Integrating AI Responsibly: Setting Guardrails for Mental Health Tech

Richard Zwicky: Hi everyone, it's Richard Zwicky with High on Healthy, the show where we know there's no one path to healthy, so we explore them all. If you haven't subscribed yet, please do so on your favorite podcast listening tool. So today's discussion is about AI and AI is a topic on the tips of everybody's tongues. It's been rapidly embedding itself into every facet of our life whether we realize it or not and definitely into our health care and especially Our mental health care systems are taxed, overstretched and can't reach everybody and people need to turn somewhere and one of the places they turn is AI. And while it can offer support between sessions and tools on demand, it isn't always good. There aren't the guardrails put in place that necessarily mean it's going to give you the information you need. It's an information exchange and that information is subjective. It's based on your thoughts and your perspective and it's often reinforcing and mirroring you what you need or want to hear. Joining me today for the discussion is Erin Pash who's a licensed therapist and founder and CEO of Pashkco and she's working at the intersection of clinical care and emerging technology helping both practitioners and clients understand how to responsibly. Welcome aboard the show Erin.

Erin Pash: Thank you for having me I'm excited.

Richard Zwicky: So Erin, in years ago, I built the foundations of behavioral targeting, which everybody uses in every system. And lately, I've been focused on building a predictive analytics system, which helps people understand who they are. But to crunch the data today, you have to use AI. You can't just do it without. There's good and there's bad. And everybody says AI won't replace therapists. But will it reshape the way therapists and therapy engages with us?

Erin Pash: I think so. I mean, I think of AI the way I think of, history class from high school. There's these different industrial revolutions and everybody thinks it's alarming and it's going to redo everything we do. And it does. I mean, think about introducing the car and all and the computer and all of these things. But the best part about human beings is our adaptability. So when I think about AI and therapy, I look at it from a perspective and a lens of how can we make the human experience even better? How can we handle mental health even better versus think of it as are we competing for a technology of the past versus a technology of the future kind of thing.

Richard Zwicky: And there's a lot of fear around AI and but people you read back and you look back and you hear stories. People had the same issues when the desktop computer started invading the offices. We used to have pools of accountants and bookkeepers and secretarial pools and you name it. And all these people found something new to do and became incredibly creative in their own right and there's an economic boom. I look at AI as allowing us that same sort of transformation. We just don't know where it's going to go. But when we look at therapy in particular, have you seen cases where it's already being used to enhance therapy? Not just make it more convenient, but help people deepen the work.

Erin Pash: I mean there are some therapists out there, the adopters I would say, there's two schools of people who are like, "Yes, this is great. I can use this as a tool to help my clients get better." And then there's the people who are I wouldn't say they outwardly reject it. I think they just don't know how, the existential crisis of am I going to solve a job, is kind of how they feel about it. But the ones who are adopters, I have found some really cool software that integrates into the electronic health record that can help you better track progress looking at behavioral patterns. I mean, it's certainly things that we can do as a therapist from session to session, but notes and analyzing notes so that you can pull out very specific patterns, especially for long-term clients where maybe they have the certain high highs and low lows at certain times a year, certain patterns. It can help us recommend if this is happening the same time every year and we're working on cognitive behavioral therapy, but maybe this is attached to a trauma, maybe I can refer them for doing some deeper trauma work because I'm seeing a recurrent pattern and analyzing data. So, there are certainly people who are using it and those are the ways I see it being really healthy in integrating into therapy and mental health. I would be lying if I didn't say that I don't sometimes use it as something I can bounce an idea off of, almost like a clinical consultation. My favorite AI to use is Claude. I feel like there's just a little bit richer academic background and actually even more reps if I've challenged it quite a bit to be pushing it above its boundaries and I feel like Claude knows the boundaries a lot better than some of the other AI tools out there.

00:05:00

Erin Pash: So, I think it really depends on the person not only the client sitting in front of you, but the therapist and the style of therapy they do to figure out how you can really use it to benefit what it is you're trying to accomplish and how it could also be used as a detriment.

Richard Zwicky: No, it can be used in those ways and claude is really good in terms of it tends to have the best output as far as written language and explaining things and taking concepts and putting them to terms that we can all understand. I also get frustrated because it does have that self-confirming bias where it tells individuals, you're right and you're smart and good catch where you're bringing something up, which can be problematic because then an individual can go down the wrong path because the system will confirm their own bias and their own interests because it's obvious in its own way, which isn't helpful. But the technologies have decided and the drivers have decided that they can get better market penetration by telling people what they want to hear a lot of the time or leading them down that path. And that's a challenge in therapy where it might interfere or dilute the process or not always conflict but if you second guess it may confirm that second guessing when it's going the wrong way. How do you help people know what not to do?

Erin Pash: I think therapists have kind of created this message of just kind of generally rejecting AI as a part of therapy. And again, I think it's about the fear of the unknown, not because it's innately bad. And so I think what we need to do as a mental health cohort of people is widely let people know it's accepted because what we need is for our clients to bring it to us so that if they're using this tool and this tool is leading them down a path, we can help them as a trusted external supporter interfere and help them get on the right path. And so, I think that the only way therapists can help people is by making sure that the people feel like they can bring their honesty and their truth to a therapy session. And so, that's the only power we have. And so, if you're working with an AI that has a lot of confirmation bias or is truly leading you down the wrong path based on, your two o'clock schemas, two o'clock in the morning schemas of, self-destruction, we can only be as helpful as what you bring to us.

Richard Zwicky: And I mean that's part of the challenge. And I'm a big supporter of AI, but it's one of, I write when I write prompts to do analysis some of my prompts are 15,000 20,000 characters for an output that is supposed to be 6 to800 because it's putting in place the right guard rails to make sure the system knows how and what to analyze. And that's not something that commonly people do. It requires a lot of training to understand how and what to do to drive the outcomes towards a clinically objective or a standardized process. And it always strikes me with the output and where we have to train AI therapy is partially about teaching us how to experience and live with and tolerate comfort and discomfort and uncertainty. AI is not about that. It' actually does the opposite in so many times. How does that, and it's also funny the therapists themselves where they say don't use AI. It's like telling a child don't go on the internet. It's going to happen. Don't go on social media.

Erin Pash: It has.

Richard Zwicky: They're just going to change their age and use it, right? People will go and access the information. They won't. So, that discomfort is all around. How do we use it without shortcircuiting the actually good part of the process which is to help people understand and feel that discomfort is normal but it's also part of the healthy process of healing. Right?

Erin Pash: I don't think I have all the answers on that, but I often sit and dream about what the perfect combination could look like. I sometimes think about what if the collaboration between therapy and AI was that the therapist could be part of that prompt or that those guardrails I know this client, I know this person, but I can't be there on the moment in a moment's notice AI can. So, how can I help make this tool be effective if they know they're going to use it? Kind of like using a internet managing app instead of telling your kids they can't putting guard rails, right? We just talked about. So, I fantasize a lot about what if there was a way for the AI to read the therapist notes and then create a system of guard rails and the therapist then gets a say of "Yep, those are the right guardrails I want for this client based on my relationship." And then it's like, "Yep, now use your AI coach and companion in between sessions because I know we've established the guardrails that are going to be in alignment."

00:10:00

Erin Pash: And then look into What if the AI can report back to the therapist to say, "Hey, here's what some of the conversations have been like that could be relevant to the treatment plan that you guys have created." I think this would really help Richard in session if we talked about some of the kind of fears he was bringing up or the anxiety he was having in some of his relationships or how he was asking questions over the weekend. So, I don't want to say that in a way that seems like AI gets to and the therapist gets to decide someone's life, but it's more of like putting these guard rails in place for accountability so that you're not losing the progress of real life and what we know can only be accomplished within a real human relationship. Yeah.

Richard Zwicky: And going back on something you said earlier with regards to you'll hear people respond differently at different times a year. That was one of the challenges I dealt with how do you establish what's a five? If I ask you how are you today? You may always say, "I'm great." And somebody else may say, "I'm okay." But if that's your standard, that's always a five. And as a human, you're able to interpret that and see that change over time. When you use tools that are doing the transcriptions and analysis, often they don't look at intonation. They look at the words. And when people sit in front of Claude or Chat GPT or any of the other tools and start entering in their thoughts, there's no intonation. But intonation is a huge part of the picture. We say a picture is worth a thousand words but the intonation with which some one speaks is equally important. How do you see that gap being overcome? Right.

Erin Pash: I wonder a little bit is AI going to be a more video style where it can try and read, people's facial expressions. I know that when my kids will mess around with chat GPT, they love hearing it talk. And therefore, we'll talk to it, where you can get a little of that intonation just by pitch and tone of somebody's voice. But truly like what the magic of the human experience is the mirror neurons. It's how we literally connect human to human. Am I feeling what they're saying feels real or am I sensing something different? And frankly that is the biggest reason I know that AI will not eliminate the need for human connection and therapists to kind of interject. You can't have empathy for something that doesn't give empathy and mirror it back to you. So, I think that's kind of one of those limitations. So, I don't know. I want it to continue to get better because we talk about if it doesn't get better and start reading that and we start going down these rabbit holes of confirmation bias and it feeding the wrong information because it can't read intonation and it can't understand the depth of nonverbal human cues, then that's where it gets scary to me. So I think as long as there's this intermediary whether it's a therapist or we're con committing to this technology with the guard rails of we know it can't reflect these things because it's not a person. I'm always reminding people will name their chat GPT and I'm like it's not a person It's an it. It's a program. But people become obsessed with a personality because they think it gets to know I'm like nope it's just repeating what you've already told it about who you are. And so I think it's those soft subtle cues too and even just like social infrastructure that we need to make sure that we're all kind of holding each other accountable to. It's a tool not and It's not a person and it's not meant to place. He

Richard Zwicky: And it is an ongoing struggle for the designers of the AI systems to make the systems appear more personal but they're learning from moment in time snapshots with regards to the questions you ask. And actually that was something I tackled as well and I've written a patent about that should publish sometime in the next few months which is how do you train an AI system on true human emotion without asking a whole pile of questions without a person being able to spoof it based on I'm asking these few questions and that allows the system to personalize but actually understand this person has the following crises or challenges respond appropriately which is key because Otherwise, the systems do too much harm. But patients we're chatting about use it before after and they believe to help make their therapy sessions more effective. We've chatted a bit about, some therapists say they stay away from it, some integrate it. You advocate for integrating it in the right way. We don't want to outsource too much of clinical judgment or

00:15:00

Richard Zwicky: Intuition. What does AI do well in mental health that really people should focus on? And where do you find it really falls short that even it appears competent stay away?

Erin Pash: My favorite way that I even personally use AI is that feelings and what's going on with you can be hard to pinpoint, that is probably the hardest part of the human existence when it comes to communicating. This is the experience I'm having, but I'm ineffective at trying to paint the picture in my head into something that you can understand, And so what I use AI for is I'll put out a ramble and I'll say, "Help me make sense so I can communicate better to somebody else in this context." So that is where I find AI to be incredibly helpful. I actually made a whole list this morning of you statements because that's how I was feeling attacked and defensive and I said, "Will you turn these into I statements for me?" Right? And so that is the best tool, So that a human being doesn't have to I find a lot of times when people are feeling misunderstood, they spend so much time just trying to say what they need to say that they lose the stamina to actually do the work. And so that's what I like. It's a tool like it can help you clarify your thoughts. It can help you say things that create more ease of communication instead of reducing it. And to your point, it's the art of the prompt. You have to kind of know and precursor like I say "Hey, I'm just going to ramble a little bit and I want you to help me make sense of my thoughts, right?" but that has to be the prompt. And before you just throw in a bunch of thoughts because you never know if the answer the question it's going to ask you is, "Are you okay?" or wow that was really insightful or it sounds like that person sucks or where they're putting kind of ascribing clinical judgment to it. So that is my favorite part about A lot of people will come to therapy too and be like I had all these things I wanted to talk about and now I don't remember any of them. So it can also be a great journal keeper of if your prompt is hey I'm having this feeling I don't want to forget to talk about in therapy. It can hold on to some of those thoughts for you and you can use it as a comp compartmentalization skill, to just be like, I don't want to forget about this because it was really hard for me in that moment and I want to process it with a human where I do Yeah. Sorry. Go ahead. Okay.

Richard Zwicky: I was going to say no that's a fantastic use case and you made me laugh because I actually have just finished the first draft on a book and so my first critiques and reviews was three different AI tools, Right? Is the language consistent? Is the story consistent? Is the use but you write the prompts to explain to what to do and to see the comparisons because they all do have a different output but the use case you do changing the statements the fashioning of the language. It's the sort of thing where we don't have access to all the tools we need. We could accomplish so much more and all of a sudden this tool is there used properly. It expands our potential or it allows us to reach our potential faster than we would otherwise. But what happens to accountability especially in therapy, if part of the work is offloaded to a tool that really isn't relational,

Erin Pash: I think it allows the therapist to focus on the work instead of the pre-work. And so I think when one of the things that really bothered me is there's two kind of cultures within the therapy community. It's one where we get patients better and one where we keep patients sick. And I'm not saying it's not intentional, right? But sometimes a client could stay with you for 10 years and see you every week and that's clinically indicated. And other times even the therapist and the client don't know am I still working towards treatment plans because you can't look like as a therapist if you're sitting across from me and you want to work on just getting healthier marriage skills. I probably know there's a lot more things that you should or could work on, but if that's not a part of the treatment plan, then that's not what we're there to address. And so I think if people can come in to therapy, we know it can make a big difference, but if they can have a little more clarity when they come into therapy around here's what I thought, here's how I reacted, here's how I can share, then you can work with the therapist on how to reform without having to get so lost be in the problem behavior. And so I see it as a tool that can be helpful to move people through a discomfort of what's going on in their life or the things they're trying to fix faster because our life doesn't happen in one hour a week. It happens in 170 or something 168 hours a week something like that. So if you can have this person or not this person or if you don't have a friend or somebody else you can kind of connect with like I say I always use the 2 o'clock in the morning because that's clearly when thoughts wake most people up. If you can resource and name that feeling then so you can fall back asleep, you can put a pin in it, you can find a healthy way to kind of still be able to wake up on time and it not impact the rest of your life. Then you can really take that to a therapy session and tackle it effectively.

00:20:00

Richard Zwicky: No, absolutely. And when you look at that, how do you see differences in how high functioning or high achieving clients engage both with you and with AI versus more vulnerable populations?

Erin Pash: So the vulnerable populations is the one that we have to worry about because there's often more complex layers, right? People who are experiencing active hall hallucinations and delusions, paranoid thinking, all of a sudden can have this companion that with that confirmation bias can now say make them hate their therapist, if they are having a paranoid belief or start shutting them out. And so everything is that's why I love the therapist being a part of it, to help create these guidelines and parameters because I do worry about vulnerable populations who tend towards loneliness and are using it as an unhealthy companionship. And that is what we're seeing. I mean, and quite frankly, even high functioning folks are becoming unhealthily dependent on this thing that they've interpreted is a companion.

Richard Zwicky: And the challenge there of course is the business interest of the person creating it,

Erin Pash: And I actually did some consulting work where people wanted to call it an AI companion. And I really challenge that. I'm like, it needs to be a tool in your toolbox, not a companion. Because that's a slippery slope of a high functioning person who can see it as such and somebody who maybe has a very diffuse boundaries on what a companion really means to them. And so we have to be creating the guard rails for both.

Richard Zwicky: The companion will sell better and faster and that's going to drive the economics behind it. So it's an interesting challenge there. Where do you draw the line? What should never be augmented, assisted or touched by AI in therapy?

Erin Pash: Crisis management. Risk I think there are certain things that a therapist can discern that are appropriate and not and that just happens in human behavior. For example, I have taught thousands of therapists how to deescalate suicidal ideiation as a crisis. And I have tried to create guard rails for different companies on how we can assess a crisis truly. And it cannot. Because a lot of times therapists are like somebody comes to you and says, "Hey, I think I'm going to take my own life." And that's not always a crisis or an emergency. You have to pay attention to those cues and soft skills to say, "What are they really trying to say to me? And where are they stuck? And where can I help?" And calling the police or getting people involved when they're not actually suicidal when you have to assess that can be just as damaging as not calling when they really are. And so I don't think it's possible truly and

Richard Zwicky: And that's a really tough one for AI. And as we were talking about before with lack of intonation, understanding or understanding beyond the words on page or the context, that's an incredibly tough one for a system to adapt to.

Erin Pash: And even if it were I don't think it's worth the risk. In mental health, we take every human life seriously and that is not something that AI is necessarily trained to believe. It's repeating and it can be searching the internet and finding information on suicidal affirmation for all we know, Especially if it's being trained that way. So crisis management can absolutely not ever be tackled by AI. I think that's always going to require human intervention because it's a human life at risk and we can't risk that. I think some of the other things that AI will always struggle to do is understand because of the intonation issue the depth of relationships. I think there's a bias that every client brings to therapy that a therapist has to be checking which is I can only know what you're telling me and I'm trained to read between the lines. But if you do that with AI and say, "My wife's this nagging brat all the time." And all you feed it is negative things, it could spit something out that says, "Are you sure you want to be with this person? Are you sure this is the right choice?" And can really destroy a life because it's using your information in a way that's not just I'm venting and not knowing what questions to ask.

00:25:00

Erin Pash: And especially with being able to keep track of that over time and getting to know without a human to be able to interpret what happens in a relationship it can be really quite damaging to somebody's life.

Richard Zwicky: It can be and you touched on something earlier which is loneliness and people not having friends and the data suggests that 50% of adults do not have a close friend they can actually open up and talk to that they can turn to in a moment where they're feeling emotional doubt or insecurity and believe they can confide in them or and believe they have somebody they can talk to and that's a horrible situation because we are social creatures. It speaks to our inability to build depth and relationships on so many levels and these people they are turning to a system

Erin Pash: There. Yeah.

Richard Zwicky: Because they know they need it but they don't know how to find it. Where do we need to attack that problem? Because it's too late just to go to therapy. It's never too late but it's what I mean is it's something which should be tackled before people really hit adulthood before they're starting to isolate through their careers.

Erin Pash: So, I was just reading about a study that was specifically around mothers and children and them being kind of in crisis. And they created anxietyprovoking things like kids who were 10 had to do a presentation in front of people. Some of them got a hug from their mom after some got a phone call, some got a text, and some just watched a 75 minute video that was kind of neutral after. And the text message was the same as watching a neutral video. So, we know like that when you talk about that intonation and that ability to hear somebody's voice is second best to being physically present. And so what I think we need, and I know it's happening because I know people are creating these tools, but we need to be able to find a way to adapt to our current environment, which is to find people to be more on demand by video and if we can't find those connections in person, and that's just not where we are in our life for whatever reason, let's just meet the problem where it's at, which is 50% of adults are lonely or don't have that connection. We know that sense of belonging has to come from a human voice or a human face. And so I've seen apps and programs where it's like therapists on demand video calls. But where do you draw the line of balance and making sure and keeping it affordable when these other programs are free? And so I think they're problems that just haven't been solved yet. Like I said, I live in this dream state where therapists are the ones who are running the AI programs in the background. So that it's not just an avatar. It could be your actual therapist to try and simulate that actual connection instead of a computer. But it's tricky and I think we're still at the beginning and there's going to be a lot that we have to learn about how we can both prevent bad things from happening with AI while also nurturing the best positive outcomes from it.

Richard Zwicky: And that's the thing, the tool is not going to go away. So, we have to figure out how to nurture the best possible outcomes. That's key. We are running out of time, Erin, but people who want to learn more about you, the work you do, consult with you however you'd like to engage or they'd like to engage, how can they reach you? How can they learn more?

Erin Pash: My company So, it's pash company.com. You can email me aaronpash company.com. We're an incubator for social health initiatives. So, when you talk about that 50% of loneliness for adults, we are actively trying to combat that by using tech, AI tools, community engagement, getting people connected in ways that maybe they haven't felt like they could especially since co. And I'm also all over the Instagram and the internet. So, it's Aaron Kelly Pash. And I love to talk to people about this stuff.

Richard Zwicky: Wonderful. It's been a really great discussion. Thanks for joining us today on High on Healthy and

Erin Pash: Thank you so much.

Richard Zwicky: Thanks to everybody for listening. I'm Richard Zwicky and we'll be back again with you soon.