Episode 01: Will AI Replace Therapists?
Dr. Jonathan Gratch
Illustration by
Alexander Bustamante
In the first episode of Heal With It, Maytal sits down with Dr. Jonathan Gratch, the associate director for virtual humans research at the USC Institute for Creative Technologies, a research professor of computer science and psychology, and the co-director of USC’s Computational Emotion Group. Dr. Gratch’s work involves developing artificially intelligent virtual humans, like Ellie, who can listen to patients talk about their experiences with PTSD. In this electrifying discussion about the future of mental health, Maytal gets to ask the burning question: will AI inevitably replace therapists?
TRANSCRIPT:
Maytal: Hi, and welcome to Heal With It, a podcast about healing in its many, and sometimes unexpected forms.
[Music Interlude]
Maytal: In 2021, things like robots, AI, and automation, no longer feel like illusive science fiction phenomena. These technologies are here, and they’re very much here to stay.
[News Sound Clip: “Tesla CEO Elon Musk announced he’s building a humanoid robot] [Soundbite: Robot noises)
[News Sound Clip: “A robot dog that’s hounding city streets.]
Maytal: Take robots, for example. They’re delivering our pizza on the streets. They’re helping elderly folks move from their wheelchairs to their beds. They’re performing life saving surgeries, and singing duets with Jimmy Fallon on TV [music clip: Jimmy Fallon and robot singing, “Say something I’ve giving up on me”]. If you do a quick YouTube search of the term robot, and you will be inundated with videos of impressive, and strangely cute metal machines with names like Spot and Atlas.
It seems inevitable at this point that as time moves forward, these technologies will become more and more integrated into our daily lives. And not only will they become more integrated, but in some instances, they may even start replacing the need for humans altogether.
For example, multiple reports point to professions like truck driving, customer service, and retail on the inevitable path towards automation. But my curiosity, is about these technologies’ capacity for emotional intelligence. Can these technologies become our friends, our partners, what about our therapists? Most directly, what I’m wondering, is can these technologies heal us?
These are some of the questions I got to ask Dr. Jonathan Gratch today. Dr. Gratch is the director of the Virtual Humans Project at the University of Southern California. He’s a research professor of computer science and psychology, and the co-director of USC’s Computational Emotion Group. I learned about Dr. Gratch by way of learning about his creation, Ellie, a virtual human designed to interact with people and make them comfortable enough to open up about mental health issues like PTSD.
Our conversation today takes a lot of different paths, and brushes upon topics like: therapy becoming automated, AI becoming a helpful hand to humanity, and what it might mean for our collective mental health to live in a world that is becoming increasingly more digital, and far less tangible.
Without further ado, I bring you Dr. Jonathan Gratch.
Maytal: Well, thanks for being here today. Really excited to get started. And I figured we should just launch right in there. Um, so I'm really fascinated by the virtual human project and from my understanding, it's basically training synthetic humans to emotionally interact with people. So, first of all, I'm wondering if that's an accurate understanding of it. And second, I'm curious to learn more about the project and all of its applications.
Dr. Jonathan Gratch: The general focus of the research is, you know, how can we create interactive characters that can have conversations with people, but also understand and respond to their emotions. And then also more from a psychological point of view, you know, how do people interact with these kinds of sort of human, like things, you know, do they treat them like people, they treat them differently than people, and if they treat them differently, you know, is that a good thing or maybe a bad thing? And if it's a good thing, how can we take advantage of that for different kinds of applications?
Maytal: I read something about how anthropomorphizing these machines can have positive consequences and then negative consequences. So I'm curious about what those positive and negative consequences are.
Jonathan: It seems like people sort of treat machines like people. And the more that those machines both look and act like people, the stronger those effects are. And so those are sometimes called social effects in psychology. So an example would be when I'm talking to a person, I want that person to like me. So I may engage in what's called impression management to convey a good impression. And that might even mean I'll be less honest about myself. I'll smile more than maybe I feel, and so forth. Those can also have negative effects. People often have implicit biases. They treat in group members better than out-group members. And the same can hold for computer that looks like an in-group member or an out-group member. And so just, you know, how people treat each other, there's good and bad things about how we treat each other. There's also sometimes a problem if there's a disconnect between what the character looks like and what it can do.
So if the character is conveying, it's a person, but doesn't actually have the AI behind it to carry out that charade, then you're basically creating false affordances. You're telling somebody something about this thing, which is not true. And sometimes that can undermine the quality or perceptions about the interaction. Like for example, if a machine says, it's sorry, but it's not sorry. Often when people say, they're sorry, they're communicating that they understood they did something wrong. They'll try to do something better. And sometimes people just sort of paint those social niceties on a computer program, like lipstick on a pig. And if it doesn't actually enact that actual function over time, you'll realize, okay, that doesn't mean anything and you'll start to tune it out.
Maytal: That's so fascinating. How do you make it so I would think a machine is genuinely sorry?
Jonathan: At the end of the line, it really depends on does it actually change his behavior? So you could imagine a GPS system of the future takes you down a road that's not safe for your car, point that out, it says, sorry, and then in the future, doesn't take you down that road. That would actually mean enacting that function. Similarly, often we use emotional expressions to achieve a certain function. Like I may look confused and then you might recognize I'm confused and change the way you're explaining something, or I might look angry, and you’ll realize, something's wrong with the relationship and you'll try to change that relationship. And so, you know, if those computers can recognize those expressions and actually do something about it - they actually change the way they're talking or the way they're acting based on that kind of moment to moment feedback, then you'll begin to realize, you know, it's actually kind of like a person is actually responding. Like people respond and then you'll treat it in a more natural way. But if they're just randomly nodding at certain points, then you'll just sort of tune that behavior out.
Maytal: So behavior change almost elicits a sense of trust.
Jonathan: Yeah, exactly. I mean, people have to calibrate, should I use this? Should I trust this? And it's a continuous judgment. If the behavior matches the expectations, then you'll have a better sense of what this thing really does.
Maytal: Super interesting. So I'm thinking about all of this in the context of mental health, psychology - the focus of this podcast is looking at technologies innovations programs that can revolutionize or transform the field as we know it right now. And so I think where I want to start is I want to ask about Ellie, um, who is Ellie? How did she come about, what is she capable of?
Jonathan: Yeah, so Ellie, uh, and I guess the scientific name we call it, the SIM sensei system. Ellie is the name of a character. If it came out of an effort to find technological solutions, to address reluctance of particularly soldiers, but in general, people who have PTSD, uh, in seeking help. So often people with mental health issues are reluctant to seek help. And so could we find technology that might address that? And there was actually a number of different teams exploring different solutions. Our solution was to focus on these virtual humans. Part of the reason we took this approach is that we thought that these characters might have some unique characteristics that could address that problem. People are reluctant to seek help, because when you're talking with the person you want to give a good impression, you don't want to be negatively judged. When you have a disability or, or mental illness, you may have, what's called fear of negative evaluation.
You think that the person you're interacting with will think badly of you and, and some prior research has suggested that people don't feel that fear when they interact with a computer, a non-anthropomorphic computer. And that's why often certain sensitive questionnaires are delivered by say a computer web form rather than a person, uh, because people feel they're less watched. They're more comfortable revealing truthfully, you know, are they having nightmares? Are they feeling that they can't cope just on a form? So computerness, in a sense, can get past fear of negative evaluation. But on the other hand, one thing that people are really good at is establishing rapport, making a connection with people and using language and nonverbal feedback to draw people out. And some of my research before the Ellie project was basically teaching computers to do that, you know, getting them to nod and smile at the right point to show concern and draw people out.
So there's sort of two pathways to get to honest disclosure about mental illness. One is anonymity which computers can help with and the other is rapport, which usually only people can do. And so we were curious, you know, could we somehow get the best of both worlds or might it be the worst of both worlds? And so fortunately it turns out that it seems like it’s the best of both worlds in the sense that we've been able to find that people feel drawn out and comfortable talking to Ellie because of her rapport building abilities. And yet they don't seem to fear being judged as much by her because she's a computer as they might if they were talking to a person. So this character is able to make people feel more comfortable and more willing to disclose evidence of a post-traumatic stress disorder than if they were having an interview with a person, say, over web chat, like we're doing right now.
Maytal: Could you walk me through, if I'm a soldier experiencing PTSD, I'm interacting with Ellie, what is that going to feel like? What is that going to look like?
Jonathan: Sure. So you come in, there's essentially like a kiosk and you'll see a 3D computer generated character, a woman named Ellie sitting in a chair, much like you might do if you walked into a clinician's office, the character starts to engage you in conversation. And the way it's structured is first, it just engages in social chit-chat. You know, where are you from some of the, about yourself that are considered safe and then begins to transition, essentially, following a script, much like a clinician might do the following, say that physician's checklist of trying to elicit symptoms of depression and PTSD, but asking a series of questions. And as those questions are answered, the character is trying to convey that it's listening attentively so it, much as you're doing now, sort of nods at the right points, tries to wait the right amount of time before taking the turn and interjecting in.
If the person talks too briefly tries to ask some follow-up questions, to draw them out more, and then kind of moves down the checklist to go through a series of questions. And then at the end tries to go back to more softball questions to kind of lead the person back into the more positive state cause sometimes people, you know, get quite emotional when they're talking about their mental illness. It's not a super smart conversation. The character doesn't deeply understand what you're saying, and so sometimes can misunderstand or ask the wrong questions. And we try to, you know, set that expectation up front. When we introduce people to the system, generally people seem to really like the interaction and, you know, they'll talk easily for half an hour. And some of the studies, we actually had to kick people out of the room [laughter] because they just kept talking and anecdotally they would report things like it was just great to have some, somebody that they actually referred to her more as a person. Um, but somebody listened to me in a non-judgemental way. It seemed to serve its purpose in drawing people out.
Maytal: What's really striking me is this notion that these virtual humans like Ellie are less judgmental than a real human. So there's this comfort, there's this willingness, maybe, to be even more vulnerable than you would with an actual clinician. It just makes me think about the implications. And I can see right now how this would be - like, I would love this technology in my work. I mean, I think of all the intakes and assessments I do, it would be so helpful to have this technology doing that, but I can't help but think about the next steps and the future implications Could something Like Ellie evolve into doing the work that a human therapist does the beyond the mental health screening the week by week building emotional rapport and so on.
Jonathan: Well, first of all, the current state of the technology is most limited in sort of understanding the nuance of language. And so it's good at eliciting language from a person and doing some shallow analysis to suggest there may be some issues and then you really have to call on a person that point to do a more thorough analysis of the content. What I think is possible is to basically use this as a sort of glorified self-help book. So, you know, there's lots of, you know, workbooks for say working through cognitive behavioral therapy, where you are writing down your thoughts and then there's, you know, text written questions about those thoughts that trigger you to think about that. Um, and there's some evidence-based research to suggest that alone can have certain benefits, but in terms of what you would experience in a real cognitive behavioral therapy session, where the clinician is actually challenging, specific thoughts you might have and get you to, you know, think of different ways to, to frame that situation, that sort of meaning work is I think very challenging and fraught. And I don't expect it in anytime in the near to mid future, that language technology will be up to the task of really helping people work through the specific meaning issues that they're bringing up in a session. So I would think it's more appropriate to think of these things is, is really kind of glorified workbooks [laughter] that have some additional stuff to help draw you out and get you to engage and motivated to do that task. But they're not any much smarter than that, that book.
Maytal: Which I think that in alone is incredible. But what does mid to near future mean? What timeline is that?
Jonathan: Yeah, I mean, I’m very hesitant to make predictions. I wouldn't expect you to - and there's of course, ethical privacy issues around this too, that have to be worked out. But I wouldn't see in the next 10 years, you having a fully automated system that people would really feel comfortable saying, “Hey, this is like a digital clinician.” You know, the technology can definitely show evidence and be useful for mental health, but I wouldn't try to compare it to the important and deep thoughtful work that human clinicians do.
Maytal: It's really interesting because I read a fascinating article about a couple of weeks ago, I think, and it was written by the director of engineering at Google. And he was saying like, look, the year 2020 going into 2021. So many monumental events occurred, so much change occurred. And we think that this is unique, but in reality, this is a demonstration of what life will look like moving forward, because the rate of ideas and technology growing and evolving is growing exponentially. And so he was saying in this article, by the year 2040, by the year 2050, we could be living in a world that looks very different beyond our current understanding. And he really meant this in the realm of technology. So I, I really respect you saying I hesitate to make predictions, but when you think of the world in 2040 or 2050, is there anything that comes to mind where you could see where this technology goes, especially in the context of mental health?
Jonathan: I think we sometimes underestimate the difficulty of certain things that people find effortless, you know, there's enormous explosion, uh, and capability with things like deep learning, with pattern recognition. A lot of what AI has gotten really good at doing is sort of stimulus response kind of behaviors. It detects a pattern and can take an action. And there's a lot of things you can do really well with that, you know, computer games. And in some, some cases, a lot of what's involved in driving a car is this more stimulus response: see a pattern, take an action.
What is really, really difficult and the progress is much, much slower is when you really have to think to strategize, to plan, or even the hardest is to sort of reframe a situation in a new and novel way. Just an example, you know, we do stuff in mental health. You can practice negotiation skills with an agent and some of the same advantages apply there. People don't like to negotiate, they're afraid to negotiate, or they find it really aversive, but they feel more comfortable practicing with an agent. Part of a successful negotiation is to find creative solutions. Just an example, two family members want to make their own recipes in the house and they both need an orange. Turns out there's only one orange left and they think there's a conflict. But if they come to realize, oh, this person wants to appeal because they're trying to make some spice and the other person needs the juice, well now they found a solution sort of a win-win solution. And you can think of a lot of - actually therapy sessions is really the therapist and the patient are negotiating. Meaning what is the significance of this event for you? Are there ways for you to rethink of this event in a, in a new way that makes you feel less debilitated? That creativity aspect is what computers really uh, fail with and you're going to be failing with for quite some time. And I think even 20 years is going to be a challenge. The way the systems work is that humans have to sort of anticipate the representations that the system reasons through and creating new representations that allow creativity aspect is just - we don't know what's going on, how to do that yet. It's something that's bedeviled AI since the beginning.
[Music Interlude: “More with Dr. Gratch after this short break.”]
Maytal: Can you tell me about something that, in your work with Ellie that has surprised you or has felt like a breakthrough?
Jonathan: I don't know if it constitutes a breakthrough. You know, we were really stunned to the level that people open up to the character that, you know, we've had people crying in the office. We've had people, you know, revealing things that are very difficult to listen to. And perhaps that's the part of the advantage of the technology is it doesn't know, you know, how difficult that is to listen to, and maybe that people recognize that and makes it easier for them to relate to it. We thought that perhaps people would find this insulting, you know, like, I've got this serious mental issue and you want me to talk to a puppet? And that was a real fear. And we had concerns about taking on the project to begin with for that reason, because, you know, up to that point, a lot of what we do is, you know, research in the lab, sort of toy problems, uh, cancel and talk to this thing. And here we're addressing a significant societal issue. And so I guess it was striking and rewarding to see that people didn't find it insulting, that they found value to the technology.
Maytal: Yeah, that makes a lot of sense to me that there was this fear that people might be insulted by the notion of speaking to a robot. And I myself am surprised to hear that the opposite happened, that people actually felt more comfortable being vulnerable. That feels like such an advantage of an AI Therapist. And what else, I think, feels like an advantage of an AI therapist, is the fact that they don't have bad days. I mean I think of myself, in my own practice, and sometimes I have bad days. Because I'm human. Because life is hard and it goes up and down. But that's just not the case for AI. They can show up, to every session, with every client, without any sort of human struggle.
Jonathan: Mm, interesting, yeah. I mean generally you probably know the concept of emotional labor that certain jobs in, I assume clinicians, but definitely customer service. People are required to portray a certain kind of emotional state, which they may not necessarily feel. And that discrepancy between what they're supposed to portray and what they feel can lead to long-term stress and burnout and have people quit. And so at some level, yeah, these characters that don't have feelings, so they can't have a discrepancy between what they show and what they feel. On the other hand, you know, there may be, you know, ethical concerns about, you know, does it make sense to have this character that can perfectly show whatever you want it to feel, to influence the customer when it's sort of, in some sense, inauthentic. So we've talked about the benefits of this and the context of mental health, but we've also done studies using essentially the same technology showing that we can get people to be more comfortable, revealing sensitive financial details about their situation. Which is, you know, probably not a good consequence or at least good use of the technology.
You're right that these things can avoid burnout but the flip side of that is in a sense they're in authentic yet. Can't possibly convey that they're authentic depending on the use that might be deemed to be problematic.
Maytal:Is there research out there or people working on this about humans ability to develop attachment of some sort to these machines?
Jonathan: In fact, people have been exploring the long-term consequences of these. We've actually done some research in our lab that looks at how these things in a sense could fulfill social needs of people, such that they're less motivated to interact with real people. People have a need for belonging. It's kind of like a hunger. And when you are separated with people and you had a need for belonging, that motivates you to reach out and connect with other people and connect with friends. And we've, uh, been able to show that if you have a good interaction with a character like Ellie, that satiates your social hunger, uh, you don't feel the need to reach out to other people. And there's the short-term effect cause we didn't study it over time, but there's a sense in which these softwares can become a substitute for social interaction.
In Japan they do a lot of work with these robots and the context of elder care where they use the robot to satisfy people's need for social interaction. And so in their cases, there, there are no people. So I guess some, the robot is better, uh, than a person, but other research, for example, looking at say, Facebook has suggested that Facebook can serve as a way to satisfy your social hunger, which might lead you not to reach out to other people and become perhaps more lonely. So some work has shown that people who spend a lot of time on Facebook tend to be more lonely. Now, the causality is not clear there. Um, but there's some suggestion that if technology becomes intimate in the sense that it satisfies a number of our social needs, might that in a sense kind of serve like a drug, we might turn away from real human relationships. And so there are within some of the professional societies, like IBBB that that tries to help inform our research. There are panels trying to construct ethical guidelines for the design of emotional systems like we've been discussing. And so one set of guidelines is around this idea of intimate technology and, you know, to what extent must that be regulated or what are some of the factors that might lead to unintended, but bad societal consequences and how can we better at least study, uh, possibly address those?
Maytal: That's fantastic that there's these ethical considerations. I have to admit. I'm like a huge sci-fi geek and I just recently read the book, um, by Kazuo Ishiguro the Clare on the sun. I don't know if you've read it yet.
Jonathan: Yeah, yeah.
Maytal: So I don't know, I can't help, but think of a world like that, that he paints of kids hanging out with their AI and that being their closest friend and caretaker and what the implications of that on human connection are. Do you hold any fear around that or concern around how this technology might evolve even to the generation beyond us?
Jonathan: Yeah, yeah, of course. So another sci-fi movie is Her. And so often, when I teach a course on effective computing, which is building emotional machines. And so, it's getting a little dated now, but, you know, I used to make them watch Her and sort of use that as a stepping point for thinking about the ethics of that technology. And there's actually different perspectives. There's, uh, a view of that kind of technology that it could be developed to kind of socially diminish us in the sense that if it gives us just what we want or what we think we want, you know, tries to satisfy our social needs for happiness, that compared to - uh, and you could actually argue what was happening in her was the technology was challenging him in certain ways, not always give him exactly what he wanted, but perhaps give him things that might be forcing him to grow as a person. And so one movie critic was arguing that actually, you know, in sort of classical movie sense, that the technology forced him to confront his personality limitations and grow as a person at the end of the movie, he reconnects with humanity.
You know, I present this, these contrasts, you know, do we give people what they want or do we give people things that challenge people to grow and as ideas for how to think it through. But I think the technology takes a shallow view of trying to sort of localize and maximize our happiness. I think that would be the worst possible outcome. But I think to the extent that technology can be designed, to help us grow, uh, particularly grow in our human social relationships and the way we engage with the world, then it can be a valuable tool. What I worry about is, you know, what are the incentives? You know, so currently, you know, our incentives where you look what's going on with Facebook, it's in the hands of profit maximization, which may not be the tool to lead to the best world. So in that sense, you know, I worry that this kind of technology in the hands of, you know, marketing people may lead to a dystopian future.
Maytal: Yes, I love the way you just laid that out because it made so clear to me what the very positive implications of this technology could be and then what the negative implications could be. And that's exactly where my mind was going. If this technology gets in the hands of people who are only concerned about making money and profiting, then they're incentivized to just give people technology that satiates them, that keeps them wanting more, that keeps them addicted to circumstances as they are. It's, it's scary.
Jonathan: Yep.
Maytal: Are those ethical guidelines you were speaking of before intended to prevent something like that from occurring?
Jonathan: I forget what the context was, but I was walking through those guidelines with the group of people and, you know, they have the intention to prevent that from happening. But a lot of the guidelines, I think, speak to things, you know, for example, that advertisers are trying to do today or doing today without - you know, a lot of them says things like people need to opt in to any form of manipulation. And, you know, if I turn on a TV, I'm getting in any commercial, some form of emotional manipulation to try to get me to vote for a candidate or buy this particular product. And there's no - I don't know to what extent there are, besides, you know, some requirements of being minimally truthful. You know, I don't know that there is much to prevent that kind of manipulation. So it's not obvious that just because I had this new technology, that's going to change the status quo. So I think that the intent, the ethical guidelines are well-intentioned, whether those guidelines can actually affect change, uh, I'm not so sure.
Maytal: Where would you like to see the future of this technology go primarily in the context of mental health? I feel like you hold such an expertise in this intersection between psychology and technology. So what would you hope for?
Jonathan: Several things. One is I hope, I mean, there's kind of two different metaphors for how AI is considered to be deployed. One is really as a replacement - replace cab drivers, replace GrubHub people, you know, replace, uh, lawyers with software. There's another view of AI as a partner. There’s even a term called Sentar. And so like when people play chess, in the world of chess, they've been shown that, you know, a computer program working with humans together and partnership can be the grandmaster. What I hope to see is that we deploy this technology in partnership with people so that the technology can emphasize, you know, things that, that machines are good at - memorizing things, recording, pattern recognition, data analysis - and the people use that as a tool to augment their own ability, rather than seeing Ellie as something that replaces a clinician. You know, maybe at best it should be viewed as an initial pre-screening system that can help then, you know, bring a clinician up to speed in general. I hope that we see these techniques as - and conceive of this software as - tools to augment human ability, rather than a way to replace it.
Maytal: I like the way you laid that out. And it kind of is helpful for me. It's a paradigm shift because sometimes I sit around fearing this idea of replacement, but I like this idea of partnership. Final question. If people listening to the podcast want to learn more about, you want to learn more about your work, where would you direct them? Is there a specific publication or website for my specific
Jonathan: For my specific work, you know, if you just Google Jonathan Gratch, you'll find our lab's research and we have a group, The Ethic of Computing Group. But I mean, specifically about emotion and technology, you know, there are conferences and journals like the Effective Computing Conference, Effective Computing Journal. You know, I always tell students, read abstracts, go to the conferences, and read the abstracts and that'll tell you, you know, what's hot and what's the new topic in the field. So all those things.
Maytal: Awesome. Well, thank you so much. This conversation was very stimulating. I have so many thoughts and ideas for now other podcast episodes I want to do, including what you were talking about in Japan with the elderly folks and AI helping them. So thanks for being here. Thanks for the great conversation. I really appreciate it.
Jonathan: Thank you.