← ABA Inside Track
ABA Inside Track
Episode 296 - Artificial Intelligence and ABA w/ Dr. David Cox
01:10:35
Though many of us likely use AI to auto-filI search results and find the name of songs, behavioral health has been steadily increasing the use of artificial intelligence in new and interesting ways for years. But is the tech outrunning our ability to...
Loading summary
Transcript173 lines
- [00:00]
A
Foreign.
- [00:14]
B
Hey, everybody. Welcome to ABA Inside Track, the podcast that's like reading in your car, but safer. I'm your host, Robert Perry Crews, and with me, as always, is one of my fabulous co hosts.
- [00:27]
C
Hello there, Rob. It's me, Jackie.
- [00:30]
B
Oh, hi, Jackie. Welcome. It's so nice to see you. Unfortunately, Diana, our other fabulous co host, could not be here. She'll hopefully be here a little bit later, but we had some. Some transportation issues that needed to be dealt with, so. Oh, well. But this isn't a podcast where we give excuses about where our co hosts may be. This is a podcast about behavior analysis and behavior analytic literature where every week we pick a topic and discuss it at length. And this week we're going to be talking about artificial intelligence and aba. And since that is a subject that we don't know anything about, we are very lucky to be joined by someone who knows a whole Bunch about it, Dr. David Cox. David, thank you so much for coming on the show, which otherwise, without you, would not be an episode.
- [01:16]
A
Yeah, I appreciate you all having me on. Yeah, really excited to be here. And I suppose you could always use NotebookLM or something. Just use AI to cock that AI. Maybe that would have filled the gap.
- [01:25]
B
Damn it. I don't know what you're talking about. It's very confusing. A lot of technical jargon coming from you right now. I gotta kind of dumb it down a little. I need. If only I had AI to dumb down the content of AI for me.
- [01:36]
C
I can, I can now almost spot AI in writing because it always has some. A certain. Like, like there's a lot of alliteration in it.
- [01:46]
A
Right.
- [01:46]
C
And there's like a lot of. But I'm, I'm not good at it would say. I have a colleague that's like very, like, he's like, let's use AI for everything. And I'm like, let's use AI for only generating multiple choice questions.
- [02:02]
A
Efficient. I like it.
- [02:04]
B
So we're going to be talking about. Since we never talked about AI before on the show, we're going to be talking a little bit about what AI is in general, how it can support or how it is currently supporting behavior health fields, including aba. And then maybe we'll glimpse into the far future where AI Helps has dominated all of our things, perhaps, who knows? But, David, first let's get to know you a little bit more. Would you mind telling the listeners a little bit about yourself and how you became interested in AI?
- [02:33]
A
Oh, sure. Yeah. So I got started in behavior analysis, kind of like probably Many listeners undergrad looking for any kind of job in mental health and found a in home there. I think there was called an autism habilitator. Didn't have RBTS at that time. Yeah. So just kind of fell into AVA that way. Really enjoyed the clinical work. Thought I was going to do kind of mental health, psychology, psychiatry generally. So I got my bcba, had a case that was doing that for about eight years. And then right there towards the end I was looking, looking for doc programs and I got really interested in clinical decision making. So I've seen, you know, two BCBAs, very similar clients. One was saying like 30 hours, the other was saying 15. One had once kind of like a suite of programs and goal they're working on that different from what the other one was. I just got really interested in how can we talk about or think about clinical decision making from this kind of operant respondent perspective, which got me into kind of choice research. So went did a PhD at University of Florida in behavior analysis. A lot of basic research again focused on kind of choice, judgment, decision making. And then the kind of translational work was in the applied decision making and ethical decision making, which I've done quite a bit of research in. And right at that time when I was getting into that kind of heavy clinical decision making research literature, I started popping up everywhere. You know, it's better at doctors than diagnosing cancer, better than doctors that, you know, technically you have a broken femur, all these kind of fun things. And so I just thought like, what is this tool? Why is it so good at making clinical decisions where humans might be failing? And in particular in the clinical decision making literature, they had this idea of suboptimal choice is a big one. Right. We don't always make the best decisions, things like that. So that's purely how I got interested in AI was how can we maybe use this tool, analyze large data sets, understand when we might be able to make a better decision clinically and then kind of pull that back in as like a decision support aid for behavior analysts or other behavioral health professionals and then just kind of ran down that track, you know, got further and further into the weeds of building these things. Then you see the scope of what's possible and how you might be able to bring it over. And I think full time. My first AI related presentation was 2016. So I've just been kind of running at that intersection of choice, decision making, behavior analysis and AI kind of.
- [04:43]
B
So David, there's something about a, about a. It feels like one of those technologies that seemed like a fake technology, like Star Trek technology or, you know, oh, there's a rogue AI or in video games, you've got to fight the AI computer people.
- [04:56]
A
Yeah.
- [04:56]
B
And then it feels like maybe two years ago, all of a sudden it's like, no, this is a real thing now. There's a business here. This is what we're going to be using. This is the wave of the future in a way that it hadn't been, even though it's not. Not brand new. So when you talk about AI, what, what actually is AI, what are AI models? Because it's not just one monolithic thing either.
- [05:17]
A
Yeah, yeah. It's not just, you know, magic. Yeah. So AI, you can kind of think about it like ava, where it's this umbrella term that refers to really any kind of system that hits at those two words that are in it. Right. So artificial. So it's something created by humans. Maybe we'll kind of leave it there. Silicon based often. Right. Computer or system or stuff like that. Software, hardware and then intelligence. So depending on kind of the creative designer, researcher, what they're often trying to do is create this artificial system that does something that we might consider intelligent in some capacity and that can be incredibly broad or incredibly narrow. So for example, I guess the example before diagnosing whether an image has cancer in it. Yes or no? Intelligent decision. We can create a system that can take images as input and learn how to make that intelligent choice. Yes. No. ChatGPT humans, you know, we emit textual stimuli when we type, right. All that kind of fun stuff and we string them together and it's coherent that AI there is a artificial system that's generating strings of text. So every AI system kind of has those fundamental components, right. It's non biological, typically. And then it's trying to imitate, mimic, improve upon some kind of intelligent behavior usually of humans. But there are also some really cool applications that play around with like swarm technology, with insects or birds that improve, like traffic flows on interstates and things like that. So the intelligence could be really anything, often biological. You know, humans were, we like ourselves. So a lot of it comes down to like things humans can do. But yeah, that's kind of broadly what AI is. And so it's, you know, as soon as you say that you can then sort of just like aba. For some people you could say like, oh, I use aba and they say cool, they walk away. But for all of us we'd say, okay, Right, but what do you mean? Are you VRA or like, what's your contract? We go through all that stuff. Same with AI. Right? Here's my system. Cool. What does that mean? How are you building it? What algorithms are you using? Kind of what data goes into it.
- [07:12]
B
So I know one of the things I really liked about, about the articles that we'll be talking about, which since Diana's not here. Jackie, will you introduce them in a minute? Sure, I'm queuing you up. I'm queuing you up. But not quite yet. But one of the things that you go into is sort of describing the different kind of branches of AI. You know, so we sort of have like the generative AIs, we have our, you know, kind of the more, more visual AIs or the robot. I never thought of robots exactly as AI. I thought, I've just thought about them either as like killer robots or more likely the things that like build cars, you know, the big arms. That's all, that's all they do. But really the idea of them moving, you know, their AI for movement and moving without crashing into things.
- [07:51]
A
Yeah, yeah, yeah, absolutely.
- [07:54]
B
What, what, what, what other ones am I missing? Because I think we talked like kind of the text, the textual AI or the visual AI, motor AI. And then I guess, you know, what's kind of the AI that you think of when you think of this AI will be good for ABA or all of the above?
- [08:09]
A
Yeah, I think all of the above. I think, you know, if there's one thing that I would love listeners to walk away with, it's really understanding that difference between generative AI and discriminative AI. So I kind of mentioned that example of the. Does this image have cancer? Yes or no? The system is learning to discriminate between. You have conditional discrimination, just like we would use. And then generative AI. The goal of those systems, I'm going to generate something new, novel, unique. And that's where when I think about the types of AI that are most likely, I think long run, going to be beneficial for ava. It's getting back to that discriminative AI. There's certain things called unsupervised machine learning, supervised machine learning. We kind of get into those things, but it's really kind of getting down to, you know, what are the decisions where more data, more information, something could be added to make our decisions better. And usually we're going to train a system to be really good at that thing. And it's different than, you know, chatgpt, it's been trained on the Internet. You don't Quite know what you're going to get back every time. I'm not sure it should be writing treatment programs just yet and those kinds of things. So you know, it's sexy, it's novel, people really like it and I think it was their first exposure to AI. But long run I think it's really going to be these alternative types of AI that are more beneficial. Computer vision. I think that's another good example. You know, if we could start automating data collection in sessions, that would be incredibly useful. That frees up all sorts of time and things like that. And that's not going to be a generative AI approach. It's going to be, it's kind of more focused discriminative AI approach.
- [09:35]
B
That's the AI I want not, not to jump ahead to like maybe talk is because there's something I, I love having my data later on. I absolutely hate doing anything with a client and also collecting data at the same time. I could do one or the other I'm very bad at doing. But like I, I've been trying to do more, you know, do something, do assessments and I'm sort of like, did anyone get that? Anyone say any data? I was, we're the client, we had a great time. But I don't remember what happened. I need that data. Right. You know, it's very sloppy. I want the really. I want AI to just get it all for me. That would be really cool.
- [10:05]
A
Oh yeah, yeah. And then you can imagine too. I mean we have like all these wearable technology sensors and things. We've been restricted historically to data that someone has, can like look and see and collect and whatever. You know, we think proactively. Oh, here's how I create my data sheet. Like what if you had everything like literally like matrix style, right Ones and zeros and everything. That's where I also get excited like what kind of new information or insights about behavior might pull out of there. Because that's the, the stuff we don't know. We don't know yet because we just haven't been able to collect data on it.
- [10:33]
C
Yeah, it's fascinating. One thing I want before we talk about the articles, one thing that I just tried recently with AI is I use ChatGPT too because I have. But it was the paid version, it wasn't mine because I don't pay for anything because I'm cheap. But somebody else paid for it is they asked AI to do a role play about a supervisee coming into a supervision session in a, in ABA and, and Said that they're nervous about this, but they need to advocate for themselves. And you're the supervisor, and the AI is the supervisee. And the roleplay was amazing.
- [11:13]
A
Oh, yeah. It's incredible, right?
- [11:15]
C
I couldn't believe it. I was like, this is not going to take the place of, like, real role playing. And I was flabbergasted.
- [11:24]
A
Yeah. And it's cool, too, because you could go in and create these Personas. Sometimes they're referred to, and you can tweak certain variables. Like, I want you to be really angry now, or like, I want you to be incredibly hard to get information from. Start. Exactly to your point. You can have people contact all these scenarios. They're going to contact out in the wild, but a safe, controlled environment and manipulating the things that you, you know, they're going to contact.
- [11:47]
C
I just. I just thought that was fascinating. Oh, my gosh. When I did that, I got rid of your articles. Okay, I'm here now. I can tell you what the articles we're going to be reviewing today are. I'm not as good as Diana, but I will surely try. The first one is starting the conversation around the ethical use of artificial intelligence and applied behavior analysis. This was published in 2024 in Behavior Analysis and Practice, and it was written by Jennings and Cox. And the second one is also written in 2024 big year. And this one was titled the Promises and Possibilities of Artificial Intelligence and Delivery of Behavioralic Services, also in Behavior Analysis and Practice and by Cox and Jennings. You guys switched it there. So both of you got to be first author.
- [12:31]
A
Second author.
- [12:32]
C
Second author. First author. Those are the two articles that we'll focus on tonight.
- [12:40]
A
Yeah.
- [12:41]
B
So, David, when we talk about AI, we kind of talked a little bit about how it feels like one of those technologies that's been around forever, but really only recently has sort of become kind of like VR. The same thing of, like, that's been around forever, but it's only been in the past, you know, decade or so that anyone's actually been like, I think we can make money on this, or we really need to double down and focus on these technologies. So when you think of where AI is on, like, the scale of, you know, like a speak and spell to a terminator, where. Where is AI on that scale?
- [13:11]
A
Yeah, around the corner. Oh, yeah, yeah. It's interesting, I think, how I often think about. Have you heard of the Turing Test? Yeah. Yeah. So there's the Turing.
- [13:24]
C
I haven't. I haven't.
- [13:25]
A
Oh, perfect. Yeah. Explain it really fast. Yeah, yeah. So the Turing Test is this basic. Alan Turing is the guy that came up with it. And it started way back in the day with artificial intelligence, where it was like, how do we know if this system is truly intelligent or might kind of pass the smell test, if you will. So the idea is, all right, you set up some kind of scenario. A human is interfacing through a computer with two different agents, if you will. One of them is a human, one of them is a computer, and at the end of some fixed period, 5 minutes, 10 minutes, 30 minutes, depends on who you chat with. Can the human reliably distinguish between which one is a computer, which one is the human? And to the extent that you essentially can confuse the human, you would say, oh, this thing passed this kind of Turing test. Since then, that's kind of like the original version of this. Like, can you convincingly communicate with a human and be a robot and they think you're a human? There have been some additional layers added onto that. So I think that original Turing test, I think we're past that. I think anybody that's engaged with Chad, GPT, or some of these LLMs, easy to be confused if you maybe don't know what's going on. Kind of the next level up and this kind of gets more into the different types of AI is like, you can imagine also that we don't just type in a computer. Right. But I can also talk. I can other types of verbal behavior. I can walk and move around my environment, and I can emit all sorts of different intelligent behaviors, if you will. Right. We have a whole behavioral repertoire. So the kind of next wave of Turing Test is, are you not just good at one thing, but can you start stringing multiple of these modalities together? Are you good at talking and, you know, doing math, and can you move around your environment? Let's say I think we're getting closer to that kind of second level. Many of these systems are getting better and better at more than one task that we would consider intelligent. For example, ChatGPT and other large language models. They can do math. They're not always great. They can, you know, code, write computer code. They can talk to you about, you know, Shakespeare or whatever. The next level up really is that kind of embedded in a. Like an organism that's living some kind of life? Not there yet. We're. We're certainly not there. And if you look at the total behavioral repertoire of humans, not. We're not even close to that. And then you kind of keep going up the chain of These levels of Turing tests and you know, eventually you have this system that we looks like a human, walks like human talks until you crap open the head and like CEO, it's all computer parts inside. Wouldn't know. And we're certainly not there yet. So I don't know if that kind of answer. You know, we're between, between speak and spell. You know, we're not a terminator. It's not moving around killing us well beyond the speak and spell, I think. But there's also, I don't know, we're also around the corner. We have cars that can like self drive and things, which is kind of like cool, but they're not going to, you know, make us bread or anything. But yeah, it's wild, wild time.
- [16:13]
C
I. Oh my gosh. I'm just saying, I would pay so much money for AI to make me a good sourdough every day. Did you say we're. We're not around the corner from making bread?
- [16:25]
A
Oh, no. I say that the self driving cars, that system cannot also make us bread. Oh, but you might be onto something. I mean. Yeah, yeah. Create a little boss that makes sourdough. Right.
- [16:38]
C
I would pay so much money because it's so high maintenance to take care of sourdough starter.
- [16:44]
A
Yeah.
- [16:46]
B
So, David, is the development of AI one of those things that has just sort of been chugging along for a long time and it really hasn't spent because it's one of the things that I always kind of was wondering about AI was has the pace accelerated due to or has it just been the increased focus because now there's money to be made? I mean, I kind of think it's the acceleration to some extent because even in your paper or one of your papers, there's a footnote. I thought that was very amusing where you note, since we started writing this paper, there's been another iteration of this. We're going to assume a lot of the principles are the same because we're not going to keep up with writing paper. Like papers take so long. AI generates itself so fast these days, so it seems like there's a real acceleration happening. Is that your sense too, or is it more. No, it's always been there. It's just been moving along and we're just paying attention now.
- [17:34]
A
Oh, sure, yeah, yeah, yeah. It was interesting to that footnote, we submitted both these papers technically before ChatGPT came out and the world like knew what it was, which is crazy then to just see it anyway. Yeah. So AI technically did get started, I think 19 at Dartmouth College. There was kind of this AI summer program. Bunch of people got together early. There was kind of an initial wave and then there was what's called the AI winter. And then it kind of kicked back up again in the 90s if you will. And then it's kind of blown up since then. If you look historically there have been a few technological advancements that have helped. And then I think more recently it's been money because of ChatGPT and other large language models. But the big developments were early on. Essentially a lot of humans figured, you know, I have this system, I want to make intelligent. I don't know, let's bring in an expert who knows how to do this thing and they can just walk me through like task analysis style. Like you do what, then you do what? And I just like try to code it into a system. These were expert based systems, kind of rules based systems. They largely flopped because we can't account for all the nuance and whatnot. So there was kind of this try. It failed. And early on too there were claims, you know, in three years we'll have robots that can replace humans and all fast. And then that, you know, we couldn't even get it to solve basic math. We I wasn't alive then. They. So that that approach kind of failed. Then you get a little bit later on, 90s early thousands and a couple of big things happened. One was the development of neural networks, deep learning techniques. So this was an alternative approach to kind of crunch data and solve these problems that we hadn't had before and it turns out as incredibly effective. So that was kind of one big thing. The second, and this one always kind of blows my mind is technically a lot of the machine learning algorithms. And so what these do is look at a large data set, identify patterns and then you can kind of use that information. We haven't, we've had these algorithms for hundreds, thousands of years. And some instances we just didn't have the data to actually run it through and teach a system. So it also kind of came 90s thousand. We have this kind of big data revolution. All sorts of technological advancements, both in the production of data, people started collecting it on everything and then storing it. So you kind of had this interesting interaction of all these things that kind of happened at the same time. We now had data to test out some of these old machine learning, more data driven algorithms. We have neural networks that are solving problems that we couldn't before. Where people are interested, they can save this stuff for cheap. So now Businesses are like, maybe there's value here. And then all really kind of blew up. And that was really the wave that I kind of caught in the early teens. Is that what we call it? 2010? And then I think the game changer in the last few years was really Chat GPT putting their model online in an interface that everybody could interact with. And that was big because I remember 2017, 2018, interacting with the earlier versions of that model because they've been building it for, you know, six, seven, eight years. It was interesting. It felt like kind of this toy example. I didn't see the use case for any of my research. And so when the one that was released online came out, it was just kind of the next one in their iteration. I was like, whatever. But I used to have to interact with it through Python, like through code. Putting it in this kind of chat interface opened it up to everybody. And then you have, you know, billions of people with ideas on how to use this, and it just took off. And then venture capitalism, money sees opportunity that gets dumped in. But you can look also, I take a quick step back to if you're looking at the advancement of this stuff over time, like it beat the world's top chess player, then it beat Lisa Doll and go. So there's like these systematic advances. If you look historically, you can see it getting better and better. And then, you know, 2022 happened and floodgates, money dropped in. And yeah, here we are today talking on the podcast.
- [21:18]
B
So. So when we look at. We'll be focusing on. Because it's the focus of your articles on behavioral health and ABA kind of in. In the healthcare system, which includes us in aba. But it does appear like AI hasn't just become something brand new because certainly, you know, one of the things that was great about the two articles is it covers, you know, here's a nice overview and, and here are some ethical questions that we still want to answer. But your citations, I mean, there are studies you'd done before using AI previously. So, I mean, how. How long or sort of how is AI in health care been. Been used? I mean, it's been around for at least, you know, a number of years, right? Because, I mean, you've been studying it for longer than just, you know, the time it took you to write that last article.
- [22:05]
A
Oh, yeah, yeah, absolutely. And part of my motivation with that was I kept going to these behavior analytic conferences and be like, what are you into these days? And I try to explain it, and no one had any idea what the hell I was talking about. So I was like, you need to like show people that others are doing this. And I think there's, in the promises and possibilities paper, there's that figure one that I think kind of captures. If you think about the, I think we call it the prototypical patient journey or something like that in that article, from diagnosis and assessment through designing an intervention, in some instances delivering it, monitoring it, and then maybe withdrawal and kind of the admin systems that wrap around that. Every one of those points involves all sorts of different decisions and choices that we make as a clinician, behavioral health clinician. And that's exactly what AI does, right? What is that intelligent task, that thing, that decision that you're trying to make, and how can I use data to help you do that faster, more accurately, with less cost, things like that. And so it's looking at these and across all these other areas of literature, but I just thinking, my goodness, it's really kind of touching all of these different areas. There's no reason we can't start trying to bring some of that stuff over into behavior analysis to take advantage of it.
- [23:16]
B
I really did love that chart. It captured, I think, what could be done with AI in a very succinct and practical manner. Like you said, sort of what's the healthcare journey and examples of all of the ways that AI could play into it. Because certainly I know peek behind the curtain, you shared a number of resources that could discuss on the show and we picked the two articles, A, because we do, you know, articles on the show and B, because I started looking at some of the other links, some of the other presentations, and, and I had no idea what it, it was because it was very, very advanced and very, very smart. And I didn't know I would ask you questions other than, so what's that? Tell me about it.
- [23:57]
A
Yeah, yeah, that's fair. That's fair. Yeah. And, and I think when I, when I talk to people that are might be interested in this area, it really, especially right now, feels like the sky's the limit. You know, kind of like you were saying before, like, what's your pain point? I don't like collecting data and trying to interact with a client. Great, let's build a system to do exactly that. Another one that I really have been interested in for a long time, going back to when I first started getting into clinical decision making, is that idea of how do I pick how many hours a week of therapy a kidney? How do you make that decision? My, my history and training was kind of like I don't. We do it this way, you know, ask for as many as you can, because insurance is going to negotiate you down. But that's another area of work for me. I was like, you know, we have all this rich information on clients as well as past kids that have gotten ADA and how many hours they received. Why don't we start bringing that together and letting AI crunch those numbers? That was my interest. But again, anybody in the audience here, I'm sure, has something they're interested in, some kind of painful decision that they wish they could make better. Then the question becomes, cool, what data do we have on that? How can we start using data to make some kind of more informed decision or creative system around it? It's cool. Yeah.
- [25:06]
B
So that actually brings up a great point, David. How do we create these systems? Because certainly we go through the journey. Okay, well, nobody, you know, everyone wants to know early on when there's a problem, something is, you know, they're ill. Like, like you mentioned, sort of, you know, AI looking at your, your lung scans and seeing is there is not cancer. We have wearables that could sort of, you know, send data somewhere. And AI does something and it's, hey, there's something. Your, your gate is different. Where it's changed over the past two weeks. You should go see the orthopedic surgeon or, you know, so we sort of want to have that. We want to have something that can collect data. I mean, I want the data, collect all my data so I can just focus on the interactions. Even like administrative tasks, you know, the flow of scheduling. You know, we. I've recently had to take one of my children to a hospital, and, you know, you're sitting in the waiting room going, there's got to be a better way to do this. Like, we're just hanging around, like, did they not schedule enough people? Did somebody call out? Could AI have avoided all of this and got everyone in and out so fast? But how does all of that generate? You know, I, I sort of just imagine that, well, we put it in the computer and the computer does it.
- [26:11]
A
But.
- [26:11]
B
But it's not. AI is not a computer. It's. It's a program or an algorithm. It's. It's kind of all of that. So how did, how did these systems develop? Like, how did, how do they get created?
- [26:21]
A
Yeah, yeah, great question. And I think this is the critical thing, like AI literacy that I hope everybody really starts to fundamentally understand, because it'll allow us, when we think about ethical AI, to, I think, start asking Questions in some of these companies, but it's very intuitively simple. So let's take the, maybe the cancer example, right? So I have two images or. Well, sorry, one image, one image. And you know it's going to have a bunch of pixels in it. You can imagine the grid of squares. Each one of those pixels is associated with data. What color is it? How bright or dark is it? Maybe what are the orientation of lines? Maybe it will just kind of stay there. You can imagine I have this kind of. Now this list of numbers that describes this picture. I can then put a label next to that and say, yes, I know this is cancer in it. Cool. Pull up another image, a new, slightly varied set of data, and I can say, no, this one does not. Awesome. And then you can imagine I just do this repeatedly. And what the computer is essentially being fed is specific patterns of data associated with the label. Yes, no, cancer or not. And what it's doing is mathematically we're kind of modeling or create this representation of those images that's associated with the label. And then over time it learns, ah, this pattern means yes, this pattern means no. And if that mathematical modeling sounds a little abstract, if anybody's played around in Excel and you have your data path and you right click and say, add a trend line and it just fits a line to it, that's all this stuff is doing under the hood now. It's more complex models than this linear regression thing. More data and things like that. But at the end of the day, that's really all it is. I have some set of data with a known output and I'm teaching the system to match those things up. And then once I have that, once I reliably know, let's say going back to our linear regression kind of trendline example, if I predict out into the future, you know, next one, two, three, four sessions, I can start to make guesses about where you're going to be. And so the same idea with a new image, if it were to come in, I'd have this model run it through that and it would kick out based on information that it trained on. Yes, no cancer or not. But why that really critical and I think important for people to understand is that these systems are only as good as the data that were used to train it. Right? If I take, say your behavioral patterns and I create a trend line to fit that, it may or may not apply to me, may or may not apply to somebody else. So when companies are out there selling products of, hey, we built this AI system That does xyz. My first question is, cool. Whose data is in there? How well does that match the patients that I might work with or even the clinical scenarios that I work in or the type of ABA therapy that I provide? Because to the extent that there's a mismatch, you're necessarily going to get kind of inaccurate prediction.
- [29:12]
B
Yeah, I, I, I think you and Adrian had that good metaphor. I think of the, it's like the, the black box. But then thinking about AI as you know, can we think about it like it's a, it's a chef or it's a recipe, you know, where you go to the restaurant and the chef can show you. These are the things that go into my awesome recipe. But there's some kind of chef magic that has to happen in between. But you, you still know that the ingredients being put in are not going to kill you. It's just exactly. There's something about how it all goes together that maybe you don't get that information or you don't even need to know that information. You just need to know everything else is, is, is above board, whether it's ethically or whether it's the data that would be relevant to the problem you're trying to solve.
- [29:47]
A
Yeah. Yeah. I'm glad you like that analogy.
- [29:50]
B
It's very good. Hey everyone, Sorry to interrupt the conversation, but we're going to take a little break and then we'll be right back talking about Artificial Intelligence with Dr. David Cox. We'll be right back.
- [30:13]
C
Hi. Do you want to be a bcba, also known as a Board Certified Behavior Analyst?
- [30:19]
D
Sure. We all do.
- [30:20]
C
Now you can come to Regis College in Weston, Massachusetts to get your graduate degree just minutes outside of Boston.
- [30:27]
D
Choose from any one of these courses.
- [30:29]
C
Masters of Science in Applied Behavior Analysis.
- [30:31]
D
Master of Science in Special Education, Dual degree in Special Ed and aba, or.
- [30:37]
C
Be eligible for your Postmaster certificate.
- [30:40]
D
You can complete your degree and be ready to sit for the exam in two years.
- [30:43]
C
And our 2022 graduates had a 92% pass rate on the BACB exam.
- [30:49]
D
Come enjoy approved fieldwork, placements, ethics mini handbooks, PhD levels, professors, small class sizes, and a service trip to Iceland if interested.
- [30:58]
C
And don't forget, our program is accredited by the association of Behavior Analysis International or ABAI as a Tier 1 master's degree program.
- [31:07]
D
Don't delay. Supplies are limited. Learn more at regiscollege. Edu.
- [31:12]
C
Again, that's www.regiscollege.edu regiscollege.edu One more time, www.regiscollege.
- [31:22]
D
Edu See you there.
- [31:24]
A
Bye.
- [31:31]
B
Hey, everybody. Just wanted to pause really quickly to remind everyone that ABA insidetrack is ACE and KWABA approved. By listening to the show, you are able to earn one learning credit. All you need to do is finish listening, then go to the website ABA inside track.com or click on the link in your podcast notes player and, and enter in some key information there, including two secret code words. And the first code word is welcome. W E L C O M E. Welcome. A message of greeting when someone arrives somewhere, like AI arriving on your computer. Welcome. All right, and now back to the show. So Diana is joining us, so she'll have some, some questions, I'm sure. Later. Danny, you already missed the part where we did the articles. Jackie. Jackie did that.
- [32:24]
D
Oh, that's okay. So I'm sure you did great, Jackie.
- [32:28]
B
So we talked about a lot of examples now, David, of how AI could sort of hit some of those points within that behavioral health journey. Are there any that you mentioned in the article that we haven't sort of talked about that you're like, this is the one. Like, that's the one that gets me excited. Just talk AI every day.
- [32:42]
A
Yeah, yeah, definitely. Yeah. So there's a branch, I guess, two of them. I'll just talk about my favorite. Let me, if you want to get going, to the second branch of machine learning called unsupervised machine learning. So everything that I've described so far, we essentially have a data set where I'm information about whatever observation and then the right answer. And I, I'm teaching the system just like we would with the learner. Right. You know, what color is this? And then you give them feedback. Yes, no, unsupervised machine learning. I don't know the answer. I say, hey, I have this complex data set based on, usually we test a variety of algorithms. I want you to go in and find interesting patterns and kind of report back to me what you find. I love that area of artificial intelligence, mainly because it starts to get at those questions of what don't we know that we don't know? And can these things present us, even if they're wrong, with interesting ideas, avenues to kind of test go down and what have you in future research. And one kind of example of this, I think I may have sent a link to this, maybe not. But I think we go back to kind of this idea of how do we create the best interventions for the people that we provide services for? They're coming to us with this incredibly complex dynamic learning history. All sorts of things. And if there's a way that we can start using unsupervised machine learning to analyze the totality of that history and start identifying different types of patients or clients almost, if you will. And we'll be able to get, I think, to more and more precise interventions and how we come up with things. And mainly where I get kind of feisty. If you look in the literature around different groupings or clusters is the technical term of individuals with autism spectrum disorder. Most research studies will say like three, four, maybe five. But if you've worked in this field, there's, you know, there's certainly more, more variability than that. There's that fun kind of saying that every individual with autism meet is, you know, unlike every other individual. And that's where I think some of these tools and unsupervised machine learning might be useful, is we can say, hey, we have data on that complex, rich history of who you are. Let's see if we can start identifying patterns that relate to how we deliver services. So we can do better than not using that information and just not knowing like we kind of do today.
- [34:49]
B
Yeah, I did, I did enjoy sort of thinking about the idea of, you know, learner profiles. How fast do they learn? Who do they learn best with? You know, and then matching your clinician to your client in a way that everyone is maximally using all of their strengths and, or, or, you know, helping each other with weaknesses. That, that did seem like a, A real fun combination of some of those, like personality tests, maybe.
- [35:14]
A
Yeah, yeah.
- [35:15]
B
Then your algorithm from that, perhaps. I don't know.
- [35:17]
A
Oh, yeah, absolutely.
- [35:19]
B
Now, are there any sort of AI areas where you have people talk to you about AI? They say, oh, I'm, I'm putting together an algorithm to solve problem X. And you sort of just roll your eyes and say, that doesn't feel like a problem worth solving. Like, I think that's one that we just want to let humans just do as is. Or is it really one of those. Are we at that blue sky phase of like, every idea is the best idea when it comes to AI?
- [35:41]
A
Yeah. No. So I don't know that there's necessarily a particular. Well, there's one. One topic that I think is overplayed, perhaps. Maybe we'll talk about that. But in terms of, like, good and bad ideas, I think the one thing a lot of people don't think about is the cost to build these Systems. So, like, ChatGPT is good because it's swallowed as much quality data on the Internet that exists that's incredibly expensive. If you look at a lot of their financial reports, they're in the red, they're losing money even though they have like these billions of dollars of investment. And so that's where, you know, sometimes people will come in and say, oh, I have this crazy idea to build, build xyz. And you start asking like, well, where are you going to get that data? Like how are you going to get the compute to actually train these things? And you start stacking up dollars and the benefit is trivial, like no one's ever going to pay for it. It's not going to be self sustaining. So those are the instances when I engage with people about ideas that it's more kind of like the reality check of what it takes to build these systems. And then just because something's a good idea doesn't mean people are going to buy it. And if nobody buys it, then it's, yeah, you're never, it's not going to survive.
- [36:43]
C
So David, do you think. So like with Chat gp they have like a paid version and the free version. Do you think that's like not great for them, right?
- [36:53]
A
Oh no, I think, yeah, I think they need to do it because if they don't make money there, and that's where I get really interested is like we're in this weird phase where I think a lot of people have started to rely on some of these tools for certain skills. And I use AI throughout my day to day life in many different ways, but there's a very real possibility that these are not financially successful and the plug gets pulled on them. Right. So like what then? And there's some companies building their businesses on these tools so that, that's dangerous, risky to me but they need to figure out some way to make money. So I appreciate that they offer the paid version and whatnot, but it's. Yeah.
- [37:30]
B
And I mean, David, do you ever worry when we're talking about AI and I know we're going to get into ethics pretty soon, which is its own can of worms when it comes to AI, but that just the amount of data available, readily available at costs that are, you know, can be swallowed by these businesses are just going to result in these AI models that like, look, this AI model is amazing at detecting cancer in specifically this age range, specifically this, you know, sex at birth, specifically this part of the country. That's the only data we got. We have no idea what it's going to do for everyone else and just make it just so limiting in how, how well we can use like, it's great at this, but that's it. And you can't make money off of of, of something like that.
- [38:08]
A
Yeah. And it costs a billion dollars a year. Yeah, yeah, yeah, I do. So I think there's like an interesting, almost like democratization that I think would be beautiful. And this is my secret dream and behavior analysis as well. It's that we know these Large language model ChatGPT, right. Cost very incredibly expensive. Realistically, we're never going to build like the behavior analytic version of that. There's just nobody that has that kind of money and no one cares about our field enough to spend it. But you know, if we get everybody, every clinic, let's say, has like a data scientist and then you start using the data from your own organization to build a model that's really good for you and your company. Then I think you kind of get in this interesting space where it's. We're building models that are really only good in a narrow space, but that's all they'll ever be used in. So it might be cost effective for the company to have these kind of smaller models specific to their use case. Whereas it's like these gigantic ones are just cost prohibitive to actually build. But that also takes, you know, you have to pay for the data scientists, pay for the compute and all that kind of fun stuff. So it's not as though there's no cost and stuff to return get return on it in some way. But yeah, it's wild west.
- [39:17]
B
So that'd be something like. I think it was. Was it central reach? I think that did AI for scheduling. That'd be kind of one of those test cases of, well, we know what employees we have, we know what clients we have, we know where everybody lives. Let's put that date. We know what the price of gas is every day. Let's get that together. We can make something.
- [39:34]
A
Yeah, yeah.
- [39:35]
D
The thing I was the most excited about because I used to do, I used to have that job where I had to schedule everyone to go to everyone's house. And it's really, really tricky. And so like we would have, you know, three, like three or four therapists trained for each child. And so if that person was that sick, we would try to rearrange the whole schedule so that everyone was with someone that they knew. But it could, it was this whole cascading effect and it was really, really hard to do. And at the time I was like, if only there was a computer system that could like figure out all these permutations for me, but there wasn't. And it could take like an hour every night if someone called up sick, which happens a lot.
- [40:14]
B
And then. I hate to tell you this, but if there had been a computer system, it probably was way better than you.
- [40:19]
D
Oh, I don't doubt it. I don't. I don't doubt that at all. So I was excited by that. Hopefully it works well for everyone who's trying to. Who's living in scheduling hell. Like I was.
- [40:29]
A
Oh, yeah. And even if it saves, you know, like you're saying it takes you an hour every night, even if it saves you 50 minutes now you get 50 minutes back when you're like. But I think it also highlights a lot of the. I think short term use cases of AI and ABA are these kind of low risk, high pain point, boring administrative things. You know, they're not the crazy sector. Like, we're going to collect data on every behavior for everybody.
- [40:51]
D
Yeah. And in the ethical realm, I feel like it's on the very safe end as well.
- [40:56]
A
Yes. Yeah, absolutely.
- [40:57]
B
Yes.
- [40:58]
A
Oh, man.
- [40:58]
B
My c. Sorry. My. My model only took an extra five minutes in your drive. Like, oh, it's. Sorry about, you know, it's just sort of annoying more than unethical.
- [41:07]
C
And you can blame someone else then too.
- [41:10]
B
Like, like the ghost in the machine.
- [41:12]
C
Or like, oh, man. So sorry I did this. Not me.
- [41:17]
A
Exactly.
- [41:19]
B
So, Dave, this probably a good time to move into the ethics. Into the ethics of AI and when you, when you and Adrian were sort of talking about, you know, putting these papers together and sort of, let's do an introduction, was it always the thought of we should have a paper that's. That's so excited about AI and then we need to say, and everybody, slow your roll. There are a lot of questions that need to be answered. Or was there ever a version where you're like, what if we just wrote a 50 page paper that just shoves it all into one place?
- [41:44]
A
Yeah. Yeah. Probably more the latter, I think. Kind of like I was explaining with my history, ethics is always just kind of on the edge of everything that I'm doing. Think about these things. And so the original paper was this kind of blend, but what we found we were doing to kind of reduce space was either leaving out some of the use cases stuff or we were cutting the ethics stuff a little bit short. And so we finally, at one point were just like, we just need to separate this thing out. And what was really great about Stephanie at Behavior Analysis and Practice is she was great in working with us where we submitted, we said, hey, these are kind of companion pieces. Can they kind of be considered, come out together and all that kind of fun stuff? And they were great working with us to do that. It's kind of unusual for two papers that go through like that, but, yeah, they're just. I mean, there's just so much to talk about in this space. And as you can see, I get excited and I'll just talk forever. So, you know, easily cute papers for.
- [42:33]
B
And I think I. And I don't know how much input you, you, you had versus just the, the editing board, but I believe the Ethics came first in the journal and then sort of the, the looking into the future of AI I don't have.
- [42:45]
A
Oh, that could be.
- [42:46]
B
Yeah, but, but. And that's how I read them. I tend to. I like to read papers chronologically because it gives me a sense of sort of where we were and where we're going. And I really liked reading about Ethics first. Or else I just read ethics first, whatever the reason. But I really liked reading about Ethics first because it made me very cautious. So when I got to the second paper and was looking at all the exciting ways that AI could be used, I wasn't, you know, down on it or saying, well, this isn't because I know you. You, you co written both of them. But it definitely tempered my opinion of. All right, well, these are exciting and we need to be careful about it because I think one of the challenges with AI that you don't necessarily call out like VCs in the paper, but I assume that's going through your head as well when we're talking about these systems of, there's money to be made, so dump it all in. Don't worry about it. Who cares? It's a computer. Don't worry. I don't, I don't. The Belmont Report, like, like the horse race. I don't know what you're talking about. Money, money, money. Let's. Let's forget about, you know, where the data is coming from, whether it's equitable. That must always kind of be on your. On your mind. So I appreciated that. Context.
- [43:51]
D
Ethics was published second.
- [43:52]
B
Oh, Ethics was second.
- [43:53]
A
All right.
- [43:53]
B
Then I just read them the other way around then, for some reason.
- [43:56]
A
Yeah, I appreciate your preference, though. Yeah, I'd like ethics, too. Yeah. You know, it is interesting. And I, And I think what's unique, a lot of ethics and behavior analysis, we have the BACB Code. Most certificates are, you promised to uphold the code and whatnot. And so there's Some contingencies there. Every tech company that's building AI, there really is no regulation right now, no motivation to follow any ethics code whatsoever. It's bad for business if I make wrong predictions. But you also have the motto of Silicon Valley, move fast and break things. And as we kind of mentioned before, a lot of these systems, incredibly costly to build. There's an upfront cost that they need to recoup if they want to continue to try to build and do creative things. And so their job at the end of the day, regardless of how good or bad it is or the blind spots of the model, is they want to try to sell it to as many people as they can to get them to buy it and use it. And that, you know, that worries me kind of like we talked earlier about how these models are built. If people aren't being transparent about who's in the data, where the model performs well, where it doesn't, is it biased in some way? And I mean that more in, like, the mathematical sense. At the end of the day, every model is going to be biased to the data that it was trained on. So it's just like, just be upfront about that. Like, you don't have to pretend like it's not biased. And that's where I get nervous. How much are they communicating? And if they're not, why are you trying to get me quiet? What's the catch here?
- [45:17]
B
Yeah, I mean, you certainly did, you and Adrian did a great job, I think, laying out sort of the pillars of where AI could be unethical, because certainly, I think we all go to the idea of, well, it could be broken. I think Jackie brought that up. It just doesn't work. And when you're doing something like scheduling, maybe that's not a big deal. But when you are making treatment recommendations, if your model's off and everyone's getting the hours that they should be getting sliced in half because of a rounding error or whatever. I know that's not how AI works, but excuse my limited understanding then, that that's, that's going to be a huge, a huge danger. But you also go into sort of the areas of autonomy related to AI beneficence with AI and injustice, which, you know, these are areas that I certainly think of in terms of ethics, but I never really thought of them as to how they could be applicable to, to AI models. Do you mind sort of, kind of given some, some broad strokes as to some of the areas that you're most concerned about when it comes to AI and ethics? In aba.
- [46:16]
A
Yeah. I think one of the ones that I've brought up a few times thus far really is that transparency kind of explainability of the model, which goes back to really that harm benefit. Like, all right, you're, you're giving me this tool, you're telling me you can write a treatment plan, prove to me that it's not going to cause harm, or simultaneously that's going to kind of optimize or maximize benefit for my client or at least do better. I also recognize, kind of going back to the suboptimal decision making humans, we make a lot of mistakes. So the baseline isn't perfect. It's like, is it as good as a human or better? So that's the one that, that really I think concerns me the most. Another one that you kind of brought up is this idea of equity. As I mentioned, a lot of that stuff costs a lot of money, which means that especially as it's rolled out early on, clinics that have the money to pay for it, can pay for it, their clients or patients are going to be disproportionately better off. Right. Than those that can't. And you already have a little bit of an interesting scenario in ABA now it seems between kind of smaller ABA companies, these large VC backed companies in terms of resources and things like that, this could be another kind of pile on that, that brick that further separates out some of these inequitable distributions, if you will. And then the last bit that I don't think anybody's really solved in AI ethics generally is this idea of kind of autonomy, data ownership. You know, data data is a really interesting resource in that it can be used infinite amounts and moved around while remaining in one location. So at any point in time, if I choose, like I'm not comfortable with my data being in the system. There's really no good answer to that. Right. Your data is already popping. I can't go in and like adjust the model weights back by that little hair. Just your example, a kind of horse has left the barn, if you will, and we don't know how to solve that. And I think we also don't have a great system in behavior analysis right now of talking with clients about how do we make clinical decisions. Are there colleagues that we consult with? Do we, you know, a particular area of the research literature, rt, act or like, it's whatever that influences how we think about delivered therapy and we don't really bring that into the informed consent process now. And this I think is another tool. How do we start talking about These things, you know, I'm not perfect. I'm going to use tools to help me help you. How do I talk with you about that? What am I using? How do I use it? Are you okay with that? Yeah, we still have a lot of good language around that are processes which is opportunity for research.
- [48:40]
B
Yeah, I just thought it was, I never, I never thought about it in terms of informed consent because I mean just talking about AI, we're like, oh yeah, AI, yeah, you know, AI the thing, AI, the computer thing. And it's very different when you're talking about, you know, what is Chat gp, How does it work? I don't know, it answers questions than someone saying, wait, you want me to, to sign over my data to you? You're asking me to pay you money to provide a service using this? Well, tell me about AI. And I would guess the majority. Not to, not to knock the majority of BCBAs out there, but if you were hard pressed, like describe how AI works. I mean you get a lot of blank stares back and just how, how.
- [49:17]
D
Awkward that is whether any of us understand how well your data are protected if they've been fed through.
- [49:23]
A
Oh yeah, right. Yeah, yeah, Very, very dangerous.
- [49:27]
B
I like your example, David though, of the person saying yes, I'd like to remove my informed consent for my Data that's in 50 different models now. Could you hit delete back backspace, backspace, backspace, please.
- [49:37]
A
Yeah, undo, undo, undo. Yeah, Impossible. Yeah, it's crazy and there's been some really interesting and I don't know why anybody would ever target an ABA client. I guess you never know. But there have been some interesting scenarios where companies have used the public version of Chat GPT. So there are ways to engage with these large language models that are in like a walled garden state that requires the ability to like pro program and write code and stuff. But there have been some instances where people have put large sets of company documents in this thing, the public version. And then other individuals, rivals or whatever have been able to extract that information out because it's now in the database in the system. So that's one of those, you know, anybody's out there, if you put. Every time you put PHI or FERPA protected data, you're technically violating the law, right? Yeah. These are things people should be aware of. It's now in that system, someone could extract it out if they knew how to do that. Again, I don't know why anyone would care about a random getting information about somebody, but it's there, it's technically in that system forever now can't want that back.
- [50:32]
D
Which is problematic.
- [50:34]
A
Yeah. Yes. Incredibly scary. Yes.
- [50:37]
B
And you mentioned in the articles we kind of talked about AI, one of the other challenges, and you sort of mentioned it while we were talking too. The idea that there aren't really AI ethics exactly. Like there are computer ethics people who've been studying some of these issues. But there's no. I think, I think you mentioned the AI ethics have no teeth. You know, it's not, I mean even just looking at our code, there's no AI statement in there at all. And if there were, probably something like please don't use AI to, you know, lie and generate reports or something, you know, like, like you get a lot of schools, you know, starting to write those in there. I mean, when you think of an AI ethics body or an AI ethics rule set or list commandments, what are some of the ones that you would probably push the most for? Like these are, these are going to be key to AI in the future. Ethics in the future.
- [51:23]
A
I mean, I think the theme, and again that's kind of maybe highlights my bias in some of my research really goes back to that explainability and transparency. I think those regardless of, regardless of the area, healthcare, finance, education, grocery stores, those seem like things that should be present everywhere. And I think in particular, because a lot of the end users of these systems kind of like in a doctor patient relationship, I can't evaluate the expert knowledge of this system. I have to trust it to some extent. So like help me trust your system with transparency, explainability, things like that. Same way like when I go to the doctor, I can't really push against them when they're interpreting my blood work as like, oh, you have this thing or not? Or certainly I can kind of trust them and they can hopefully explain it to me. So those are the big ones there. And then I think after that the. It kind of then becomes kind of domain specific. Right. So healthcare, we have a pretty robust framework for how to think about ethics and clinical decision making. Same in education, same in finance, same in law, stuff like that. And there's a little bit in some of these fields around how can we ethically think about tools generally what we don't have in any really system and I think this will end up being like a discipline by discipline development is how do we handle tools of this impact that can make decisions. In one of our studies we had ChatGPT take the BDS exam, passed. So how do we have these systems that now can make decisions at the level of A bcba, like, again, asked in a particular way. So it's not. You talk about Turing Tests, it's not like a BCBA walking around, very limited. You have to prompt it in the right way. But these are real systems and they're going to get better. These are the stupidest AI systems we're ever going to interact with in our life. And those are the, I think, going back, like, the starting the conversation. Those are the conversations that, as a field, as a collective, because I don't think any one of us can say, like, this is the rule you all have to follow, I guess, except the bacd, because they literally can do that. But, yeah, ethics generally, I think about as more of a profession group decision. Right. So we have to. We have to talk about this. How are you using these tools? Where has it helped you? Where have you kind of run into some oopsies, you know, and it's okay. We're also trying to figure it out. Let's just be open and honest about what we're doing. And then, yeah, it's only going to get better. So how do we try to stay on this path, going back to rapid development? Another kind of related example, we sent one paper in around this kind of AI system that we had created. We got the reviews back in, like, 30 days, which is incredibly fast for academic publishing. By the time we got the reviews back, we were already two iterations beyond the system that was reviewed. So they were asking us to make these changes. And I was like, I don't. Like, do I do that? Because by the time I then submit it and it comes out Now, I'm like, 10 iterations on. I talked about the new one. Like, I don't know. So I just gave up. I just didn't start recently. Like, I don't know, whatever.
- [54:07]
D
Liked that footnote. That was good.
- [54:09]
A
Yeah. Yeah.
- [54:11]
D
By the time this thing's finally published.
- [54:13]
A
Yeah. And it kind of kind of goes back to this conversation because it's like, not only do we have to talk about how we're going to use these tools, but also not about, like, a specific tool because it's changing so rapid. It's like, how do we come up with our essay, our methodology for handling these things that we can generalize kind of over time? It's crazy. It's incredibly exciting. So just crazy to think about.
- [54:38]
C
I find it terrifying.
- [54:41]
A
Yeah. I get terrified a few times a week or so, you know, I don't know. Yeah. Hold on and ride the wave again.
- [54:49]
D
Our colleague, Dr. Alan Carcina, is very into AI. I don't know if this has already come up, Jackie.
- [54:55]
C
It has.
- [54:56]
D
Okay. Sorry, I'm jumping in late here. But he asked. He's played around with it a lot of different ways, but he asked it about like I told them.
- [55:05]
C
I told him about the script already and the super.
- [55:07]
D
I wasn't gonna talk about that though.
- [55:09]
A
Okay.
- [55:10]
D
He asked it like, what are your concerns with behavior analysis? Or like, what did he ask it like? Talk. Talk to me about behavior analysis through like a compassionate framework or something like that. And it gave back a very good, detailed answer that was complex.
- [55:30]
A
It was.
- [55:30]
D
It hit on a lot of really important nuanced points about the field in the larger context of the current agenda. Talked about cultural responsiveness, compassionate care, neurodiverse practices, etc, in such a way that, you know, I, I was impressed that it could come back with all of that. Which only means, first of all, that it's been fed good content. Right. By very, very smart behavior analysts. But I do see that as a positive. Right. Because, you know, we're here trying to disseminate this type of information and you are as well. But not everybody has necessarily access that. And if you can have something like, you know, a conversation partner in. In chat gp, where. Chat. Chat GPT, where you can ask it a question like that and get a response back. Right. That does hit on some of these, you know, much harder, I think, to pinpoint answers. That's surprising to me that. That the robot can give you that back. Right. But in. In a way that is encouraging in terms of ensuring that that type of content is potentially getting to a wider audience. So I also am extremely wary of this whole area, but I found that to be kind of hopeful.
- [56:50]
A
I love that. And I agree. I do think there's a tremendous number of positive ways that AI can impact us. And in particular, if the companies creating the systems can validate on their side or control the accuracy of that output, then it also becomes incredibly easy to try. And I can say, hey, Supervisee. Yeah, Chad with chat GPT about it, because you could also imagine another one asking the same system. I want you to tell me the top 10 reasons why compassionate care should not be followed or something like that.
- [57:21]
D
Yeah.
- [57:21]
A
And it'll kick out a convincingly in the opposite direction. And so those are the things like, you know, tell me why the earth is flat or something like that. Oh my goodness. Then I'll give you the. So those are the things that worry me. But I agree, I mean, my career has pivoted all to AI. And behavioral health. I'm obviously a fan. I think it's going to do a lot of really wonderful things for our field.
- [57:42]
B
So with that optimism, we'll do the optimistic dissemination station and kind of start getting some summary points and looking into the future. We should probably be on a more. Do a futuristic chat. Jackie T. Do a futuristic train sound.
- [58:01]
A
Yep.
- [58:01]
B
That's the sound of the future train.
- [58:04]
C
Because people are dancing on it. It's like a dance train.
- [58:08]
B
Why are they dancing on the train?
- [58:10]
C
It's a dance train.
- [58:11]
B
What is this science fiction world you've created with your AI train? It's really drawing me in.
- [58:18]
C
People love to go to work on the AI train.
- [58:23]
D
Amazing.
- [58:24]
B
So, David, we look five, maybe 10 years down the road. You're thinking about how AI has changed the field of ABA. Like, what would you see as realistic? Ish. Because I know you can't, you know, 100 predict what the future will hold. Lives of the, of the bcba, of the ABA clinic. You know, what are the areas where the jobs have gone? We don't have that job anymore. Instead, we've moved this person into a different job because AI handles that now. Like, where do you see those being kind of the big, the, the, the, the big moves ahead with AI, specifically in our field?
- [58:57]
A
Yeah. So there's, I think, a lot of those tedious, repetitive admin tasks gone, or at least very light. And how a lot of these systems have been implemented elsewhere is that it's not as though humans aren't involved. It's the AI handles the 99% that are obvious and it kicks the 1% to humans. So it kind of just reduces that admin time. I think that's going to be one big area. It gives us all kind of time back in our life. Another big area that a lot of people are talking about is this idea of automating data collection in many different ways. Whether that'll be fully solved or partially solved. I think we'll have a lot of incredible tools that kind of do that work for us. And then where I also get excited and this is kind of where most of my kind of work falls, getting into the complexity of human behavior. I think back when I was practicing and it was like program by program, graph by graph, and I kind of analyze, make a decision or whatever. I never looked at that thing in total. Right. The complex interaction of all the programs I'm creating, how all the different graph boards are kind of playing off each other and things like that difference across school, home and AI will start pulling out those patterns. And I think nudging us, offering us advice on clinical decisions that we didn't even know were possible before. And then where this, I think, gets interesting is that it allows then behavior analysts, the phrase is like, to practice at the top of their license. Right. So it's kind of, again, going the same with, like, this admin, admin idea. There's a lot of probably obvious rote decisions that could be made by these systems. And then that gives us the time to focus on what are those really complex cases and how can we pull in different information from different locations. I don't think we'll get rid of the RBT in the next 5, 10 years. I think we still need a human delivering those services. I know there's a few people playing around with robots to deliver therapy. That seems incredibly unlikely. And at the end of the day, we're trying to teach humans how to interact with humans. So it seems counterintuitive to them to put a bot in there. But maybe, maybe I'm wrong. I don't know. That'll be there. And then I think the big thing that I hope this all solves really, is that access point. We know there's a lot of kids that don't get therapy. Rbts can't solve that per se. But if I can scale up a BCBA so they can effectively ethically handle 50 clients because there's a lie, can just handle more information, making decisions more quickly and whatnot, then you can maybe tackle some of that access issue. But I think that's realistic. Five, five, ten years, maybe.
- [61:13]
B
Oh, that's exciting. So, you know, David, for folks who are either saying, yes, I love AI, can't wait to get involved, maybe not so much folks who are like, nope, that's enough. I don't need AI in my life because I think it's probably coming to some extent, whether you want it to or not. How do we, as clinicians who have varying levels of computer knowledge, get involved in AI? And you certainly, I know you and Adrian put out kind of like three tiers, which I found very helpful in the paper. So you can certainly cite those or, you know, if you've had other thoughts on it since. Since writing that paper. But, like, what are the ways we get involved?
- [61:50]
A
Yeah, no, I think that that framework still. Still kind of holds because I think there's some people that hear this that are like, dope AI, that's my career. They're gonna need to learn how to code. Like, that's what they're gonna that's just gonna be their path in life. Amazing. They're learn math and like computer science and things like that. There's another group of people, probably the majority that's this is cool. No chance am I gonna learn how to code, like whatever. And I think those are gonna be the individuals where we need them to start seeing the suite of tools that are available and it's almost like applied research. Right. Let me start bringing stuff in, let me see how it works, where does it fail, let me give that feedback back to researchers, things like that. But they'll become experts in how to use the tools that exist. And I think that's probably going to be the majority of people. And then there may be some people that are kind of in the middle that maybe want to do beta testing or something like that. And they're really on like that cutting edge and want to get involved, but they don't really want to know how to code. I think that's a smaller set of individuals, probably harder to do because you have to connect with researchers or tech companies that are building these things, but certainly a viable job path. I have seen a number of these tech companies building tools for the ABA space and they hire behavior analysts because they know the field, they know how these systems work. That can be a, perhaps an alternative career path. But yeah, starting to try these tools, seeing how they do and don't work, learning about I think again the suite of types of AI. I think the world right now is heavily focused on these large language models ChatGPTs. But I think the real power than some of these other types of artificial intelligence systems that people are still using, they've been having a huge impact on medicine for decades now. Starting to have a big impact in aba. But I think those are going to be the ones where it's kind of like the example before. Like I need something for scheduling. Right. Or I need something for X or Y or Z and finding what are the tools that do that for me.
- [63:41]
B
Excellent. And if someone were to say, no, I don't trust AI, I don't think I'll ever use it. And you wanted to say here's one way, try this tomorrow and then come back and let me know if, if you're warming on the subject. What, what would be the, the, the program or the usage case that you might, might recommend for the, the AI, I guess. I don't know.
- [64:00]
A
Yeah, yeah, that's a good. So I think the one, the one tool that I recommended that seems to land well with everybody is a, A Product called Motion M O T I O N. It's a very simple. It's a calendar calendar app. It integrates multiple different calendars. I have four. Right. So just kind of pulls them all together. But what's amazing about it is that when I move right, something pops up. I have to move something to my schedule. It automatically rearranges the rest of my calendar and it takes. I integrate it with Asana, but you can put tasks right into it. And then it also blocks off chunks in my day to make sure I get my tasks done by their due date. So people can't, like, schedule meetings over it. It's a very simple. But it's using AI to similarly, you know, I move a meeting, I'd lose 20 minutes just trying to figure out when the hell am I going to do everything else. And this just handles it and it's incredibly useful. And, yeah, not scary. It's not. If it makes a mistake, it's not going to kill anybody. And you can see it like, I can just move the schedule if I don't like it. Right.
- [65:00]
D
You had me in scheduling.
- [65:03]
A
Love it.
- [65:04]
B
Oh, well, Dr. David Cox, thank you so much for coming on the show and talking about this very broad and fast moving. I'm sure while we were recording, there's two new AI algorithms that we missed.
- [65:16]
A
The alerts come in.
- [65:19]
B
If folks would like to reach out. I mean, certainly reading, reading the papers more at length because we didn't get into every single piece of information because this is a very broad and large topic. I think those papers are really an excellent, excellent source of information. Good summaries. I appreciate that some of the citations you had, I think you had one that was like, here's some citations. Their books about computer science or AI, that was a little scary. So I'm glad. So if. If someone wants to reach out and learn more about AI, but maybe doesn't have the time for a computer science degree or to read a whole book on it and they just want some. Some advice from someone who's been studying the field for so long. Is there a place folks can reach you?
- [65:56]
A
Yeah, definitely. So people can find me on LinkedIn. That might be the easiest spot. And then I'm also faculty at Endicott College, so they can find my email address there on the faculty webpage. And then I'm. I also work for Rethink first and they can find my email information there as well. So any of those three will be easy enough.
- [66:14]
B
Excellent. All right, big thanks to Dr. David Cox for joining us on the show to talk all about artificial intelligence, but my new AI calendar has told me there's one more part of the show we're going to do, and that is pairings. Diana, take it away.
- [66:30]
D
All right, everybody, it's time for pairings, which is the part of the show where I tell you about past episodes. You might want to check out if you enjoyed the current conversation. So there were a few times in the past where we have talked about varying aspects of technology as it relates to our field. And I want to tell you what those episodes are. They go all the way Back to episode 15 where we talked about technology and safety skills with Dr. Nick Vanlo. Times we talked about. We talked about virtual reality, episode 25, and then also episode 93. And for that episode, we were joined by Doctors Berglin, Sinbjorn's daughter, and Casey Clay. Episode 188. No, I'm sorry. Episode 88, where we talked about ethics of telehealth. Episode 211, we talked about variety in ABA with Dr. Matt Normandy. And then episode 224, we talked about teleconsultation with Dr. Aaron Fisher. So you can enjoy going back and checking out all of those.
- [67:35]
B
That was the closest one. I remember we talked about the AI Aaron talked about the AI robot that would kind of go around and look in the windows of the iPad. I think that was the closest to what David was saying about, you know, could. Could you have a therapist be a robot? Yeah, that was the closest.
- [67:51]
D
And then also, I like to recommend a snack to go with this episode. So I'm going to recommend two things. The first one is nuts, and then the next one is some cookie dough bites. But bites is spelled B Y T E S, so you can imagine why.
- [68:08]
A
Please.
- [68:08]
D
Clever.
- [68:09]
C
Clever.
- [68:10]
B
Is it European? What? I know.
- [68:15]
A
Sorry.
- [68:16]
B
Anyway, thanks everyone so much for listening to ABA Inside Track. Please, if you like the show, leave us a review on Apple Podcasts, Spotify, or wherever you like to get your podcast. Feel free to subscribe as well. You can find us online as ABA Inside Track. You can go to our website, abainsidetrack.com to find links to all the articles that we discussed in this episode as well as all of our previous episodes. You can also find a place to purchase ces. Speaking of purchasing ces, you probably want the second secret code word. It is overlord. O V E R L O R D. Overlord, like the leader or something that's over you and in charge. Think 1984 kind of thing if you'd like. Even more ABA Inside Track content. Why not check us out on patreon@patreon.com Aba InsideTrack where you can subscribe at any level to get our episodes a week ahead of time. But if you subscribe at the $5 levels, you can join our quarterly listener choice episodes and get a free CE. And if you want all the ABA InsideTrack content, joining at the $10 level will get you access to our full length book clubs a full year before everybody else, as well as 2 CES at no additional cost.
- [69:32]
D
Full year before we recorded them?
- [69:34]
B
No, we, we recorded them and then you get them early before everybody else. And that will give you a chance to hear our most recent book club on the Science of Consequences by Dr. Susan Schneider.
- [69:45]
D
Sometimes we do the book club before the book's even written.
- [69:48]
A
That's.
- [69:48]
B
We're not that fast. That's. That's the AI book. Yeah, I want AI AI. Read this book for me and do a podcast. Save me some time. A lot of time.
- [69:56]
D
You're welcome, Robert.
- [69:57]
B
Again, that's patreon.com ABA Inside Track some final thanks. Thanks again to our special guest, Dr. David Cox. Thanks to Dr. Dim Carr for recording our intro and outro music, Kyle Sturry for interstitial music, and Dan Thabit of the podcast Doctors for his amazing editing work. We'll be back next week with another fun filled episode, but until then, keep responding.
- [70:18]
A
Bye bye bye Sam.