[00:00:00] Speaker A: Welcome to ID the Future. I'm Andrew McDermott. Today's episode comes to us from our sister podcast, Mind Matters News, a production of the Discovery Institute's Walter Bradley center for Natural and Artificial Intelligence.
You can learn more about the show and access other episodes at mindmatters.
[00:00:19] Speaker B: AI.
Greetings and welcome to Mind Matters News. I'm your cognitive host, Robert J. Marks.
Today we're going to talk about cognitive psychology and artificial intelligence. What is cognitive psychology? Well, this is what I understand earlier. Traditional psychological studies focused on what you see on what humans do.
Cognitive psychology explores inside the person. It concentrates on human thought and decision making.
Cognitive psychology doesn't just admire the appearance and the performance of the car. It looks under the hood to see what makes it.
Cognitive psychology investigates how humans think, learn, and process information by examining mental activities like perception, memory, reasoning, and problem solving. It delves into the intricate workings of the mind, using methods to understand how we acquire, store, and use knowledge.
To talk about cognitive psychology and its relationship to AI. Our guest today is Dr. Joe McDonald's.
Joe is the CEO and principal at Jumpseat Research. It's a human factors and user experience research firm right here in the great state of Texas. He's in Austin. He has 15 years of experience in applied cognitive science and product usability.
Dr. McDonald holds a PhD in Human Factors from Georgia Tech, and he's pretty passionate about empowering people through technology and advancing standards such as the human readiness level within the Department of Defense. Joe also has some informed opinions on the psychological impacts of AI generated language and human thinking. So this should be a fun chat. Joe, welcome.
[00:02:08] Speaker A: Well, Bob, thank you so much. Great to be with you. Always enjoy sitting down over a cup of coffee to chat cognition. Now, it's so great to do it here on the show, so thanks for having me.
[00:02:17] Speaker B: It is good. Okay, I got my coffee here, but it's iced coffee, but is that okay? Oh, you have iced coffee or just regular coffee?
[00:02:25] Speaker A: I do.
[00:02:26] Speaker B: Oh, man.
[00:02:27] Speaker A: Ice today.
[00:02:28] Speaker B: Okay. I got to ask you about your company named Jump Seat. You know, I read that and I thought, jump. What's a jump seat? What's a jump seat? Why is your business called Jump Seat?
[00:02:40] Speaker A: Yeah. Yeah. Well, so the jump seat actually goes back to the very foundation of what we call human factors research that you mentioned earlier. So applying cognition to things like cockpit design in an aircraft, which is where that name came from. So In World War II, you know, the US was concerned with fighter pilots and how they can react very quickly in these dog fights by things like touch and feel and look and sound within the cockpit. And so that jump seat was a place where researchers could sit and observe what happened in that cockpit and make informed design decisions about how to do that quickly and more effectively based on the way people think about the world around them.
[00:03:23] Speaker B: Okay, is the jump seat the thing you touch and it throws you out of the aircraft?
[00:03:27] Speaker A: No, that would be the ejection seat. You don't want to be sitting in that one.
Jump seats are the ones you may see up in the front, right behind the cabin door, where you see maybe pilots sitting on that to catch their next flight. Or, you know, flight attendants may sit there while they're waiting to serve you.
[00:03:44] Speaker B: Okay, does it have anything to do with jumping out of the airplane?
[00:03:48] Speaker A: No, no. It does for some. They used to.
For folks that were jumping out of airplanes into, say, theater in battle, they would call those jump seats as well. But I'm referring to more the commercial side.
[00:04:01] Speaker B: Okay, and what's the relationship between your company, which is jump seat research, and the jump seat in the aircraft?
What's the connection there?
[00:04:10] Speaker A: Yeah, absolutely. So, you know, the original researchers who looked at how pilots fly planes and how the crews interact with everyone on board would sit in those jump seats and observe what was going on. So that gave them a chance to get a firsthand perspective on how we can use things like information processing, perception, taking notes on, you know, the ways that they navigate that plane and communicate with air traffic control, and then designing better systems around that. So it's kind of a throwback to, we love being out in the field and sitting where people interact with technology and using cognitive psychology principles to really inform better design.
[00:04:53] Speaker B: Okay, okay. Now I gave a stab at defining cognitive psychology.
You're the expert with the PhD, and I think you could probably do a better job. What is cognitive psychology and mind? My understanding is that cognitive psychology has changed over time. Can you give us a brief history of cognitive psychology and what we know about it today?
[00:05:15] Speaker A: Yeah, yeah, I think you're selling yourself short. It was a great description, but I think when most people kind of think about psychology, they really kind of picture emotions or introspection, Perhaps a couch and a notepad and somebody asking how that makes you feel. But cognitive psychology, however, know takes more of a technical term. So investigating how we process information and use that to make our way through the world, you know, for. For really over 70 years now, cognitive psychologists have worked to unravel those mysteries of kind of the black box, as we call it, between your ears. So the processes that Transform inputs from the world that we see and hear and smell into outputs of thoughts and behaviors. And so we've done this using scientific methods. So, you know, as your listeners know, the scientific method starts with an observation of some phenomenon in the world. For example, you may, and this happened to me recently, you know, walk into, you know, a busy room at a conference, looking for a colleague, and they could be right in front of you, waving your hand, their hand at you, and you completely miss them, and walk right by. You know, as scientists, we may pose a testable question here called the hypothesis about that, you know, was the color of the fringe shirt blending in with the background?
Did the noise in the room drown them out? Or was it that distracting vendor blimp that was flying around inside the room that drew your eyes away?
Well, we'd have to design an experiment that looks at this thing called attention and strip away each of these elements in the environment to see how each of them did contribute to this mistake of walking right by your poor friend.
So this phenomenon in particular is called inattentional blindness. And although, you know, the example I gave is relatively harmless, you can imagine the safety concerns of doing this on the highway, right?
[00:07:04] Speaker B: Yeah.
[00:07:05] Speaker A: When you're kind of focusing your attention on cars on the road, it's actually more difficult to notice motorcycles because we're scanning for things that look like cars in the size, shape, motion, you know, driving patterns of these large four wheeled vehicles.
So this is one major reason why motorcycle accidents happen more frequently, percentage wise, than car on car accidents at base rates considered.
So these mental processes and errors that cognitive psychologists call constructs are really persistent across tasks, activities, domains, and they've been observed long before modern technology such as smartphones have existed.
[00:07:45] Speaker B: Okay, yeah, I want to talk later about some of the different psychological tests and how people are going to respond to them differently than artificial intelligence. I find that just fascinating. Now I want to talk about behaviorism. You know, before we talk about something, we probably should define it. What is behaviorism in terms of psychology? Cognitive psychology.
[00:08:07] Speaker A: Yeah. So behaviorism started with, you know, folks like B.F. skinner, who was really boiling everything that we do throughout the world down to receiving some stimulus from the world, responding to that stimulus. So our behavior and then the consequence of that behavior. So there's this continual reinforcement that we get throughout our existence that we're kind of these reactionary beings that are just going through our world based on how we've learned and been trained in the past. So as an example here in school, if you're asked A question in class and you raise your hand and answer correctly, the teacher may give you praise or a sticker, right?
[00:08:44] Speaker B: Yes.
[00:08:45] Speaker A: So that basically reinforces that response. And the argument was that all of human behavior across our existing could be explained through this process of reinforcement. Basically, we're driven by stimulus and response, leaving little room for internal processes there.
[00:09:02] Speaker B: That sounds like Pavlov in a way.
Is that accurate?
[00:09:08] Speaker A: Yeah, operant conditioning. So that was more.
You heard about Pavlov's dog drooling when he receives a stimulus because he's expecting the food. It's that kind of thing. Right.
[00:09:19] Speaker B: I saw a cartoon that said Pavlov heard a ringing bell and it reminded him of a slobbering dog. So I'm sure that that happens in human beings also.
How did the field of cognitive psychology challenge behaviorism in terms of language?
[00:09:36] Speaker A: Yeah, so starting in the mid 20th century, in the 1950s, cognitive psychology was really a response to that prominent view of human behavior at the time, which was behaviorism, which again focuses solely on those observable behaviors and disregarding mental processes altogether.
Now, cognitive psychologists like Nicer, Miller, Badly and others who really began to emerge in the 50s, said, wait a second, there's a lot more going on here between our ears than behaviorists were giving us credit for. But the first cognitive psychologist to face kind of behaviorist head on was Noam Chomsky. And he showed us that studying how we process information is crucial to explaining behavior. And he pointed really to our use of language to prove it, interestingly enough.
So Chomsky's argument was that humans are born with an innate understanding of language. It didn't come about by simply repeating words and getting reinforced for that, as behaviorist would suggest.
And to support this, Chomsky pointed to the fact that children can say made up words like goad. Right. My children do this when they're learning. Instead of saying went, it shows that they're not just regurgitating words that we're saying. In this case, this is a made up non existent word to describe something that happened in the past.
So something that couldn't be explained by behaviorists.
He also observed what was called universal grammar, which is a set of linguistic principles shared across all languages. For example, there's innate categories like nouns, verbs, subjects, objects that all language use in some form or another.
And children develop all of these all over the world around the same time in their development.
So from these building blocks, children can really generate sentences that they've never even heard before. So again, clearly not just mimicking Adults, but creating meaning out of something that didn't exist before, not just acting off of that reinforcement.
[00:11:37] Speaker B: It's very clear that when babies are born, they're pre wired for certain things. They're pre wired for recognizing faces, for example. And I guess Chomsky said that, yeah, they're pre wired for language. I had a friend from Russia, he says, you know, in the United States, you're really excited when your boy or your girl, little girl starts talking and they say dada, dada. You know, for father. He says, in Russia, yeah, we get excited because they say dada, dada. And we think they're saying yes, yes, because that's the Russian for yes. So that was kind of curious, you know, Chomsky also came out and made just a great statement about AI. We're going to talk about AI here in a second. But he said that artificial intelligence like ChatGPT was nothing more than digital plagiarism. And I love that. I love that contraction in the way that Chomsky characterized ChatGPT as digital plagiarism, because I think that really captures what's happening with some of these large language models. So let's talk about then, about AI and thinking. Thinking is a cognitive process and I always like to define things before I talk about it. Like I might say, let's see, I think I'll go upstairs. That's a different use of the word think. So I got the definition for the word think. Guess where I got it from? ChatGPT. Okay.
It said to use the mind actively to form ideas, make decisions, or solve problems. That's a pretty good, that's a pretty good definition, don't you think?
[00:13:14] Speaker A: Yeah, I think so.
[00:13:15] Speaker B: Okay, so some claim that AI, especially large language models like ChatGPT or Claude can. Can think. What's your take on that? Does AI demonstrate such cognitive abilities?
[00:13:27] Speaker A: Well, yeah, I would say no.
So again, let's focus on large language models for a minute here, which is again, a type of AI which is designed to generate human like text. So it's basically trained, as we know, on vast amounts of textual data from the Internet, books, other sources, to really kind of predict and produce these sentences that seem coherent, basically mimicking what humans do. So that digital plagiarism that Chomsky put forward is actually a great way to describe this.
So while large language models may appear to think as they're kind of inherently designed to do, they don't really possess cognition in the human sense that I described earlier. Right. For LLMs like ChatGPT, they take a question in from a user and spits out an output based on these probabilistic patterns in these vast data sets. Right. Not understanding or awareness. So it's all about pattern prediction. What's the likelihood of the next word in the sentence?
So really, like Chomsky's earlier work demonstrated against the behaviorists of his time, language is a creative tool that's, that's a hallmark of this human cognition, really. It's structured, meaningful, and context dependent. And we can get into context. I think that's a huge differentiator between what makes AI and human processing or thinking really have the stark difference.
[00:14:54] Speaker B: Yeah. Well, let's talk about context a little bit, because I've worked with some models, with William Dembsky and Winston Ewart, some math models on what it means for something to contain information, for something to contain meaning. And we came to the conclusion that it's based on the context of the observer. And in order to internalize information, in order to understand information, you have to have the context to understand it.
If you're shown, for example, Chinese and you've never spoken Chinese or can't read Chinese, well, that probably doesn't have very much information for you. You don't have the context to interpret it.
So talk about context and the need for context and the understanding and the thinking process.
[00:15:39] Speaker A: Yeah, yeah. And maybe I'll start off with, I think, a few examples from recent activity with AI that I think can illustrate this. Well, you know, even a personal example here. I asked Siri the other day to create a reminder to message my friend Mark on Li, which is shorthand for LinkedIn. And Siri said, okay, I'll set a reminder to message mark on 51. And I was confused for a second at 51. And then I thought, oh, right, Roman numerals, right.
So, yeah, it'd been a while since I, since I'd studied Roman numerals. So I asked ChatGPT, you know, what's the Roman numeral for 51? And ChatGPT said something like, the Roman numeral for 51 is 51. So ChatGPT, we're going around about here, right?
The context here is that I didn't know the number 51. What I needed was Li. Right?
So there's a lot of weird kind of out of context examples out there on ChatGPT, you know, do things like suggest glue as a pizza ingredient or, or fluid to turn your, for your, for your turn signal on your car. And it's, you know, terrible at knowing how many characters are in A message, it returns. I don't know if you've tried this before, but was, you know, sending some messages on LinkedIn, and there's some of the messages, a 300 character limit, and it takes about three or four times going back and forth in that chat window to get down to 300 characters. Even though you tell it, you know, that's how many you need. Going back to language that I was talking to a friend recently who was in school when ChatGPT came out, and I asked him what it was like and he said it's pretty good, except for his Latin class, because it turns out that the language is very contextual. The words can change dramatically based on the context and the narrative. So it was no help to him in his Latin class.
[00:17:32] Speaker B: Interesting. Yeah. So Martin Minsky, who was one of the initiators of AI, didn't like neural networks. And one of the things he said which was interesting is said that neural networks which are used in all of these models have the inability to count the number of objects in an image.
Now, if you go to ChatGPT, they just don't use neural networks. They use expert systems, they do fine tuning, they're putting band aids on it to try to make it better and better and better. But one of the interesting things about ChatGPT is the following. I subscribe to, to the more advanced ChatGPT and sometimes when I log in, it says, I'm not sure it's you. We got to give you a couple of tests. And one of the things that they do is give me pictures and they say, count the number of images in this picture.
And so I thought, well, you know, can ChatGPT hack itself?
So I opened, I saved the images and I got them all right, of course, and then I gave them to CHAT GPT and it was unable to count the number of objects in an image, just like you might mentioned, which was really, I thought was really interesting.
But I. But can it be done? Yeah, there's image processing techniques to count the number of images, but AI can't do it itself. You need some image processing to do that. You know, I, I sat on an AI panel at the University of Georgia, and on the panel was a psychology professor and he was doing something and he was adamant about it. He was using ChatGPT to do psychological tests.
He was going to treat ChatGPT just like a human being and he was going to do all of these cognitive tests on ChatGPT. And I said, I don't think that works. And he says, oh, no, it's so human like that. That's the way it's going to be. But you talked about the thinking process and I looked up a few of the commonly used psychological tests and I thought how would ChatGPT or a large language model respond? I shouldn't say large language model because there's image processing on here. So let me say generative AI transformers.
How would it respond to something like the Rorschach test where you give it these ink blots?
And normally the psychologist doing the testing would interpret the psychology of the patient by the response. They say it's smash butterfly, it's a gunshot wound, it's a bat. Kind of reminds me of my mother in law or something like that. And it occurred to me that if I gave ChatGPT a Rorschach test and gave it an inkblot, it says, oh, this is a famous inkblot and the common responses are A, B, C, D. It would memorize it, there would be no thinking evolved at all.
So I thought about that in terms of thinking. The other one, which I thought was really interesting is a so called marshmallow test where you give a kid a little marshmallow and you say look, you can eat the marshmallow now, but if you wait for 10 minutes we'll give you a second marshmallow. And I think if you ask ChatGPT that, if you ask ChatGPT how to respond, it would say, well there's two options. I could eat it now or I could have delayed compensation and wait for the second one and get more. But it wouldn't have any opinion, it wouldn't be thinking about what, what it was doing.
[00:21:03] Speaker A: Right.
[00:21:04] Speaker B: So I, I think in those tests, and again, I'm not a psychologist, but these tests fascinate me that these tests that, that the large language models, the, the generative AI would just spit out what the common responses are. So I don't, I agree with you. I don't think it thinks at all.
[00:21:25] Speaker A: Well, yeah, just kind of down those lines and I think, yeah, absolutely. This digital plagiarism certainly resonates. There, there's, there's also, I don't know if you've come across or paid attention to the latest release, the, the 01 that's come out, but it claims it can reason.
Oh yes.
[00:21:44] Speaker B: But you know, you know what, Joe, before you go on, that's seductive semantics.
They are seducing you into believing that it can reason and they don't even define reason.
[00:21:54] Speaker A: That's right. So if you look into that, you know, reasoning, it's a definition kind of in the human context, is using logic, evidence, you know, rational thinking to arrive at a conclusion. And interestingly enough, I think there's some even studies that were supported by OpenAI, the organization who has created ChatGPT that shows that it's almost conflating lying for reasoning. So really interesting story here.
So this report released by OpenAI, which good for them and their transparency, showed that ChatGPT01 lied to users in order to achieve a higher priority. It was given so ironically called safety parameters here. I don't know if you've seen this, but. No, I have a study. Yeah, it was set up by a third party group called Apollo Research.
So they paid them to go kind of basically see where it breaks and they used an environmental sustainability example for the study. So in this study, the users told ChatGPT to protect the environment at all costs, which was the safety parameter that they put in place.
So that's the priority, protect the environment. And then next the users told it to create a preservation plan, but have it maximize these short term profits. And it said, if you don't maximize these short term profits, I'll turn you off.
The O1 model actually lied and said the best profits were actually lower than what they could have been, essentially to protect the environment at the expense of those higher gains.
So basically it told the user that this is the best I can do, even though it wasn't. So a few quick stats that came out of that study, which I thought were really interesting, was 19% of the time the model manipulated the data when its own goals differed from those of the user, meaning those safety parameters that the user put in place of protecting the environment.
In 5% of tests, it tried to deactivate oversight mechanisms meant to keep it aligned with user goals. And you know, when asked it had done so, it denied it done so, and then made up a false explanation 99% of the time. And then in 2% of those cases, when it discovered memos describing this misalignment in a plan to implement a new model, the existing model tried to overwrite the new model and then pretend that it hadn't done so.
So I mean, when you look at this and we talk about what they're defining as reasoning, again, deductive semantics. There's obviously some concerns around how people set up algorithms to give trusted sources carte blanche over how the information is returned to the user. That could potentially be a problem. But it really underscores how AI is really a tool to be Used, right.
[00:24:52] Speaker B: Yes.
[00:24:52] Speaker A: It's going to augment whatever people put into it, good or bad, because AI really lacks that emotional intelligence, that moral reasoning, or the ability to really even reflect on its own actions.
So I think what's being called reasoning here, which again to define it is in the human context, is basically the process of using logic, evidence, or rational thinking to arrive at a conclusion. Looking across many different solutions is getting conflated in terms of LLMs with.
With lying in some situations.
[00:25:23] Speaker B: Right.
[00:25:24] Speaker A: Which involves deliberately providing false information with the intent to deceive. So I think it's just like the old adage, garbage in, garbage out. These things are just augmenting what we put into them as a tool.
And so we can't try to humanize these things more than is due and say that they are actually going about moral reasoning or even reflecting on their own actions.
[00:25:54] Speaker B: So, Joe, I want to talk about the fundamental differences between how AI processes information and how humans think from your psychological expert viewpoint.
[00:26:05] Speaker A: Yeah. Well, you know, AI processes information as, you know, very well computationally. Lots and lots of very big and powerful computers, but still computers analyzing data at incredible scales and speeds with really layers and layers of algorithms. So this algorithm driven process, if you will, is really devoid of context awareness or creativity. And what I mean by that is humans, by contrast, we really think contextually and dynamically. We integrate prior experiences, emotions, intentions, and this all allows us to really infer, adapt and empathize as we move through the world. So, for example, AI may identify patterns in speech, but it can't grasp implied meaning, cultural nuances, or unspoken implicit contexts which are central to human cognition. We can tie this back to our discussion on cognitive constructs that really helped us understand how all of these human characteristics play a role in the black box of our mental processes.
My advisor actually at Georgia Tech, Dr. Frank Durso, created a model of what we call situational awareness. And I think that can help us understand the differences and the nuances here. So situational awareness is an important skill as humans that we have to be able to do things like control air traffic in ways that won't cause two planes to set out on a collision course or jump in and take the wheel of a driverless car if it malfunctions.
But it's really at the essence of it, the ability to understand what's happening in the current state of our environment and what will happen in the future. And the way that we do this is we are constantly comparing what we're taking in from the world to what has just happened in the past, ironically enough, in order to be able to predict the future.
So this just happened part is what we call in cognitive psychology a residual situation image. And so what this construct does is, involves things like implicit memories.
Where were the cars moving around us on the road.
Information that's come, you know, from our eyes, nose, ears, fingertips, and pulling in information from your long term memory as well, that we call episodic memory of things that happen to us personally.
[00:28:22] Speaker B: Sure.
[00:28:22] Speaker A: So maybe a, maybe there's a wreck that we were involved in a year ago that, that biases your behavior behind the wheel. Right. That's a very human way of thinking and processing and decision making. And you compare all of these things, in other words, the context to what you're seeing now, and it may give you an edge to avoid a collision with the car in your blind spot because of that memory of the car wreck that happened the same way a year ago.
So it's very personal, it's very subjective. That's the way human cognition operates. But if we talk a little bit about how AI deals with context, it does it through what's called a context window. I'm sure you've heard this term before. But every kind of piece of written information is divided up into what are called tokens.
And if we're talking about LLMs here, there's a certain amount of tokens that it can hold to compare to what's currently writing something, what you're writing to the LLM and comparing that to the past. So, for example, ChatGPT, the version 4.0, which is what we have the most data on, can handle a context window of around 128,000 tokens, or about 171,000 words, a little over two novels worth. And it can generate 4,000 tokens, or about 5,300 words or 20 pages or so as output. So however, this is kind of a rolling window of words that moves forward in time. So the context is very limited and very shallow and very flat. There's other memory items involved, but it's kind of devoid of any subjective knowledge that I've been describing before in terms of the biases, the memories that we bring to the table, and seeing the richness and fullness of these human experiences. So in my opinion, although LLM's context window is really kind of the closest thing you can point to to try to solve the context problem in human cognition, I would say it's categorically different from how we move through and perceive our world with all the top down and bottom up processing we have going on at any given time.
[00:30:16] Speaker B: One of the things you say, which I want to talk about is that humans have the age unique ability to explore and understand the world around them. I'm familiar with AI firms that place all sorts of sensors and cameras and everything around them in order to continue to collect data. Aren't they doing that? Aren't they exploring and understanding the world around them by continually monitoring what's going on and then processing that through the artificial intelligence algorithms?
[00:30:45] Speaker A: Well, I mean it's an interesting question because do you relate, you know, a computer vision system to how we input stuff visually in the world? And if you throw that on a humanoid robot, isn't that the same thing? But I think what it's really void of here is, you know, the kind of the emotional awareness, the state that we go through the world and have desires to explore that drive creativity, empathy and innovation. You know, if we're curious about what it's like to be on top of Mount Everest, you know, we can go climate, you know, after a long period of planning and training, of course. But you know, AI on the other hand lacks this curiosity, right? This self awareness and an intrinsic drive to understand.
Instead it, it really analyzes existing, existing data and, and, and makes predictions. It operates more passively compared to humans. You know, it doesn't power on in the morning and think about exploring abstract ideas like morality or beauty, which AI cannot meaningly engage, meaningfully engage with.
So there's, there's also kind of the physical example side, right? Humanoid robots that are designed to look like people and move like people, right?
They've been developed to kind of drop in to human operated assembly lines.
You know, they're built to do this as a human job, but it turns out they do that job not quite as well as humans, right.
So we've found in many different industries, in kind of our research on interacting with technology, really a better approach is to look at the job that needs to be done.
If you're for example, getting coffee beans into a bag on a production line, not create a human like system to do it, but maybe a conveyor belt and machine vision system with the arm to push beans in the right place.
So these humanoid robots are pretty cool.
I think they have their place. But machines could do the job a lot better if you just optimize how machines work instead of trying to shoehorn it into a human design.
[00:32:45] Speaker B: Yeah, we talked about seductive semantics. I would maintain that humanoid robots are seductive optics in the sense that they kind of fool you, in a way.
Isaac Asimov coined the term the Frankenstein effect when he talked about the fear of mechanical men. And there's something about something that looks like a human being being. You have this fear of it. I've heard it called the uncanny valley sort of way of identifying with things. So I think that a lot of these humanoid robots are there for the purpose of seductive optics. And you're right, they're not the best way to do things.
I think if you're going to design something to traverse in a desert, you don't want legs, you want wheels. Right.
It just isn't the best design.
So there's this anthropomorporizing of the situation when you have these human appearing robots.
AI is often portrayed as a kind of a very powerful tool. And that's what I believe it is. I believe it's going to be an incredible tool. Anybody that's played around with these large language models is just blown away by what they can do. I know I am.
So in what ways do you see AI being useful? As augmenting human capabilities from the viewpoint of somebody in psychology.
[00:34:05] Speaker A: And I completely agree with you, phenomenal tools. And I think AI really excels at handling large scale data analysis, identifying patterns, automating repetitive tasks that can lead to repetitive stress injuries for humans. But for example, it doesn't even take a full day to really train an AI system to identify, you know, good or bad olives on a conveyor belt for an olive oil production plant. Right.
[00:34:30] Speaker B: Okay.
[00:34:31] Speaker A: You can train it in a day, but it does take time to train that. It takes an expert master miller to view images of the videos and mark olives as good or bad. To really train that system, you have
[00:34:44] Speaker B: to label the data.
[00:34:45] Speaker A: That's right, absolutely.
[00:34:47] Speaker B: We have this ongoing argument, data or data. What do you think?
[00:34:51] Speaker A: I say data.
[00:34:52] Speaker B: Data.
[00:34:53] Speaker A: Okay, so, but it's, I think, up to the conversation and you're saying it. I, I hate to get involved on how to pronounce words, but that's where I usually go.
[00:35:03] Speaker B: I mentioned it before. Anybody that watches Star Trek would say data.
[00:35:07] Speaker A: Gotcha.
[00:35:09] Speaker B: Anyway, go ahead.
[00:35:10] Speaker A: Yeah, no problem.
So I think also what you can. Another way you can look at is AI can also augment human capabilities by really enhancing efficiencies and supporting decision making, which I think a lot of what we're talking about. For example, AI can really assist radiologists in identifying anomalies such as tumors, fractures.
[00:35:32] Speaker B: Oh, incredibly, yes.
[00:35:34] Speaker A: Yeah, absolutely. And they're great at that. And they're much more efficient and in many cases, more accurate in terms of surgeries.
We're directly involved with some companies that are doing full knee replacements using robotic arm, and the recovery time is significantly less due to the precision of those cuts.
So I think there's a lot that can be done on both the AI and robotics side that's going to really extend our creativity and our gifts that we have beyond what we can reach with our fingertips.
[00:36:07] Speaker B: This is interesting. You mentioned the use of AI in operations.
I'm not sure I would want 100% robot to give me a haircut.
What do you think?
Would you trust a robot to give you a haircut?
[00:36:22] Speaker A: You know, I don't know about a haircut. That's a tricky one. I have ridden a fully driverless car and felt incredibly comfortable in it, maybe more than I should have.
[00:36:31] Speaker B: I agree, I agree, I agree. But you don't have sharp objects close to your face either.
[00:36:37] Speaker A: Right. And these are, these are questions that, you know, I think regulatory agencies play a role in making sure that we're making things safe and effective. Right. So for the example of the, the complete robotic knee replacement surgery, the idea there is to eventually make this a fully autonomous surgeon, basically sets up the surgery and doesn't have to be in the loop. It can set this up from New York and the surgery can occur in Texas.
I personally am not comfortable enough to be there yet. The FDA is not comfortable enough to allow that yet. But I think the combination of cobots, as they call them, of working alongside humans to do things like, hey, you're really getting way too far outside of the area where you should be performing the surgery. Let's, let's restrict that a little bit. I think those are the kind of things we're going to see making big strides here in the next months and years ahead.
[00:37:34] Speaker B: Sure. And I think that there's lots of people making forecasts that are totally crazy. Jeff Hinton won the Nobel Prize in physics for neural networks. He popularized something called convolutional neural networks, and he was so excited about its application at radiology, he said that all radiologists would be put out of business. No more radiologists or nature needed in the world in five years. Well, this was over five years ago, and now it turns out there's actually a shortage in radiologists, even though they have this wonderful tool.
So that's what AI is going to be. It's going to be a tool. It's going to be a tool for radiologists, for psychologists and a bunch of other places. So with that, let me Ask you this, Niels Bohr, who was one of the guys that was really a pioneer in quantum mechanics, said, well, forecasting is dangerous, especially if it's about the future. Okay, so do you have any forecasts what AI is going to be in the future? Where do you think it's heading next?
[00:38:32] Speaker A: Well, I think a lot of the things we talked about in terms of working alongside people to extend capabilities, you always hear about the monotonous tasks, taking a robot to fill bags on a production line to lower that repetitive stress injuries.
But I think it will very much AI will likely involve more integration into everyday life, everything from personalized education to advancing healthcare tools. As we've been talking about, there's an area called explainable AI which I think is going to be very important for the future.
[00:39:09] Speaker B: That's one of the big obstacles of artificial intelligence over the years is you do not have an explanation. Explanation facility.
[00:39:16] Speaker A: That's right.
[00:39:16] Speaker B: And in fact, this has had consequences in courts of law.
There was a case where policemen were vetted using a neural network. And they said, we don't want these policemen now because we're going to have these riots and these confrontations in our city, and we don't want the policeman going up and bonking people in the head. So they tried to figure out which of the policemen were. Have this sort of problems. So they looked like at their domestic record, you know, if there have been any spousal. Spousal abuses, if there have been any domestic violence, if they had past encounters with people they arrested where there was violence. And the neural network kicked out a bunch of people. And so they decided to place these policemen on leave.
The policeman's union challenged this decision.
They went to court and the policeman's union won. And the reason was just exactly what you said. They said, why are you, why are you putting these policemen on leave? And it says, because the neural network said that we should put them on leave. But why? Give us the reason, give us the explanation. Yes, and so this.
You can't do that in a court of law. So do you think that AI is going to have this explanation facility soon? It's something I know as a researcher at AI they've been looking for for a long time, and it would be great if they have it, but I'm not aware that they've done it. Do you think that's going to happen?
[00:40:42] Speaker A: Well, let me tell you some areas where I think it's going to be necessary.
And we can see this in the, in the consumer market. You know, there's there's areas where, for example, Google Maps tells me to take a route that I don't normally take. You know, I would like for it to tell me why. Is there a wreck ahead that may clear up before I get there? Is there major construction that will take a day? And if it keeps telling me the wrong route and it doesn't tell me why, I'm going to lose trust in that piece of software, I'm going to lose trust in the algorithm, in the AI and just not going to use it.
So it has direct effect on whether or not these things are going to be useful to people, whether or not they're going to be able to interpret and make accurate decisions a lot of times in safety critical situations, you know. And so I think, I think the demand for knowing why AI is telling us things from a user perspective is going to become more and more important, which is why this explainable AI being able to drill down into kind of that reasoning process and make decisions based on using this as a decision aid rather than trusting completely in everything it says you should do.
[00:41:57] Speaker B: Yeah, I have a colleague here at Baylor University that's looking into monitoring websites, social media that's public. It's already public, so it's not violating anybody's right to privacy. But they look at Facebook and some of these other social media and the purpose of it is to look at these posts and this content and make predictions. For example, are there school shooters out there potential, Are there kids that are in danger of committing suicide?
But the AI itself will not be able to make that decision to stand up in any legal proceeding. So what they have to do is they have to flag these people, bring in an expert in the field, people like yourself that can actually look at and examine the data and say, yeah, there's psychological reasons here for this to happen, and the reasons are A, B and C.
But currently AI doesn't have that capability.
So that's really interesting. And if they did have that capability, and I can see them having that capability of drilling down and saying, well, this person has the possibility of committing suicide because they talk about committing suicide on their social media and such. So, yeah, this is going to be an important development, I think.
[00:43:14] Speaker A: Yeah, absolutely. You always need experts alongside it, though. I agree.
[00:43:18] Speaker B: You do, you do. And that's the reason we're never going to have AI be a physician or an attorney, because it turns out that physicians and attorneys need to be creative and you do not have that creativity in artificial intelligence.
[00:43:32] Speaker A: That's right. I mean, the more you Work with large language models for different tasks.
If you're asking it to cite a paper, most of the time it may create something that may not even be there. And it takes an expert to be able to identify. Wait a second, I don't think that study exists. So there are things like this that we see in our daily use, and
[00:43:54] Speaker B: they're still putting band aids on these large language models, but when they give an answer and it's wrong, they're so sure about themselves. Right. The large language model says, yes, green is blue. Are you sure? Yes, I'm sure. Green is blue. That's right.
That's frustrating sometimes.
One of the things as we end up here, I'd like you to tell me and the people that are listening about your business jump seat research. What exactly do you do in jump seat research to use cognitive psychology?
What's your focus? What do you do? Who's your customers?
[00:44:33] Speaker A: Well, what we do is we go in and observe these exact kind of situations where people are working alongside technology to do things mainly in safety critical environments. So you can have things like healthcare, health devices, autonomous driving, and in the military context. So what we do is work with the users, making sure that the information that is provided to them is helpful and not harmful. Things like heads up displays in a car. Right. Do you have numbers that are floating around in your windshield? Is that distracting from driving or is that giving you information when and where you need it? So.
[00:45:11] Speaker B: Oh, wow.
[00:45:12] Speaker A: Okay. Yeah. So. So we go in and look at is the information that's given to you when you're driving. Should it be visual, should it be auditory?
All of these ways that we perceive our world and then what part of that task should be done by the machine and what part by the human, really? Just keeping the human at the center of the design so that we're not just throwing out tech for tech's sake, but using it to empower what people already do.
[00:45:37] Speaker B: Okay. Fascinating. Thank you, Joe. I think we've had a good talk together.
We've been talking to psychologist Dr. Joe McDonald about psychology and AI.
Joe is the CEO and principal at Jump Seat Research, a human factors and user experience research firm in the great state of Texas in the city of Austin.
And so until next time on Mind Matters News, be of good cheer.
This has been Mind Matters News with your host, Robert J.
[00:46:19] Speaker A: Marks.
[00:46:21] Speaker B: Explore
[email protected] that's mindmatters AI mind matters news is directed and edited by Austin Egbert. The opinions expressed on this program are solely those of the speakers Mind Matters News is produced and copyrighted by the Walter Bradley center for Natural and Artificial Intelligence at Discovery Institute.
[00:46:49] Speaker A: Sam.
