Hey there, you amazing therapists. Come on in, grab a drink, and take a seat. Hopefully, you're listening to this on Wine Wednesday. At home, not driving or seeing clients. I'm Eric Twaktman, and today we're going to sit with something that may start to pop up in your therapy sessions if it hasn't already: artificial intelligence. If you're studying for your licensing exam, here's what you need to know. The exam doesn't care if you can build a prompt in ChatGPT. Nobody's asking you to code in Claude or use cowork to modify your therapy notes. The exam tests whether you can hold on to your clinical judgment when an AI tool hands you a different and better answer. That's the whole game, right there. Why? Because AI is going to promise you the world. A treatment suggestion, a risk score, an incredibly clean, neat, and orderly summary of your session. And the question the exam's really asking underneath all of it is who's responsible for what? And what happens next? By the time we're done here, you'll know what the NBCC actually says, where the ethical tripwires are buried, and how to spot a wrong answer choice from across time and space. Tell me that's not an example of quantum entanglement. Well, I know you may be clueless about what I'm talking about here. I was clueless too, because I didn't get into counseling to swap my feelings and emotions and help others for some soulless chat GPT, OpenAI, Gemini, or God forbid, Grok. Let's start with what AI and therapy even is in plain
terms. It's any AI tool that takes in data and spits out a recommendation, a prediction, or some generated content that you might use in counseling. I know, that's a lot to unpack. Documentation software that drafts your progress notes? That's it. A chatbot that checks in with your client between sessions? That's it. A prompt that flags somebody as an elevated risk? That's it. A scoring engine that you bolted onto an assessment instrument? Same umbrella. All of it. Now here's the core mechanism. And this is the part you might blow right past. These systems
produce output based on patterns and data. They don't understand your client. They don't know your client's flat affect today is because her mother died on Tuesday. They generate something that looks plausible and plausible isn't the same thing as accurate. Let that marinate, because once you've got it, most of the ethics falls into place on its own. So, what's this look like in real life? You might see AI drafting a treatment plan after you punch in intake data. You might get an automated risk assessment back with a number attached to it. You might have a telehealth platform that transcribes and summarizes your whole session for you. And the signal you're watching for the one thing is the moment the AI tool stops assisting and starts deciding. That's where the ethics violation lives. Right there. The NBCC's language on this is blunt. AI mustn't be used as a replacement for professional judgment or the therapist-client relationship. Not a supplement gone too far, a replacement. That's the line. Before we go further, let me give you the five core principles to
consider. If you grasp these, you can reason your way through almost any question on the licensing exam dealing with AI. Accountability, client welfare, competence, and notice the NBCC splits that one into two flavors AI competence and clinical competence. And confidentiality, one accountability, AI recommends, the therapist decides, and the therapist's accountable for outcomes. Two, client welfare. Your primary responsibility is advancing client welfare. If AI might compromise well-being, you use an alternative approach. 3. Competence. You understand how the AI tool works, what it can and can't do, and how it handles client data before you use it. 4. Competence. AI never replaces your professional judgment or the therapeutic relationship. Your skill in assessment, diagnosis, and treatment still has to be there. And multicultural considerations get folded right in here. 5. Confidentiality. You protect privacy when using or storing client data with these AI tools, with actual policies and safeguards behind it. 5 Principles. Say them out loud a couple times. Accountability, Client Welfare, AI competence, clinical competence, confidentiality. Okay, but what about the person on the other side of it? Your
client? Clients' reactions to AI are all over the map. Some are totally fine with it. Some feel surveilled, and let's be honest, that's especially true for clients from communities with a long history of being monitored or misdiagnosed by institutions. That's not paranoia, that's memory. And the NBCC actually names this as an obligation. You discuss and address client concerns about AI tools and data usage. Don't deflect them, address them. And some clients are just clueless. That's the biggest problem of the three, because a client who doesn't get it can't consent. When she finds out later that her words got fed into some system, the rupture won't be about technology. It's about trust. You'll see that framed as a relationship or attachment injury on the exam, not a tech problem. Write that down. That's worth remembering. One more piece here that you might skip entirely. Access. Before you engage a client with an AI tool, you'd better make darn sure she actually has compatible technology to use it. If the AI tool requires a smartphone app and she doesn't have a smartphone, that's not her failure to keep up. That's a barrier you just created. Let me separate this from two things students constantly mix up. First,
telehealth. Telehealth's a delivery method. It's how you and your client end up in the same room. AI's a decision or content tool. You can do telehealth with zero AI involved. You can use AI sitting face to face in your office. Totally different vibes. That said, they overlap constantly in practice. And the NBCC has a whole telemental health section for exactly that reason. AI use gets disclosed in your telemental health informed consent, and the systems have to meet HIPAA and state security standards for digital transmission. You do have a telemental health informed consent, right? Second, standardized assessment. A structured instrument is validated, it's normed, and it's transparent about how it generates a score. A lot of AI tools aren't any of those things, and the NBCC says it out loud in the counselor education section. Some of these AI tools are unproven or experimental, and students should be told so. The licensing exam loves this distinction, and here's why. A therapist who treats an AI output like a validated test score has already fumbled the bag before they've done anything else. Ethical and legal considerations in AI use. Alright.
Let me walk you through the two areas that may generate the most AI exam questions and, honestly, the most real-world complaints and litigation. Number one, transparency and consent. The NBCC principles are pretty direct here. You tell the client if and when AI is going to be used. You explain what it's for, what it can't do, and what kind of AI tool it is. Its purpose, limitations, specific types, those three, every time. You get informed consent specifically for AI, not some vague line buried on page 9 of a 12-page intake packet. Specific. The client gets to opt out. And if they opt out, you provide an alternative service. Run that back, because it's the answer to a whole bunch of exam questions. Declining AI can't cost her care. If your only workflow requires AI and they says no thanks, you find another way. That's it. That's the standard. You also owe your clients transparency about the data itself, who can get at it, whether it's shared, how long it's kept. And here's my honest gut check. If you can't answer those three questions about an AI tool you're using, you're not ready to use it. Second area, competence, oversight, and accountability. These three
travel together. Before you touch an AI tool, you need foundational knowledge of how it works and what it can't do. Same standard you'd meet before administering any assessment. Nothing exotic about it. And competence includes knowing when AI is not appropriate. That's a judgment call that NBCC puts squarely on you. Then you review every recommendation it produces before it reaches your client. Everyone. And the NBCC calls out risk assessments by name here. Automatically generated risk assessments get reviewed carefully before you act on them. That one can show up on your exam. An algorithm hands you a risk score, and the score isn't the decision, you are. And then, if client welfare's on the line, you override the AI tool. Not negotiate with it, override it. And if the AI tool's causing problems, you don't just quietly work around it, you cease use and report it to the developer or organization. There's also a welfare tenant people forget because it sounds like a business issue rather than an ethics issue. You prioritize client welfare over efficiency or financial incentives for using AI. Read that as it saves me 20 minutes, a note is never a defense. Efficiency isn't a clinical justification. Here's the line that anchors this whole conversation. You remain fully accountable for client outcomes. The AI recommends. You decide. There's no version of this where the software takes the blame for you. Practically speaking, that means you document when you used AI, which AI tool you used, and why. Three parts instance, AI tool, and justification. It means you actually evaluate whether client outcomes improved, so you know if the thing's even working. And it means you report problems with the AI tool back to the developer and to clients where appropriate. Two more that show up constantly. Confidentiality client data going into or sitting inside an AI
system gets encrypted, access gets restricted to only the personnel who genuinely need it, and you're meeting HIPAA plus your state's digital privacy law. There's a fourth piece people often miss. You need a policy for securely deleting client data once it's no longer needed. Not just protecting it while you have it, getting rid of it properly when you don't. And bias. AI systems inherit the bias of whatever data trained them. So cultural competence isn't a side note here, it's central, and the NBCC folds multicultural consideration directly into clinical competence and tells therapists to stay informed on AI risks to maintain cultural competence. A risk algorithm that underestimates a client's distress because of their language, their presentation, or the way their culture expresses symptoms? That's a client welfare problem. And it's got your name on it, Case Study. Let's look at a typical narrative. Marisol's a counseling grad student
three months out from her licensing exam. She's at her practicum site, she's exhausted, and she's been leaning on an AI tool to barely keep it together. She sits down with her supervisor to talk through how she's been using it. Here's what comes out. First thing, Marisol's been pasting anonymized client details into a general-purpose AI chatbot to generate treatment plan ideas. In her words, I take out the first and last names, so my client's covered. The only thing she enters is the client's age, diagnosis, country of origin, and the specific circumstances of a recent workplace incident. Taking out a name isn't de-identification, it's not even close. And notice what else is missing. She has no idea whether that chatbot encrypts anything, who at that company can see it, or whether it's ever deleted. She hasn't met a single confidentiality tenant. Second thing, for one client, the AI tool generated a treatment plan built around exposure work for panic. Marisol liked it, it looked clean, it was well organized, so she used it in her next therapy session with her client and started implementing the plan. She didn't check it against her own case conceptualization, since it seemed to cover everything in way more detail than what she remembered of her own attempt. When her supervisor asked what led her to choose exposure work for panic over other approaches, Marisol said, that's what was recommended with what we use at my practicum site. Third thing, she never told her client anything about using AI. Not at intake, and not in any sessions since she first saw her client. When her supervisor brought it up, she said she didn't think it was necessary because the AI wasn't interfacing with the client directly. She also hadn't documented a single instance of AI use in any note, no AI tool, no justification, nothing. And because no client ever knew, no client ever had the chance to opt out. And then a fourth detail, this one's the quiet one. Marisol mentioned that for one client, a Spanish-speaking woman, the AI generated treatment plan felt off to her. Something in it just didn't click. But she went with it anyway because she figured AI definitely knew more than she did dealing with someone from a different culture than her own, and besides, it's what her practicum site uses. Okay, here's what you should notice looking at all this. None of what was constructed and implemented was done in bad faith. Marisol's a good intern who's just stressed. And that's exactly how ethics violations actually happen in the real world not through malice, but through fatigue and deference. So, why is this an AI ethics situation and not just a narrative about a stressed-out intern? Let me build it for you. Take the treatment plan first. That's what's recommended at her intern site. Line that up against what we said about accountability. The therapist bears responsibility for clinical decisions and is accountable for client outcomes, and there has to be human oversight of AI-generated treatment plans. Marisol didn't override. She didn't review for appropriateness. She didn't run it against her own conceptualization. She swapped the AI tool's expertise for her own judgment. That's the exact failure the NBCC principles were written to prevent AI being used as a replacement for professional judgment. Now the Spanish-speaking client. Marisol sensed the plan didn't fit and went ahead anyway. That's a competence and client welfare failure sitting right on top of a bias problem. The principles require intervening or overriding when the client's well-being may be compromised. Her clinical instinct was the signal that the client's welfare might be compromised, and she overrode herself in favor of AI. Her actions were backwards, completely backwards. The consent piece stacks on top. No disclosure, no AI-specific informed consent, no opportunity to opt out, no documentation. That's four separate ethics violations in a single omission. Now, here's the counterargument somebody's going to raise, and it deserves a real answer. You could say the actual problem here is supervision and her practicum site. She's a student. Nobody trained her on this. The site's at fault, not her. Fair. That's a genuine gatekeeping failure, and the NBCC's explicit about it. Supervisors should be knowledgeable about AI tools their supervisees are using. They should address ethical AI use in the supervision agreement. They should monitor supervisees' AI use as part of gatekeeping. And they should intervene when a supervisee uses AI in concerning ways. Counselor educators are supposed to teach critical thinking on AI integration and warn students that some AI tools are unproven. Marisol's site did none of that, but it doesn't let Marisol off the hook. The principles are explicit. Competence comes before use. You don't get to use an AI tool you don't understand and then point upstream at somebody else. Both things are true at once. The site failed her, and she's still accountable for the clients sitting in her chair. It's like she's already lost her license before she even sat for the exam. Make sure you don't end up in Marisol's shoes by making sure you've got everything we talked about covered. AI probably came out years after your supervisor was in grad school, so check on how up to date they are on AI. If you didn't have a course on AI and counseling, chances are your supervisor will be out of the loop when it comes to this issue. And if they're still using a flip phone from 1996, well, you get the picture. Let me give you an overview, and I'll tie each one back to Marisol so it sticks for the exam. Informed consent for AI. Consent obtained specifically about AI use, its purpose, limits, and type of AI tool. It's not the same as your general treatment consent. Marisol had general consent and figured that covered her. It didn't. Have a separate form for this? You do, don't you? Human oversight. A person reviews AI recommendations for appropriateness before they reach the client and intervenes or overrides when welfare's at risk. Oversight isn't a glance. It's actual clinical judgment applied to the output. Marisol read the plan, liked the formatting, and confused reading with reviewing. Those aren't the same vibe. Accountability. The therapist, not the AI tool, not the developer, not the supervisor, not the practicum site, bears responsibility for clinical decisions and client outcomes. And it comes with paperwork. Document the instance, the AI tool, and the justification. De-identification. Removing information that could reasonably identify a client, not just the name. Age plus diagnosis plus country plus a specific workplace incident can point straight at one person. This is where Marisol thought she was safe and wasn't. Algorithmic bias. When a system produces skewed output because the data used to train it was skewed, looks like there's underestimation of distress in some populations and overestimation in others. It's exactly why cultural competence is an AI issue and why the NBCC tells you to stay current on it. Right to refuse, sometimes called the opt-out. The client can decline AI, and you provide alternative services when she does. Refusing can't shrink her access to care. And one more the exam quietly loves AI competence versus clinical competence. AI competence is knowing how the AI tool works and where it fails. Clinical competence is your assessment, diagnosis, and treatment skill, which the AI tool never replaces. You need both. Having one doesn't excuse missing the other. Three things I want you to carry out of here. 1. The therapist is accountable. Always. The AI tool recommends you decide and no exam answer that slides responsibility
onto software is ever the right choice. Ever. Period. Two, consent for AI is specific, separate, and refusable. If the client says no, they still get the care they deserve. 3. Competence comes first. You understand the AI tool before you use it, and if your clinical instinct disagrees with what it spits out, your instinct wins until you've got a real reason to think otherwise. So here's what you want to look for in an exam question. Any stem in which a therapist accepts an AI output without reviewing it, in which the client was never told, or in which efficiency is offered as the reason. Those are your flags. Chances are you're dealing with an ethics violation. Here's one practical move before you go. Before you use any AI tool with a client, answer three questions out loud. Who can see this data? How long is it kept? And how does it get deleted? What happens if this recommendation is wrong? Can't answer all three? You're not ready to use that AI tool. And one last thought the AI tool's a consultant who's never met your client, it never will, and it doesn't have a clue what it doesn't know about you or your client. It's dumber than a rock, and it's your responsibility to either keep it, use it responsibly, or give it the heave-ho. The key takeaway here is that AI is an it. It's not your co-therapist, and no matter how it tries to be humanized, it'll never be a replacement for what you can and do provide for a client. You're the one in the room doing therapy with your client, not it. So when you're sitting for your licensing exam and a question drops an AI recommendation in front of you, think when push comes to shove, who will be the responsible party when you're sitting across from your client in a courtroom, and remember, it will always be you that's responsible. And as I always say, it's in there. You can do this.