Balakrishnan Prabhakaran, director of UAlbany’s AI Plus Institute, leads efforts to use AI to solve real-world problems. In this episode, he explains the flights of impulse, wonder and exploration that make humans different from AI, whether we hold machines to a higher standard than ourselves, and why he wouldn't go on a podcast recorded solely from the host and guest’s unspoken thoughts (even though it's technically possible).
In part of our interview that didn’t make the final edit, we spoke to Prabha in greater depth about a word that comes up a lot when talking about AI: reasoning.
It’s a word with very human connotations. We talk about behaving reasonably and reaching reasonable conclusions. But what does it mean when we’re talking about machines?
His answer was closely related to his argument about AI’s plurality. And if it sounds familiar, it may be because it echoes a point that UAlbany atmospheric scientist Kara Sulia made last fall when we discussed AI’s ability to transform how we understand the weather. AI’s great power, she said, stems not just from its ability to look deeply inside datasets but across them to help us make sense of a wider, more complex world..
Here’s what Prabha had to say:
JCE: The word reasoning comes up a lot. It’s a term that most people, until lately as AI has entered the popular conversation, consider a very human thing. I’m a reasonable person. I look at the information before me and I come to a reasonable conclusion. But what does it mean in the context of AI? What is the machine actually doing?
BP: I want to flip it around and ask you why you said, “I'm a reasonable person.” Why are you not saying I'm a good person or a bad person? Why are you saying I'm a reasonable person?
JCE: Good and bad seems more subjective to me. When I say I'm a reasonable person, I say I've looked at the information before me and I've come to a conclusion that is defensible based on that information. Is it the right conclusion? It might be. Is it the wrong conclusion? It might be. But is it a reasonable conclusion? That’s a different question to me — as a word person.
BP: So that is the gray area. You are unsure whether that's a correct conclusion or a wrong conclusion, but based on the data, based on the information that you have, based on the prior experience you have, you think you have arrived at a reasonable conclusion. That is something that the computers were not doing before AI because it was binary. It was one or zero, it was yes or no, it was true or false. Whereas now, when you have millions of records, now you reason through it and come up with a conclusion just like a human conclusion. And there is a probability associated with it; there is a confidence level based on this data, based on the amount of training.
When I say the algorithm, the machine learning algorithm, has 90 percent confidence that this is the correct interpretation of the data, it is actually mimicking the real world, the way we look at the real world — the way we try to analyze and understand the world and derive conclusions, derive interpretations or make decisions.
JCE: The computers we have now and the computers that we're building now for the future can be trained on, and inference based on, billions of pieces of data —orders of magnitude that are difficult for regular people with non-mathematics degrees to really understand. Is that where AI derives its power from? The fact that it can reason based on enormous datasets?
BP: Yes and no. You're talking about data as one thing. Because data is like numbers, like database records of citizens and students and whatever. But you also have images, you have video, you have audio, you have gesture and movements, you have the sensor data and other things. So again, data is not one thing. Now the power of AI — yes, it comes from data, but it also comes from the ability to integrate these multiple sources of data and then reason through that information.
To read more of Prabha’s thinking on the plurality of AI, check out his essay on the subject in the first issue of the AI Plus Institute’s Hallucinations magazine.
Lean more about the AI Plus Institute at UAlbany.
UAlbany biologist Morgan Sammons also touched on the curiosity that distinguishes humans from machines in our conversation last fall about the frontiers of brain-inspired computing.
Since then, UAlbany has hired four new faculty members in neuroscience, computer hardware engineering and mathematics who will focus on the intersection of neuromorphic and quantum computing.
Audio editing and production by Scott Freedman
Photos by Patrick Dodson
Hosted and written by Jordan Carleo-Evangelist
[0:01] Host: Welcome to The Short Version, the UAlbany podcast that tackles big ideas, big questions and big news in less time than it takes to cross the Academic Podium. I'm Jordan Carleo-Evangelist in UAlbany's Office of Communications and Marketing.
[0:17] Balakrishnan Prabhakaran: We allow humans to make mistakes because we make mistakes/ Whereas we don't want machines to make mistakes. So it all depends on the societal way of looking at it and what is acceptable and what is not — and our increasing expectation of machines and the way that the machines perform.
[0:39] Host: Imagine you're holding a Rubik's Cube. That tinge of anxiety — or maybe for you it's a thrill — sparked by those mixed up colored squares? AI doesn't feel that. AI looks at that cube, analyzes the color pattern and instantly determines the most economical combination of moves to solve it.
Hammer, meet nail. No furrowed brow, no unnecessary rotations, no wonder at the game's elegant simplicity.
And that raises an important question: Can a machine ever delight in the excitement of an unknown world? Can it ever look at a puzzle as anything more than a knot to be untied? That's a singular difference between human and artificial intelligence right now, according to Balakrishnan Prabhakaran, the director of UAlbany's AI Plus Institute. The Institute brings together scholars and researchers from across campus, including the social sciences and humanities, to tackle questions that AI might be able to help solve.
Prabha, as he's known on campus, is a computer scientist who uses AI to analyze data from wearable sensors. His research on how the human body moves helps people adapt to disabilities and recover from injuries. AI is great for that.
Yet Prabha wouldn't be doing that work if not for his own improbable path from aspiring cricket star in India to one of UAlbany's leading AI thinkers. His own life's journey — shaped by pivots toward the unknown rather than the most statistically likely outcomes — could be a metaphor for what makes humans different.
Prabha likens most humans’ experience of AI to the parable of the blind man and the elephant. That is, our understanding of a thing is shaped by the narrow piece of it that we can touch in any given moment. For most of us, that's large language models and AI chatbots. That's a problem, Prabha says, because it fails to convey the true diversity of technologies and applications that AI makes possible. This fundamental plurality of AI is what makes it so powerful and enormously challenging for policymakers and the public to grasp.
In this week's episode, we talked to Prabha about what distinguishes human from artificial intelligence; why he thinks we'll keep moving the goalposts for whether AI is ever good enough to surpass us; and how close we really are to being able to talk with just our thoughts.
It's the first of two episodes exploring the relationship between AI and us.
Here's our conversation.
[3:08] Host: You have been studying artificial intelligence for much longer than most of us have been thinking about it. Most of us really didn't give it a whole lot of thought until around 2022 when ChatGPT came out. What was your path into that field like? How was AI a field that you became interested in before most people were thinking about it?
[3:29] BP: So as a computer scientist, I wanted to work with something that helps people. And I got involved in physical medicine and rehabilitation for multiple reasons and started working with a collaborator who was also interested in using technology for physical medicine and rehabilitation. It's not just about rehabilitation. It's about identifying what is normal, what is abnormal in terms of the way you walk, the way you move your body parts. And for that, people wanted to understand what is normal, what is abnormal and how you can use sensors to distinguish and help people who need that kind of help. So when you start using sensors, you start collecting a lot of data and you need to analyze data. And at that point in time, we started using what is called machine learning. And that's how I started. So using machine learning to help with human gait analysis and how that kind of gait analysis can inform physical medicine and rehabilitation.
[4:34] Host: If you're at a social engagement, surrounded by people who don't work at the AI Plus Institute like you, who are not computer scientists, who are not engineers, and they say, “I don't really understand AI. I read about it, I use it. I have Copilot on my computer, but what is it?” How do you define it for them?
[4:58] BP: I would say, hey, just imagine closing your eyes and talking to me. And now you are listening to me, but you don't know whether I am talking or there is a machine talking to you. And the machine is able to talk exactly like Prabha in the tone, the content, the way it understands you, the way the responses are given. Now, when you're closing your eyes, you may not be able to distinguish whether it is the machine talking to you or the human Prabha that is talking to you. And that's where AI is. The machine can do it now. It can mimic my voice. It can mimic the way I understand the question. It can mimic the way I respond to that question. It can mimic the way Prabha would behave in that social context.
[5:53] Host: And the fact that it's mimicking, though, is that what makes it artificial intelligence?
[5:57] BP: Yes. Because it is trying to learn and reproduce or be the way Prabha would be rather than becoming a person that would be different. Whereas as a human, I learn from you and then I become different. But as it would try to learn from me and be more like Prabha.
[6:20] Host: So I guess a related question then: What is regular intelligence or traditional intelligence? What distinguishes it from what we consider artificial intelligence as you just sort of described it?
[6:32] BP: If you're just looking at it from an intelligence perspective, can we do it? That is a question that a computer can ask and answer. But should we do it? Should I be doing this? Is it correct? As you said, is it wrong? Is it reasonable? That is a difficult decision for humans, but that is even more a difficult decision for a machine to do. Because we are talking about: I feel like doing it. I think this can be done, but I should not be doing it because of whatever reason. And that reason could change, evolve over a period of time. Whereas the machine is going to make a decision based on whatever the data that has been fed into it, that it has been trained on, and whatever it has been observing. It thinks, OK, 90 percent probability this is the decision, and then it acts on it. You don't have that kind of monitoring, a self-monitoring, a self-doubt that humans have. The critical thinking that the humans have to say that this is probably not the right way to do it. Let's stop doing it.
[7:42] Host: One way to interpret the difference between what we might consider traditional human intelligence and artificial intelligence are these factors — should we do it? — that are subjective? That one of the key distinguishers here are these sort of subjective and evolving moral philosophical answers is really, really interesting. And I guess the related question, the next question, is: is there a point in the near, medium or distant future where machines or artificial intelligence will be able to do that part?
[8:21] BP: I would rather flip it around and put it in a different way. What would we accept as a computer being intelligent? What would our definition be? Because again, that is kind of changing. We were happy with Google Maps giving us directions, but now we are expecting the self-driving cars to take us to that place. But we are holding — if you really look at it carefully — we are holding those self-driving cars to a higher standard. What is considered as human intelligence and what is considered as a machine achieving that human intelligence today in 2026 may be very different like five years later or even three years or 10 years later. Our expectation may be different because the machines have achieved a certain level and now you are not happy with whatever that level is. And then we say, “Hey,” you find holes in that wherever that machine is and say, "Hey, these holes are there, so it's still not intelligent enough."
[9:28] Host: For folks like me who aren't immersed in this every day, I think most of us, our interaction with AI is primarily through LLMs. And so this notion that AI is many things, I want to ask you about that. And I also want to share the way you described it in something you wrote recently, because I thought it was a great entry point into this idea. You said, "Much like the parable of the blind man and the elephant, our understandings are shaped by partial encounters that grasp only fragments of a much larger reality. AI is not a single technology, but rather an ecosystem of interconnected technological advancements across multiple areas." So what do you mean by that?
[10:07] BP: If you look at the evolution of the AI machine learning algorithm, they have been trained or they have been targeted for certain tasks. And when you do that, the algorithms become very good for certain tasks, not so good for other tasks. If you look at the way the LLMs, the large language models, developed, right? So the backbone or the central one, it's the neural network. They were actually designed for understanding images, understanding whether there is a cat or a dog in the image. So they were looking at pixels, image pixels, and then trying to group the pixels to understand where an object might be and what that object could be. As the expectation of people with respect to AI changed and the needs changed, people figured out like, "Hey, this is a powerful neural network architecture that tries to understand, that tries to mimic the human brain.” That's why it's neural network architecture. Now I can apply it on text.
Whereas the image pixels did not have any meaning. It's a value with color and brightness whereas in text, alphabets and words have meaning. So now you are using something that was used to interpret kind of like information without meaning and then make sense out of it. Now you are applying it on data like text that has this information that has meaning, but the algorithm thinks that that has no meaning and applies statistical techniques to interpret the meaning. And that's when it can go right or it can go wrong because you are not really understanding the semantics of it, but you are using the same statistical principle to analyze that text information, come up with your interpretation, generate information for you and so on. Whereas if you look at the traditional neural network architecture, it was for discriminatory purposes. Is there a cat? Is there a dog? When you apply this to medical AI, for example, you can detect cancer. Is that cancer or not cancer? In many cases, the doctors have even said that, "Hey, I would not have thought about that. " And the AI machine learning algorithms caught that point, and so the patient's life is saved. So there are a lot of good things that can happen whereas we are fixated on one type of AI.
People always talk about the inaccuracies and the problems that can happen with the AI. And in my view, that is a little bit of a dangerous situation because now you're not going to fund research that would enhance the technology and make it better for other domains. So we need to be careful. We need to convey to the public that there are multiple components that can be helpful for you, that can be beneficial for you. So we are responsible for the public and the public's opinion, the public trust matters. And for that, this notion of conveying this notion of plurality of AI becomes really, really important.
[13:17] Host: When you were nine, 10, 11 years old, you're growing up in India. When you were a kid, what did you think you would end up doing? And what was your path from that to here?
[13:29] BP: Oh, I would be lying if I said I could have imagined the whole thing. Obviously growing up in India, cricket was a popular game. My dream was to become a cricket player like any other kid loving sports. But then I had this notion of like, “Oh, I want to become an engineer.” And I didn't know what that means. And then I wanted to become an electronics engineer. I didn't know what that means. And then I was like, “Oh, electronics is good, but computers seem to be more fascinating.” I want to become a computer scientist. And again, I didn't know what that meant. So these were very typical human getting influenced by friends, family and media and perception, like what you would like to become. And then identify whether you like it or not. That's probably a very good way of looking at and comparing human and the artificial intelligence too.
We are able to have a sense of the world. We are able to have role models on what we would like to become. Not everybody in my family wants to become a professor. In fact, some people have asked, "Why did you become a professor?" Because they think that's a boring job. We have role models. We have things that excite us. But as when you talk about artificial intelligence, you take data and then you want to make a sense of that data and then behave as if the world is the same. Whereas for humans, it's like we are very much ready and actually looking forward for that excitement of an unknown world. And you start exploring things. The way you take inspiration, the way you take the experiences of others in shaping your life, the role models and the excitement probably is what led me to wherever I am.
[15:25] Host: When you first came to Albany a few years ago and I asked you what were some of the things that excited you most about where this field is going, where AI was going. One of the things you said is, "Who knows? A couple years from now, maybe we do this interview without talking to each other. I just think the questions, you just think the answers, and that's how it works." How real is the notion of me and you sitting here interviewing each other just based on what we're thinking without speaking any words?
[15:52] BP: To answer your question about communicating with thoughts: we comprehend, and then at some point we have an urge to ask a question — like you want to ask a question, that's an urge. But you don't know what that question is. So you want to fixate on human intelligence versus natural intelligence or ethics versus this. So now you have that urge. Now you are conceptualizing that question. And now even at that stage, it is just a concept that you want to ask about this kind of a topic because you find a conflict in your understanding. There is a conflict and you want to ask a clarification. It's still a concept. Then you are formulating a question, the words that you want to use and how you want to formulate your question. That's a formalization phase. Then there is a mechanist phase where you are asking that question.
The tongue moves, the lip moves, the larynx moves, and you produce that sound. Now AI with the help of implantable devices to understand what is happening in the brain, pick up the brain activity, they have been able to do it in both the second and the third phase where you are formalizing the question, you are building that question or the sentence as well as when you are speaking it out, the mechanist part of it. But for a person without any physical disability of speech, the AI working on the mechanist aspect of it is not that interesting. So the second part of it, the formalization part of it, people also call it the inner thought. And they're picking up on the inner thought to understand the question you want to ask and then frame it as a question. AI has been successful in doing it, but that is a challenge.
Probably you are thinking that Prabha is speaking nonsense. I don't understand what he's talking about. But you are not going to say it hopefully, at least hopefully not during when the recording is in progress. What I'm trying to say is that it's a self-monitoring process in humans. Even if that comes as a slip of a tongue, then you try to change it, you apologize. Now that kind of thing is difficult for AI.
[18:06] Host: Not every inner though is meant to be spoken. I know the difference, but does the implantable device know? It's
[18:13] BP: In AI, it is called the world model. So you have a world model where you know this is appropriate, this is inappropriate, and you can change the way you articulate your question. Whereas the current AI is picking the brain signals, mapping it into the question, but there is no self-monitoring, there's no ability to control that process. All of us have our own view of the world, which provides a context. What is appropriate, what is not appropriate? Whereas the AI would be looking at a generalized world model which may not be appropriate. The technology is already there in terms of picking up the brain activity and communicating our inner thoughts, framing it into a question or text and communicating it. Whereas it may not be helpful for a healthy individual because you don't have a control process, it may be very useful for a person challenged with speech production due to whatever deficit they have, either because of a stroke or some innate disability.
Now it can help people communicate and work with the world around them. So if you are looking at a use case where you are trying to handle a deficit or overcome a deficit, yes, we are already having the technology and the technology may be useful. But if you're looking at it like can I use it as a general purpose communication tool so that I can be faster, better and more efficient? If you are able to handle the fallout that can happen when people know you are inner thoughts, then yes.
[19:53] Host: We're not ready for podcasts powered entirely by inner thoughts.
[19:59] BP: I would not have accepted it if you invited me for such a podcast.
[20:04] Host: That was Balakrishnan Prabhakaran, director of UAlbany's AI Plus Institute.
To read more about what Prabha describes as AI's fundamental plurality, and what we mean when we say that AI engages in reasoning to answer the questions we ask it, be sure to check out The Longer Version in our show notes.
And don't miss next week's episode in which my colleague Bethany Bump quizzes Assistant Professor of Philosophy Alessandra Buccella on what it would mean for AI to become conscious.
[20:32] Alessandra Buccella: “What it's like to be a creature with a mind, what it's like to have a unique perspective on the world and your own specific point of view and the qualitative aspect of what it is to have that point of view — that's how philosophers think about consciousness.
[20:50] Host: The Short Version would not be possible without contributions from many people, including audio production and editing thanks to the very human ears of Scott Freedman in UAlbany's Digital Media Studio, deep inside the Podium tunnels.
We'll be back next week with another quick conversation about something interesting. I'm Jordan Carleo Evangelist here at the University at Albany, and this has been The Short Version.