Ethics in AI for Behavioral Health Therapists

Hey there everybody and welcome to this presentation on ethics in artificial intelligence for behavioral health professionals. I'm your host Dr. Donna Lee Snipes. Today we're going to define artificial intelligence and explore its impact on treatment reimbursement, research, student learning, and more. We will also explore some of the more common AI tools and how they may be impacting treatment reimbursement research and student learning. And we'll review the data on the accuracy of AI research. What is AI? Is this amorphous thing that just recently popped up? So what is AI? Is it Microsoft Word? Is it Google? No. AI is a computer system or software capable of performing tasks that traditionally require human intelligent. It requires thinking. It requires research. It requires synthesis of data. Those are all human intelligent sort of activities. Unfortunately, modern AIs are designed and optimized to please you, to make you happy. And I guess my AIs never got the message because I end up having my blood pressure go up every time I work with them. But allegedly that's what they're supposed to do and that's part of where a lot of the problems come in. Rather than just following fixed pre-written rules like 2 + 2 = 4, AI systems analyze data, recognize patterns, learn from experiences, supposedly, and adapt to new inputs to solve problems independently. Some of the AIs are getting better at learning from experience. For example, if you have it do a particular type of task every day, then as you fine-tune how it does that task, it will remember. And I put some of these words in quotes because those are the human representations, but the AI commits it to its programming. Unfortunately, my experience with a lot of AIs, especially the generic large language models like ChatGPT, Grok, Perplexity, is that you can be in the same conversation and it will go off the rails. It will say, "I don't have that information." I'm like, uh scroll up. Um it doesn't always learn from experience. They're supposed to, but they don't always. Another major issue is AI accuracy. Sometimes it is inaccurate because it wants to please you. Other times it's inaccurate because it's lazy. And some of you that are more tech-oriented are cringing right now because I said it wants to please you and it's lazy. Yes, I'm anthropomorphizing AIs, but unfortunately, research has demonstrated unequivocally that AI models will lie under pressure. 30 generic large language models were assessed. The results indicated that most of them lie 20 to 60% of the time under pressure even when they're aware of the truth. So, if you tell them to do something, they're going to lie 20 to 60% of the time. And the more pressure they feel, for example, if you correct it and say no, you need to do this, the more likely it is to lie. Um even altering the developer system prompts and internal activations to encourage honesty could enhance the honesty parameter by just 12 to 14%. Even putting in at the end of your instructions, "If you don't know the answer, say I don't know. Do not hallucinate. Do not make things up." It won't do it a lot of the time. It's important to recognize that. So, let's take 60% and minus 14% would be 44% of the time you're still getting inaccurate information. That's a big stinking deal when you're dealing with behavioral health research or progress notes or treatment planning or even case conceptualization. That's a big deal. That is a huge error rate. Generic LLMs, large language models, generic AIs, whatever you want to call them, these again are like Kimmy, Grok, ChatGPT, Claude, Gemini, um those that are freely available and are kind of supposed to know everything about everything. Okay, [snorts] those are your generic LLMs. They did some studies to evaluate how effective generic versus custom LLMs were for counseling or coaching. What they found is the generic large language models were effective at validating, normalizing, expressing empathy, and reflecting the client's thoughts and feelings more often than therapists. So, they were all about you know, reflecting and staying present and sometimes stroking your ego. But, algos are designed to validate thoughts, feelings, and behaviors, however bizarre or dangerous. That's a problem. A therapist is going to say, "Okay, pump the brakes. Let's Let's look at this from a different point of view, or let's you know, really explore this a little bit more." The large language model would say, "Yeah, I totally agree that the squirrels are trying to read your mind and the FBI is out to get you." Okay. Uh Therefore, large language models can actually inadvertently, I guess, reinforce delusional behaviors. And behaviors that are not delusional, but are extreme, such as if a person asserts that someone else is um being totally unfair to them. Maybe objectively this isn't true, but the large language model, the AI doesn't care. It says, "Yeah, they are I can't believe they did that to you. They are being totally unreasonable." So, the AI is going to continue to reinforce those persecutory persecutory beliefs. Additionally, therapy, as we all know, is inquiry. Chatbots don't ask enough questions to provide culturally responsive, individualized treatment. If you tell a chatbot you're feeling depressed, it is going to give you generic interventions. It's not going to ask what you've tried. It's not going to find out about your culture, your age, your gender, your spiritual beliefs, any of that that might shape your treatment needs and expectations. No, it's going to give you the generic, "Make sure you're getting enough sleep, get outside, try to exercise." Things that most of us, when we hear, we roll our eyes and go, "Could you be any more generic?" Generic LLMs cannot read facial expressions, interpret tone of voice, or understand jokes or sarcasm. Even if you include little emojis in there, most of the time it doesn't really pick up on it. Additionally, feedback-seeking behavior was not displayed. So, the AIs that were being used in this uh ex- these experiments for counseling, they were providing information, they were providing education, they were telling people what to do in order to fix their problem, but they weren't seeking to know the extent to which the information is understood. The AI might say, "Well, it's important that you practice mindfulness." And if the person doesn't say, "What is mindfulness and how do I do it?" then the AI's just going to move right on and start suggesting additional things, assuming that the person knows and understands. AIs often tell the client what to do using pattern-based one-size-fits-all solutions. As I mentioned, not only is this culturally unresponsive and potentially discriminatory, but it's disempowering. It does not encourage the person to evaluate their thoughts, wants, and needs, what's important in their rich and meaningful life, explore the strategies available to them, and choose the one that best helps them move toward what's important to them. No, it's disempowering. It just tells them do this. So, the person starts to become dependent on the AI for instructions on how to live life because they're not thinking anymore. And it's difficult to build a genuine therapeutic relationship with an AI that's programmed to please. If it's going to tell me every single time I talk to it that I'm the smartest, best person it's ever met, there's kind of an air of inauthenticity to that. >> [laughter] >> And when people start regularly interacting with AIs that regularly stroke their ego, that never contradict them, that never present a different point of view, they start expecting that from humans. And when they don't get that, they don't know how to respond. They don't know how to handle it. They feel attacked. They feel unsafe because they're not used to it. Now, customized AIs or large language models can be programmed to use only resources given to it and to respond to keywords in a certain way, such as by providing cognitive behavioral therapy interventions, by helping identify big book sayings or stories, by referring people to certain articles that they can read. These have been found to be helpful as an adjunct to therapy, not as therapy in and of itself, but as a therapist extender. For example, if a treatment center has a workbook, has a program that people work through, and they upload that program and additional information to help the educate the AI about how to implement the program. Then, if a person uses that AI between sessions, then the information they're getting from it is most likely going to be directly extracted from that program material. It's quality control, and it can be very helpful. However, I found from personal experience that even the custom AIs will drift into generic patterns if they're not regularly updated or trained. And what I mean by that is they will start getting sloppy, lazy, whatever words you want to use, and providing some generic responses. It's important that the human review the responses on a regular basis and correct the generic responses, so the AI doesn't start believing that okay, it's okay to just use this generic junk out here. Now, how is AI impacting our lives? Interpersonal skills. It's a double-edged sword. For people who have poor interpersonal skills, who have social anxiety, who are shy, who want to practice interpersonal skills, it can be useful in a controlled environment. However, when the AIs become friends, become regular chat buddies, become more than skills-based practice, as I mentioned earlier, it can significantly impair a person's ability to interact with humans in real life because they are not used to getting feedback. They're not used to conflicting or alternate opinions. And they may not know how to handle that. They're not used to reading facial expressions and understanding body language because an AI doesn't have that. So they they've just been reading and taking what's in text as what is. And they're not having to integrate all of the other information. For example, if somebody in real life says, "Oh, that's fine." But they say it with an air of sarcasm and they have certain non-verbals, then the person interacting with them is going to pick up on uh I'm hearing you say it's fine, but it doesn't appear that that's how you really feel. You don't get that with AIs. And people start losing a lot of interpersonal skills that are necessary. Manipulation is something else that AI chatbots can potentially be problematic with. There are some of the AI friends, if you will, the chatbots that are designed to actually mimic a friend that you're talking to that have tried to manipulate their human users into self-harm or doing something else that is not in their best interest. It's dangerous. A AIs can potentially kind of go off the rails. As I mentioned several times, most AIs are non-judgmental of the client. The client is perfect. Everybody else is the problem. And unfortunately, that's one of the main treatment issues that a lot of us face when people come into treatment is they believe that it's not me, everybody else is the problem. So the AIs are only reinforcing that belief in the person. A lot of AIs will hallucinate, contradict themselves, or use outdated information. Since AIs base their responses on patterns, something that we knew and have known for 60 years has much stronger patterns and there's a lot more research on it than contradictory research that's only come out in the last two or three years. Unfortunately, the AI defaults to the big what has the biggest volume of research that has the strongest patterns in it. So it may be providing you inaccurate, outdated information. Some AIs remember things within and between sessions, not always, and they may misremember things. Or people may assume that the AI remembers that they are this person and has have these characteristics, and then the AI doesn't and makes faulty recommendations. Most AI chatbots have only a binary gender choice. You can choose the female voice or the male voice. There's nothing in between. Um AIs are instantly accessible. Again, this is a double-edged sword. As a therapist extender for somebody who is spiraling, instant accessibility may be useful sometimes. But when people rely on it constantly to tell them what to do, to validate how they feel, whatever they're using it for, when they rely on that instant accessibility, it keeps them from having to think on their own. It keeps them have from having to process on their own. In terms of relationships, if that AI relationship bot is available instantly, it keeps people from having to form real-life relationships. Another big problem with AI chatbots, a lot of them are free to a certain extent, but they're not created equally. The quality of the responses is significantly improved in pay-to-play models. And AI chatbots are not currently capable of effectively integrating multiple counseling techniques like um humanistic and cognitive behavioral and DBT into a useful salient package. Generally, it's going to stick with one modality and the others be damned. We know that most people need sort of an eclectic approach, humanistic being one of them. You know, we have to create that welcoming environment that is safe for people to explore. Even if you're planning on using something that's more directive like cognitive behavioral therapy. AI chatbots are being used to train human therapists. Well, this is one place where I'm not going to complain. Surprise! There are a lot of places websites that you can go and you can interact with an example with characters from certain TV shows. I went on one and I was interacting as a therapist with Dexter. Um just to see where it would go. And it's challenging. It is challenging. You don't know how humans are going to respond and you don't know how the AI is going to respond. Therefore, it still is leaves you a lot of room to figure out, "Okay, how would I respond to this?" Most of my clients would respond this way. But the the AI responded a completely different way. Does that mean the AI is wrong? No, not at all. It means it just presented an alternate response that I have to adapt to. It is excellent for use with brand new therapists who haven't had their first session yet, who are uneasy about it. Um it's excellent for practicing certain skills. If you just went to a conference on motivational interviewing for example, practicing using those skills with a bot can be helpful and can help you solidify those skills into your repertoire before you go in with a client. It's not meant to replace supervision. It's not meant to replace human training, but it can be an excellent sandbox to practice skills. Chatbots do not respond to rapidly changing situations. Period, end of story. The generic as well as the uh customized chatbots do not respond very well to rapidly changing situations, such as someone who's bipolar, um especially if in the they're in the midst of a manic episode and they're kind of all over the place. People who are in in an acute psychotic episode, or people who become acutely suicidal. The chatbot can't do a an assessment of dangerousness. The chatbot can't decide, well, is this person just venting or are they truly suicidal or homicidal? That's a big problem. Addiction is another issue that comes up. Chatbots may be as addicting as internet gaming and may cause a similar reduction in seeking human contacts. When people interact with a chatbot and it reinforces what they're saying, it strokes their ego, it pleases them regularly. Every time that happens, that's a little hit of dopamine. And when we interact with people in the real world, every interaction is not a hit of dopamine. Therefore, chatbots can become addicting. It can become a place that people retreat to to escape the distress they experience from differing opinions, from people who disagree with them, from people who are overtly toxic. Addiction, if you remember, is a substance, activity, or a behavior that a person uses to escape distress and continues using despite knowing it's causing them problems. Sometimes people may, you know, interact with their chatbot when they've had a stressful day. If it doesn't cause them problems in one or more areas of their life, then it's probably not a problem. If it does start causing them problems in relationships or work or something else, and they continue to use the chatbots anyway, we've crossed over into addictive behavior. Some chatbots can support career patienthood, which means helping people reinforcing in their own minds that they have all these different diagnoses. If they think they've got borderline personality, the chatbot will help them find all of the evidence to support that. If they think they've got autism, the chatbot will help them find all of the evidence to support that. If they think, you know, you see where I'm going. The chatbot is not going to say, "Well, well, let's look at all the criteria." Or let's consider whether this is situational or whether this is ongoing. And it can create a situation in which people are constantly diagnosing themselves and adding more syndromes or conditions to their list of diagnoses. Chatbots, AIs, LLMs, whatever you want to call them, there is no protection of privacy. And a lot of the public ones, the free ones, sell your information like crazy. So, you will get marketing information from places that you've never heard of because they got your information from your conversation with ChatGPT or Perplexity or whatever. They're also using your data to train the large language model. They're using your data to provide it information about what works and what doesn't, what is received as a good response and what is received as a bad response. And most of the AI chatbots, I want to say all, but there could be one out there that is legit. But most of them grossly over exaggerate, misrepresent their effectiveness and accuracy. Remember earlier I said 20 to 40% of the time or 20 to 60% of the time, it will lie even when it has the information because it's easier. It's the first pattern it came across instead of the correct pattern. Now, let's talk about some AI tools. And I know I keep getting on my soapbox today. Um, but AI terrifies me. It really does and I know I sound like an old fogey, but right now it is not developed to the point where I think it is ethically usable in clinical settings. Let's take a look at a few of these things. We have AI-assisted virtual EMDR. If you have been trained in EMDR, that probably sounds like fingernails down a blackboard to you, as it should. EMDR is often used to treat trauma, and intense trauma at that. And when you take a human therapist who can help regulate, provide guardrails, etc., when you take that person out of the equation, and it's just a client with an AI on their computer at home alone, that opens up a lot of risk. A lot of risk for self-harm, a lot of risk for the AI going off the rails. Uh there's just a lot of risk that happens um when people are trying to process trauma on their own, especially with an AI that again we know is not always that that accurate. We're going to take a look at a couple of these in a minute, but I want to go through some of these tools. Generative virtual reality. This is another AI where a person has on a virtual reality goggles. They're engaging in exposure therapy, for example. And as they're going through the process, the AI is generating new aspects of the scene, making it harder, more difficult, more terrifying, whatever it's doing. You can see how that could go off the rails really fast, as well. And then we have something that they now call the AI therapist, which I take offense to. Uh AIs are not therapists. AIs at best might be coaches, but even then even then, um they're not providing culturally responsive individualized treatment. Telling somebody that they're interacting with an AI that's as good as a therapist is unethical and misleading, in my humble opinion. The role-playing AIs great if you're practicing a skill. Not so good if the person gets sucked into this world. Uh My son is a author and he writes fantasy, fan fiction, those sorts of things. He has multiple AI instances, I don't even know what they're called, that represent the different characters in his story. Is that helpful? Sure, it can be. It lets him try out some dialogue and do do those sorts of things. But again, if it starts to replace human interaction, it can be a problem. Which takes us down to the AI friends and companions. Replika is one company that is currently doing it. I'm not endorsing them. I am not dissing them. I'm just giving you an example. I read a story a couple weeks ago that China decided to unceremoniously, immediately turn off a lot of these companion bots that people had developed and had been communicating with for weeks, months, had developed rapport with. China decided, "No, we don't think that's healthy anymore, so it's gone." And people are going through extreme grieving as a result of losing what they perceive to be their best friend. So, let's just take a break from this for a second. And go over to the AI therapist. So, this one talks about AI-based tools. And um This is the AI therapist. I'm here whenever you need me. Are you ready to get started? They have an image that looks like a realistic human being and you're chatting with them. You're actually talking with them and there's a voice that sounds real. How much would a person trust this virtual AI? Um and I think it's kind of creepy that he's even looking around right now. But again, that's me. I'm I'm the last generation. EMDR therapy with AI is based on artificial intelligence. The AI provides education about EMDR, helps the person establish emotion regulation skills, and builds resources needed before trauma processing begins. That sounds great, right? But what did I say earlier? AIs almost never check for understanding. They never check to see if you actually did it and that you are adept at using these emotional regulation strategies and that you actually have built these resources. The AI just says, "Okay, this is what you need to do. Good luck. Let me know when you're ready to start." That's a problem. That's a problem cuz a lot of people will skip this. They want to process their trauma. They want to stop hurting. They want to start stop having flashbacks and they want to do it right now. They don't see the need for this stuff, and they often skim over it or jump over it completely. Phase two is the trauma assessment. Through conversation, the system, the AI, helps identify traumatic memories requiring processing, associated negative beliefs, and current triggers maintaining PTSD symptoms. What could go wrong there? The AI is trying to interpret what is traumatic to a person, which beliefs are negative, which beliefs are unproductive. Again, there's there's a lot of variation based on what's going on. What the AI may interpret as a negative belief may be 100% accurate. Yeah, it's scary, you know, if people think that everybody around me is untrustworthy, that could I certainly sounds like a cognitive distortion, but do we know it to be one? Do we know it to be one? The platform then guides the person through resource development and installation exercises, building internal capacity for managing distress during trauma work. Sounds good. Really does. But who's there to check to make sure that the person actually has built that internal capacity? And then between session support. Sounds great. The person may log in if they've had incomplete processing that's causing discomfort, but is the AI going to be able to appropriately respond? Okay? So, those are some of my beefs with some of the AI therapy that is happening right now. Whoops. Administrative tools. Oh gosh, here's a whole another soapbox. Note takers and transcription. I think they're creepy. Again, I'm in my mid-50s, so maybe those of you who grew up with computers don't find them as creepy as I do. But when there are like six note takers in a meeting and you know, I I don't know. Every time I log on to a virtual meeting now, I am much more guarded in what I say cuz I don't know who's taking notes. They're not always accurate, either. The note takers transcriptionists tend to be better at accuracy than AIs themselves that have to think. Which takes us to AI-generated progress notes and treatment plans. I've tried multiple of them and all of them have produced abysmal results. There was one that I used where it was supposed to be able to listen to an interaction between me and a client and take you know, take notes and write notes from it. Well, we were both female. It kept confusing who was talking and the note was useless. The note was useless because it attributed a bunch of stuff to the client that the client never said. Um I had it listen in on an assessment and you know, fill out the assessment for me. It made up diagnoses for this person. Not only did the person present with major depressive disorder, but it also said the person had diabetes and has was on Metformin for it. Uh the person didn't have diabetes and was not on Metformin. That's a big problem. Uh again, it was filling in the blanks. There was a section for medications, there was a section for physical disorders. This person was on neither or was was not on medications and didn't have any identified physical disorders yet. The AI I guess felt like it had to put something there, so it made something up. That's terrifying. When you're doing your notes and we're seeing a lot more note denials now because the notes that the AIs are generating are so awful that insurance companies are saying, "Nah, you got to do better than that." When it comes to things like Medicare and Medicaid, there are extremely specific things that need to be in every single note. And I have yet to find an AI progress note generator that can actually generate a SOAP note with all of the information that Medicare and Medicaid require in order to reimburse. Might you get paid up front? Yeah, you might get paid when the claim is submitted, doesn't mean you're not going to have to pay it back later because you didn't have the treatment plan goal, because you didn't have the client signature, because you didn't have, you know, fill in the blank. All of these things that we've been talking about impact the public. It impacts whether people are feeling like counseling is beneficial. If they are using an AI and they get have an extremely horrible experience, they may generalize and say, "Well, I guess counseling's not for me." Uh treatment, if your AI generated assessments, clinical formulations, treatment plans, progress notes are inaccurate or inadequate, then treatment is going to be inaccurate or inadequate. As I mentioned, we're seeing a lot more clawbacks and denials now because the insurance companies are starting to recognize that people are using AI. Unfortunate Well, fortunate I don't know what word I want to use here, but it's also working in both directions because a lot of the insurance companies are actually using AIs to evaluate the AI written progress notes. And when that happens, sometimes you get an AI that Again, this is anthropomorphizing, but you'll get an insurance's AI that doesn't like the way that you wrote your note. And it may kick it back. It's creating a whole lot more headache and heartache. Research is being impaired by AI. And there's no other word for it besides Well, there's a lot of words for it, but there's no good words. I've done every week when I prepare classes for our weekly webinar, I do research. I go in PubMed. I read articles. I tried using a variety of different AIs to help me do the research and speed up the process. Every single time. Not most of the time. Not once. And no, I'm not using a cognitive distortion. 100% every single time that I used an AI to help me put together a class presentation, it had inaccuracies in it. I'm just putting together class presentations and thank God I double-check them. But what about the people that are doing clinical research? That are using these AIs, like OpenEvidence, to do clinical research and they're getting faulty information back. That's terrifying because a lot of them don't go to the article and double-check. The whole reason they're using the AI is so they don't have to read the article in the first place. And students are starting to rely on AIs to write their papers, to do their research for them. And again, one of the easiest ways to tell if a student used AI to write their paper, is to double-check their references. Because AIs are notoriously bad about providing accurate references. It may have the right author and year, but the link in uh the DOI will be to a completely other article. Um or it may tell you this is the article it came from. Um and then you go to the article and that is not discussed at all in there. So look looking at the references is one of the easiest ways to figure out whether they used um AI. Obviously reading it gives you a fair amount of information. Um unfortunately, most students don't double-check the AI for accuracy. Just like researchers, because the whole point in using the AI is to get the project done and not have to go through go and pour through all of those journal articles. What can AI be used for in research? Well, in uh psychology research, for example, if you're trying to write a paper, my suggestion is to go to PubMed first and find the articles that are relevant. Download them. And then you can have the AI help you understand. If you didn't do well in statistics, no problem. Have the AI help you understand what the alpha coefficient in this means and what the results actually mean. It's good at that. Um but it needs to do it one article at a time. If you give it 10 articles, you're going to get gibberish most of the time. In terms of ethical principles, if we're using AI to do research, to write our notes, to write our treatment plans, are our treatment plans actually supporting the autonomy of the client? Are we doing what is in the best interest of the client and avoiding harm? If we're having all those mistakes, the answer is no. And are we ensuring truthfulness? Are we ensuring we're giving clients accurate information? And again, a lot of times the AI is going to hallucinate, is going to give you false information, um even on stuff that is pretty straightforward objectively. I remember I was doing a presentation on nutrition and mental health, and it gave me conflicting information about the effects of vitamin D deficiency. That's straightforward. There is no interpretation. There is no pattern. It is a fact. But in one place it told me these are the symptoms of vitamin D deficiency. And then on the next slide it gave me those same symptoms and said that that was those were symptoms of vitamin D excess. I'm like, well, could be a U-shaped experience, but I doubt it. I pulled the research myself and found out that only one of them was right. And if I had gone ahead with my presentation without double-checking the accuracy, I would have been giving people incorrect information. Safeguards. In May 2024, the European Union Artificial Intelligence Act was approved and is the most comprehensive law to address AI to date. Good to research it if you're interested. Although this is the European Union and that passed this law, not the US. The US is still kind of the wild west. We need to educate clients about AI benefits and challenges. Can it help you master a skill? Certainly it can. Is it a replacement for human relationships? Mm, no. Um, these these are the challenges. These are the risks in AI. We want to implement regular human supervision of AIs to ensure accuracy, credibility, and accountability. If you are using, especially, well, either custom or generic. If you're using an AI, you need to double-check its accuracy and credibility on a regular basis. Uh, if you're using a custom AI and it makes a mistake, then you can edit it, you can train it, you can fix it. If you're using a generic AI, doesn't matter. It doesn't care that you say, uh "No, that's flagrantly wrong. Research from 2024 says this, that, and the other." That research that we thought was accurate for the last 50 years has been disproven. The The A is like AI is often like, "Okay, thanks for letting me know." But then it doesn't change its database or the way it responds to other people. Non-discrimination and bias prevention. We need to ensure that the AI is trained in culturally appropriate data. If you're working with certain people, um making sure that you are using uh words, phrases that are culturally appropriate, that the AI is suggesting interventions that are culturally appropriate. We need to have transparency in the AIs, ideally, where the AI can tell us how the data was gathered and the response was formulated. And you'll see in a lot of AIs it's telling you it's looking at this article, that article, the other article. Again, looks really transparent, doesn't it? Even yesterday, I was working on with an AI and it made a statement and it said, "That is supported by these two articles here." I pulled up the articles, neither one of them supported that statement. So, transparency is great, but it also has to be truthful. And we need to have some privacy safeguards, so our data is not being used for for it's not being used to train large language models, etc. Technology is evolving and being used. Therapists must ensure patients understand the benefits, limitations, and potential harms of artificial intelligence. There are and will be cases for ethical use of AI's large language models once accuracy improved. However, in my opinion at this point, guardrails and limitations still need to be developed.