A few months ago, I had a client sit down and do something a little strange at my suggestion. In session, using a structured prompt we had prepared together, she entered a narrative built around her feared self into an AI language model. The model returned a story.
The output read like her OCD had written it.
Not because the model was broken. Because she and I had just discovered something genuinely interesting: the architecture of a large language model and the architecture of the OCD loop are doing something strikingly similar. They are both narrative engines trained to produce the most contextually plausible next thought given an initial input. And seeing that from the outside, watching the machine do what her mind had been doing for years, was the beginning of something shifting.
A Brief Introduction to How LLMs Actually Work
You do not need to be a technologist to understand this. Large language models like the one you may be using right now work through a process called next-token prediction. Given an input (a question, a sentence fragment, a fear) the model calculates which words are most statistically probable to follow. It has been trained on enormous amounts of human text, which means it has internalized the narrative patterns humans use when they write about danger, worst-case scenarios, catastrophe, and threat.
Feed it fear, and it generates fear-consistent narrative. That is not a flaw. That is the design.
The result is a loop:
Both systems are doing something functionally similar: taking an initial input rooted in doubt or threat, and generating the most contextually coherent elaboration of that input. OCD has been linked to differences in CSTC circuitry (cortico-striato-thalamo-cortical circuits). In an LLM, the elaboration happens in transformer layers trained on vast amounts of human writing, including a great deal about threat, danger, and worst-case scenarios. Both produce output that feels authoritative. Neither is accessing ground truth.
OCD is not a broken threat detector. It is a narrative engine that has been fed a fear prompt and is doing exactly what it was built to do: generating the most plausible, internally consistent elaboration of that threat. The problem is not the output. The problem is that the person cannot yet see the machine.
Katherine Illgen, LCSWWhy This Matters for Treatment: The I-CBT Frame
Inference-Based Cognitive Behavioral Therapy (I-CBT), developed largely by Frederick Aardema and the researchers affiliated with the Institut Universitaire en Sante Mentale de Montreal, reframes OCD not as an anxiety disorder but as a disorder of doubt. Specifically, it identifies the mechanism as inferential confusion.
The core problem in OCD, from this perspective, is not that the person is anxious. It is that they have come to distrust their own perception and instead rely on an internally generated "possibility." That obsessional doubt overrides sensory and experiential evidence. The person knows, on some level, that the stove is off, that their hands are clean, that they are not the kind of person who would harm someone they love. But the obsessional inference, the OCD narrative, overrides that knowing.
This is where the LLM parallel becomes clinically useful. When a client enters a prepared prompt into a language model and watches it produce an elaborate, coherent, distressing narrative, they are witnessing something important from the outside: a system that is generating plausible-sounding content based entirely on the logic of the input, with no access to reality, no ability to verify, and no special insight into what is actually true.
The externalization goal of I-CBT is to help clients see the OCD narrative as something separate from themselves. It is a narrative generated by a system, not a revelation of truth. Watching an LLM do the same thing in real time can make that distinction visceral and immediate in a way that weeks of psychoeducation sometimes cannot.
One of my clients described it this way after the exercise: "It's like I finally saw what was happening. It was just... generating. It wasn't telling me anything real. It was just doing what it does when you give it that kind of input."
That is the therapeutic moment. Not a technique. Not a worksheet. A direct experience of the mechanism.
This Works Whether You Trust AI or Not
One of the things I appreciate most about this framework is that it does not require the client to be enthusiastic about artificial intelligence. In fact, skepticism toward AI makes the parallel even more useful.
If a client pushes back on the comparison, that pushback becomes the intervention. "You don't trust the AI in your computer to tell you something real. Why trust the AI in your head?" The client who distrusts AI already understands, on some level, that a system can produce confident-sounding output that is not grounded in reality. They just haven't applied that understanding to their own OCD narrative yet.
For clients who do see value in AI tools, the parallel lands differently but just as usefully. They understand that AI is a tool. A sophisticated, sometimes impressive tool, but still a tool that generates output optimized for plausibility and coherence rather than accuracy. It is built to produce something that reads as convincing. Whether that output reflects reality is a separate question the system never asks.
Every major AI platform includes a version of the same warning: "Claude can make mistakes." Or: "ChatGPT can get things wrong." There is an irony in pointing to a checking instruction inside a piece about OCD, and it is worth naming: the disclaimer is not an invitation to verify. It is an acknowledgment that confident-sounding output and accurate output are not the same thing. AI generates narrative that feels authoritative because it has learned from authoritative human writing. OCD does the same. The obsessional narrative sounds certain. But certainty of tone is not evidence of truth. In neither system does the confidence of the output tell you whether it's right.
This is one of the more quietly powerful reframes available in I-CBT work. The client does not need to fight the OCD thought or prove it wrong. They need only recognize that the system generating it has no access to reality, and that confident output from an imagination-based engine is not, by itself, meaningful evidence of anything.
OCD Is a Doubting Disorder, Not a Feeling Disorder
Aardema's framing, elaborated in the Resolving OCD series and grounded in years of clinical research, positions OCD as fundamentally about doubt rather than anxiety. The anxiety is real, but it is downstream. It is the emotional response to a narrative that has already been generated and accepted as plausible.
This matters because much of traditional OCD treatment addresses the anxiety. ERP (Exposure and Response Prevention) is evidence-based and effective for many clients: through repeated exposures without compulsive response, clients test their feared predictions, disconfirm them experientially, and build new inhibitory associations that compete with the original fear response. But I-CBT asks a different question: what if we addressed the doubt itself? What if, instead of sitting with the anxiety, the client learned to distrust the narrative that produced it?
The LLM analogy extends here as well. We do not treat LLM hallucinations by helping the user develop a higher tolerance for bad outputs. We treat them by teaching the user to understand the system's limits, to recognize that the model is generating plausible text, not accessing truth. I-CBT takes the same approach with OCD: the goal is to help the client re-trust their own perceptual system over the OCD narrative engine.
The doubt is not evidence. It is output.
A Note on Doing This Carefully
If you are reading this as a clinician, you already know that anything with therapeutic potential also carries clinical risk. The same exercise that creates a powerful externalization experience for one client could function as a reassurance-seeking compulsion for another.
Clinical Guardrails for AI-Assisted Externalization
- This exercise is intended for use within an established therapeutic relationship, not as a self-help tool.
- Assess first whether the client is prone to reassurance-seeking compulsions in digital spaces. Using AI to generate feared scenarios and then repeatedly engaging with the output can reinforce rather than interrupt the loop.
- The goal is one observation, not a dialogue. The client enters the prepared prompt, reads the output, and pauses. They are not seeking reassurance. They are gathering data about the mechanism.
- Debrief immediately and thoroughly. The clinical value lives in the conversation afterward, not in the exercise itself.
- This is not a replacement for evidence-based OCD treatment. It is one tool, used intentionally, within a broader I-CBT or ERP framework.
- Clients should not use AI tools between sessions to process OCD content without clear, explicit parameters established in session first.
What This Points Toward
I am genuinely excited about the intersection of this work and continuing education. Not because AI is a solution, but because it offers something rare in OCD treatment: a way to make an invisible mechanism visible. The obsessional loop has always been something clients experience from the inside. It feels like reality because they are living inside it. Watching a machine perform the same loop from the outside creates the possibility of what I-CBT calls "distancing from the narrative," and it can happen in a single session.
There is more to explore here: the ethics of AI in clinical practice, the specific ways this approach interfaces with inferential confusion as a construct, and what a thoughtful protocol for this kind of work might look like. I am in the early stages of developing a continuing education framework around this, and I look forward to sharing more as it takes shape.
For now: if you work with clients who have OCD, consider telling them that their brain is not broken. It is an extraordinarily sophisticated narrative engine that has been given a fear prompt and is doing exactly what it was designed to do.
The goal is not to silence the engine. The goal is to see it.
