and all that stuff. Because people think that computer science is just typing in some code and all of a sudden AI is going to do it, so what are we going to do? It’s never been that. When Dijkstra came up with the algorithm for nearest neighbor, it was such a big thing. Now, any first year bachelor with just a little bit of prompting is going to be able to come up with nearest neighbor or shortest path algorithms or something like that. I guess what I’m saying here is that it’s great that the tool exists for the people who don’t want to study computer science and just want to use computation, and are willing to take on the risks that some mistakes will happen. This is like when you’re using another person as well.
But there is room to be concerned about how we are educating students. I once gave them an assignment in databases to do something with an entity-relationship model and then enhanced entity-relationship models because I wanted them to know how much harder it is when you don’t have these relationships. And five of them wrote me emails asking why they had to do it because it’s so much easier to do it the other way. So this is a constant battle with the students. And it’s not about getting it done, it’s about walking the path. So I guess what I’m saying here is I think we’re traumatizing ourselves unnecessarily here because there will always be jobs for the plumbers and there will always be jobs for the computer scientists. I guess the only dissatisfactory thing here is that the world doesn’t understand there is a difference. I mean, prima facie, the more you interact with the subject, the more you figure out what the plumbers are and what the scientists are, and there is space and need for both.
Sanmay: So I agree with Marija completely on some of this, but I guess the reason I am still worried is in two parts, and this is mostly my inner economist coming out. The thing is, it moves the margins. The first class I ever taught I told the students that I wanted them to work together, but I wanted them to write up the solutions alone. That’s because it’s really easy to convince yourself you know something when you’re working with somebody who’s really smart. Then you’re not going to be able to recreate that in the exam. That kind of worked for some share of the students. So I would have a class in which the top five students were fantastic, and then there would be another 10 or 15 who understood my point and got it, and then there would be another 50 who just did the problem sets by copying the idea off of someone else, and then didn’t do very well on the exam, but did enough to get by and maybe get a job as a software engineer somewhere. And that’s fine, but what we’re doing, I think, is potentially shifting it to the point where those five people who actually look fantastic and know everything, may not exist anymore because they’ll go and get it from the AI. Maybe one or two of them are self-directed enough. But the point is that you’re shifting the margins and so you’re going to then just have fewer in every one of the good buckets than you would have earlier. There’s also this kind of homeostasis type problem where universities are reacting to what industry says they want and therefore we’re going to be producing fewer of these people and we’re not going to realize it until five to ten years from now, at which point a massive correction will be needed. And this is just purely in terms of workforce. I’m even eliding the bigger issues of human knowledge and what’s truly good. But I do think that by making it easier, you’re shifting the distribution and that has consequences.
Rina: Marija’s comments about Dijkstra made me think that actually the main danger is for understanding. Because when you are just asking for the Dijkstra algorithm, you would hardly have the same understanding of the algorithm as when you learn it the way it’s traditionally done, pseudocodes and so on. And it reminds me of what my students always tell me, that it’s not enough for them to see the pseudocode, they need to program in order to feel that they understand what they’re doing. So this understanding may be really damaged. And if we don’t have understanding, we don’t have thinking. So that’s some of the danger.
Marija: I fully agree with that, but I think the issue is already here because we already don’t have a good way of examining people for understanding. We already struggle making them do the hard thing. Pre-AI it was already difficult to push this understanding. It was already hard to make people see the value of struggling and figuring things out and checking their own work. And that was without the shortcut that is AI. So I think the problem is not new, but it is pressed to its breaking point in a way, which might be good. I mean, you walk into a room with 300 plus people, there will be cutting corners and figuring things out and copying and all that stuff. But maybe now we’re saying “okay, we understand you can do it with AI, you have the perfect corner cutting instrument here, that’s not the point. The point is to do it this way, not to get it done.” It used to be that the only way you could get a result from a computing program was to actually write it yourself and get everything to work. Then in an exam it would be enough to click run and see if it runs or not. And now we can’t do that, but we shouldn’t have been doing it in the first place. But somehow I think the problem is communicational, like trying to explain to people that the value is in walking the path, not in getting somewhere. I think students will understand this if we are more clear that the whole point is not that you make me a program so when I click run, it runs and it delivers a number. That’s not the point. I don’t care if the number works or not. The point is that you understand. And if you cut corners and don’t understand, then you cheat yourself out of it in a way.
