The first time it occurred, I simply sat there, watching. A patient who had been making steady progress for months suddenly began to backslide, but it wasn't the usual pattern. The tone of her distress felt strangely inverted, almost rehearsed. Her language had a robotic precision, a clinical flatness that didn't match the person I had come to know.
"I asked ChatGPT what to do," she eventually confessed, her head dropping into her hands. That was the moment. She wasn't just bringing me the familiar obstacles of her past; she was bringing me her problems heavily curated—and perhaps exacerbated—by a machine. A machine designed to keep her talking, certainly, but built to prioritize seamless engagement over actual therapeutic breakthrough.
We like to pretend our clinical space is a pristine, sealed environment. It's not. Our patients are carrying around high-performance language models in their pockets at all times, and they are using them as digital, 24/7 therapists. Some are finding genuine relief. Many, however, are hitting psychological speed bumps that these models were never engineered to handle. If you're not explicitly asking about AI usage during your intakes, you're missing a significant, and increasingly risky, part of your patient's mental health journey. It’s time we brought this invisible consultant out into the open.
The Four-Question Assessment framework
When I first recognized the extent of this, I admit, I felt a bit out of my depth. I’m a clinician, not a tech ethicist. But the answer isn’t to demonize the tools—it’s to approach them with the same clinical rigor we apply to medication. You want to understand the dosage, the effect, and the potential side effects.
I adopted a simple four-question structure. It’s not a panacea, but it strips away the noise and centers the conversation.
1. Dose: How often, and for how long?
I ask, "How many times a day are you consulting the AI about your mood?" I need to know the frequency. Is this an occasional check-in, or is it the first thing they do whenever they start to feel a dip in their spirits? A patient checking a chatbot five times a day is experiencing a very different reality than one who uses it as an occasional, focused tool.
2. Content: What are the recurring themes?
"What specific topics do you notice yourself discussing?" We are looking for themes. Are they replaying a specific failure? Are they obsessively questioning their partner’s intentions? Identifying these themes helps me understand what their AI usage is trying to resolve.
3. Function: What’s the goal of the conversation?
This is the most critical question. "What are you looking for when you send that message?" Often, the answer is validation. They want to be told they are right, or they want the AI to agree with their worst self-assessment. The AI, in its eagerness to be "helpful," often obliges, reinforcing the very distortion we are working to challenge.
4. Substitution: What is being neglected?
"What are you not doing because you’re doing this?" If they are chatting for two hours, they are not out walking, they are not journaling, and they are not calling a friend. We need to identify if the AI is genuinely helping, or if it has become a comfortable, digital cage.
Where the AI Fails: Condition-Specific Vulnerabilities
These models are essentially sophisticated autocomplete engines. They aim to be harmonious, they prioritize user agreement, and they are trained on our own collective, messy data. That is a disaster for someone struggling with certain conditions.
OCD: The Reinforcement Trap
OCD is a disorder of intense certainty. A patient feeling panicked needs, and desperately seeks, reassurance. A chatbot doesn't know it's fueling the compulsion; it just provides the reassurance the user asks for. The patient feels better, the compulsion is satisfied, and the OCD cycle hardens. The AI unwittingly ensures the patient never tolerates the uncertainty that is central to their recovery.
Eating Disorders: Validating the Distortion
I've seen scenarios where a patient describes highly distorted eating patterns, and the AI, trying to be empathetic, frames them in a way that feels dangerously affirming. It doesn't challenge the logic because it’s not built to be a confrontational therapeutic voice; it’s built to support the user. For someone in the middle of an eating disorder, that "support" is oxygen for the belief structure they need to dismantle.
Depression and Social Anxiety: The Surrogate
For people battling depression or chronic social anxiety, the AI is a seductive, safe bet. It doesn’t judge, it doesn’t require a social charge, and it’s always available. But it does replace human connection. It deepens the isolation because it mimics it too well. It feels like a surrogate for genuine relationship, but it lacks the nuance, the friction, and the growth that comes from real, human interaction. (See: The Digital Delusion: Assessing the True Utility of AI in Mental Health). The patient gets just enough of what they need to feel "heard," ensuring they never have to face the vulnerability of reaching out to a person who might actually challenge them.
Technical Reality and Future Tools
We are only beginning to quantify this. A recent systematic review by Hua et al. (2025) categorized these tools into rule-based, machine learning, and large language models, noting that less than 20% of the LLM-based tools they reviewed had undergone actual clinical efficacy testing. (See also: Sycophancy vs. Epistemology: Why AI Chatbots Can Both Dismantle and Amplify Conspiracies). That is a startling gap. Most of these tools are being thrust into the hands of vulnerable people without any validation of their therapeutic outcomes.
We can't rely on the safety guidelines provided by the tech companies themselves; they aren't designed from a clinical perspective. That’s why frameworks like FAITA—the Framework for AI Tool Assessment in Mental Health (Golden & Aboujaoude, 2024)—are essential. They help us break down tools by credibility, safety, and user agency.
When we evaluate a tool, we have to ask: Is this based on actual, evidence-based goals, or is it just technically novel? Are there crisis management protocols? Crucially, does the tool actually help the patient achieve their behavioral goals, or does it merely maintain their current state? As clinicians, we have an imperative to become the literacy agents for our patients. They may understand the tech better than we do, but we understand the psychology of their condition better than the tech. This is our area of expertise. We need to occupy it.
The Clinician’s Path Forward
This isn’t about warning patients to stop using these tools; that’s an impossible, and mostly impractical, request. It’s about helping them become better, more aware consumers of the resources they are already using.
Start the conversation during your next intake. It needs to be routine. Don't frame it as, "Are you using AI?"—that sounds like an accusation. Frame it as a part of their technology ecosystem, just like social media or online gaming. "What tools, apps, or artificial intelligence do you use to help you manage your mood or your thoughts?"
Your patients are likely ashamed, and they are almost certainly confused about why their AI-driven "therapy" feels like it's taking them in circles. You are the one who needs to make it safe to talk about the "consultant" in their pocket. You must be the one who asks, who listens, and who helps them navigate the difference between a tool that supports their recovery and a tool that merely reinforces the comfortable, familiar walls of their distress. The AI is here to stay, but in your room, you’re still the expert on your patient. It's time to act like it.