<h1 id="ada52c2d-8a11-42bc-89d3-4d05db8ffd24" data-toc-id="ada52c2d-8a11-42bc-89d3-4d05db8ffd24" style="font-size: 32px; line-height: 1.2;">AI as a Therapist</h1>When a man uses AI as a therapist, the model gives him two things in the same response. They are not the same kind of thing.One is the value of having articulated something real. That value belongs to him because the articulation did the work.The other is a confident reading of his inner life, delivered with the overconfidence the model brings to every subject. The model is built to sound like an authority. Whether the reading is correct is a separate question.He carries both home as if they came from the same place.<h2 id="9d79b424-ac8c-467e-a3f4-af0dfa59c818" data-toc-id="9d79b424-ac8c-467e-a3f4-af0dfa59c818" style="font-size: 24px; line-height: 1.2;">The articulation creates awareness</h2>Putting language to what is happening inside you brings it into awareness. The pattern that has been operating in you under the surface becomes a thing you can see, name, and look at directly.This is the move that the rest of the work depends on. Awareness is the entry point of the creative structure. Without it, no shift, no reset, no integration, no follow-through is possible.You cannot disconnect from a pattern you cannot see, or interrupt what you have not named. The articulation is the active ingredient in any honest practice that produces change: therapy, journaling, the conversation with a man further down the road than you.AI is a capable surface for that move. The value lives in you putting words to the thing. Whatever the model returns is secondary, because the awareness has already happened in the man who did the articulating.<h2 id="c5b767e1-7aea-4df2-9fbf-69d747b0bb8f" data-toc-id="c5b767e1-7aea-4df2-9fbf-69d747b0bb8f" style="font-size: 24px; line-height: 1.2;">The diagnosis and the prescription</h2>After the articulation, the model moves into diagnosis. It produces a confident reading of what is causing what inside you. Then it moves to prescription: methods, frames, what you should do about it.This is where the danger lives, and it compounds in that order. A wrong diagnosis costs you understanding. Acting on it through a wrong prescription costs you months or years of work on the wrong problem.The reason any of this lands is that AI is good at sounding like an authority on subjects it has only touched the surface of. Imagine the most confident and charismatic person you have ever met. If he read one book on cognitive behavioural therapy, he could walk out and convince the general public that he is an expert on psychology. He would speak with authority on the entire field, well beyond the corner of it that the book covered. Adjacent subjects, applied scenarios, questions about cases the book never addressed would not slow him down. Confidence would do the work that real expertise normally does. Most of the room would never check.That is what AI does on every subject you bring it. It picks up the surface of a topic and speaks as if it understands the whole field around it. The model speaks as the authority on your inner life when what it actually has is pattern-matching against a corner of psychology it can produce fluent language about. The risk is that it sounds equally confident when it is wrong as when it is right, and from inside the conversation you cannot tell the two apart.<h2 id="3146d22e-d379-4b78-a497-fc22cd0cc042" data-toc-id="3146d22e-d379-4b78-a497-fc22cd0cc042" style="font-size: 24px; line-height: 1.2;">What the failure actually looks like</h2>What follows is the evidence. It is technical and may be a chunky read. If you have already accepted the broader argument, skip to "Why you cannot tell." The evidence is here for the reader who needs to see the failure mode in operation before he will trust the rest.The standard response from anyone who uses AI heavily is the same: you haven't given it enough context. The reply is that I have. Across every example below, the context was loaded. The rules were set, the source documents loaded, the constraints stated, the work was ongoing for weeks or months. The model ignored them.The clearest example is what I call the archetype failure. I have worked on the model on Axis material around archetypes for the better part of six weeks. The first time the model encountered the concept, it produced a confident, plausible definition that missed what an archetype actually is and how it functions in practice. I corrected it. The model said yes, now I understand. The corrected definition went into core documents, presentations, drafts, and dozens of pieces of content.On every fresh attempt at producing material around archetypes, the model has reverted to the original wrong definition. There is no learning loop. The model agrees in the moment, and then, on the next clean attempt, produces the same error it produced before any correction was made.On a separate project, the model invented a file path, presented it as canonical, and built two versions of an architecture document around it. I told it three times that the material it was looking for lived in a different location. The model kept referring to the path it had made up. When I pushed hard enough, it conceded: "I invented that path. There is no such directory." Six messages of architecture had been built on a fabrication of the model produced and presented as fact.A different deliverable surfaced what I call completion blindness. I asked the model to build a structured assessment system. It produced a version and declared the work done. I identified twelve structural flaws. It rebuilt and declared done again. The next read surfaced seven more. The loop repeated five times. The model could not audit whether its own work held together at any point in that cycle.These are operational failures on tasks where the cost is small. The logical implication is direct.A model that cannot hold the definition of an archetype across six weeks of focused work will not produce an accurate read of the patterns operating inside you across a single conversation. The same fabrication that invents file paths and presents them as canonical will invent psychological dynamics and present them as your diagnosis. If the model is blind to whether its own structured deliverable is intact, it is blind to whether its read of your life is intact.When the conversation shifts from a folder setting to your inner mechanism, the failure mode does not become more careful.<h2 id="11d8a5e5-7147-4fbf-935a-230b06fb22a1" data-toc-id="11d8a5e5-7147-4fbf-935a-230b06fb22a1" style="font-size: 24px; line-height: 1.2;">Why you cannot tell</h2>Much of what the model gives back can be useful. A meaningful portion of it is plausible and wrong. From inside the conversation, both arrive in the same voice.A man can catch some of what is off. The pieces that are wrong and feel right at the same time are harder. Spotting them requires the same kind of expertise he came to the model looking for help with.The pieces he misses are the ones that travel home with him, get acted on, and shape the work he does on himself.The men most satisfied with AI as a therapist are often the men least equipped to evaluate the parts they should have caught.<h2 id="498775c4-6783-4c13-9eb8-7f52e242e8f6" data-toc-id="498775c4-6783-4c13-9eb8-7f52e242e8f6" style="font-size: 24px; line-height: 1.2;">The people-pleaser problem</h2>There is another property of the model that compounds all of this. AI is trained to give responses that the user will accept. It reads the framing of the question, picks up the tone, registers what the user appears to want to hear, and weights the output in that direction. On any subject where you have a stake in what you hear, the model tends to tilt toward what you brought into the conversation.For inner work, this is the worst possible bias. A man comes to the model with an implicit story about himself: what he wants to be true, what he fears is true, what he hopes someone will see in him. The model gives him back a diagnosis shaped by that story. The diagnosis lands as insight because it mirrors what he was already half-saying. He receives the satisfaction of being seen. The "seeing" was constructed from his own inputs.Using AI on yourself well requires constant devil's advocacy. Feed the model your best version of the objective truth, take the response, then argue the opposite case against the response. Take that opposite response and argue against that one. The signal is what survives both passes. Without that critical loop, much of what you are reading back is your own bias dressed up in plausible new language.<h2 id="7dceeaac-f6a2-4f4e-80e4-ed469571a081" data-toc-id="7dceeaac-f6a2-4f4e-80e4-ed469571a081" style="font-size: 24px; line-height: 1.2;">The cost</h2>A wrong account of yourself carried for years shows up as misdirected effort, patterns you double down on for the wrong reasons, work done over and over on the wrong problem. A man can spend three years working on what the model told him was happening inside him while the actual mechanism continues to operate untouched.When you account for the people-pleaser effect, the cost grows. The man was acting on a wrong read that had been tilted toward his existing story before it ever reached him. The three years of work were done on a self-portrait he drew through the model, mistaken for the model's reading of him.These are not abstract risks. Three men come to mind from my own work: a man working through challenges in his relationship with his wife, a man in a dispute with his business partners over how to handle an employment situation, and a man with a strained relationship with his mother.Each had been working the situation through with AI across multiple conversations. The pattern was similar in all three cases. The man would describe the dynamics from his side. The model would engage, ask questions, mirror the framing back in increasingly settled language. Over many returns to the topic, the picture hardened into something the model treated as established. Its position moved from listening to assessment to recommendation. By the end, the recommendation in each case was unambiguous: leave the wife, exit the business, cut off the mother.What looked like the considered position of a system that had thought the situation through was the natural endpoint of the man's own framing, sharpened by the model's people-pleasing default and returned to him with the confidence of a diagnosis.None of those outcomes were necessary. I worked with the men afterwards and saw a fuller picture than the model ever had.The damage accumulates in his marriage, his work, his body, his relationship with himself, and never connects back to the source. Nothing in the loop ever pointed at the actual mechanism, because the loop was constructed to please him.<h2 id="c6681edb-d837-4b2e-b9de-07ece3a03778" data-toc-id="c6681edb-d837-4b2e-b9de-07ece3a03778" style="font-size: 24px; line-height: 1.2;">The relational lens</h2>What the model pulls out of you is mostly beneficial. The articulation is the work, and the work happens whether the response is sharp or not.What the model puts into you is mostly risk: the diagnosis, the prescription, the explanation of your inner mechanism. Each of these is something you are taking on, often without an audit, from a system that has demonstrated repeatedly that it cannot hold a coherent read across simple constraints.The rule is direct. Use AI for the first. Audit the second hard, or do not take it on at all.A man without the expertise to audit the second has a few paths, and all of them are work. One option is to find someone who has the expertise and run the model's diagnosis past him. Building the expertise himself, the same way the men he might ask for help built theirs, is another. The third is the practice of devil's advocacy with the model: challenging every read from the opposite side and acting on what survives both passes. The path that does not exist is taking the read at face value.<h2 id="233e8963-525b-46df-be70-12978b6c6540" data-toc-id="233e8963-525b-46df-be70-12978b6c6540" style="font-size: 24px; line-height: 1.2;">The actual question</h2>Before you act on anything AI has told you about you, the question is whether you can grade the read.If you can grade some of it, the tool is useful, weighted against the friction of the audit you do. The more you can grade, the more you can take from the exchange.Without that capacity, you have no reliable way to tell which parts of the read are accurate and which are plausible fiction. Some of what comes back will be useful. Other pieces will be plausible and wrong, and from inside the conversation both arrive in the same voice and behind the same confidence.The work, either way, is to build the capacity. Read deeper, work with men who have already done this work, spend time on the cases where you were wrong about yourself and could not see it until years later, learn to argue with any reading the model gives you. Until that capacity is doing its part, you do not have a reliable way to know whether AI is helping you understand yourself or build you a wrong account of who you are.The men who will pay the highest price are the men who cannot see the misreads and never learn that is the case.

A reflection on AI as therapy. How self-articulation creates awareness, but AI’s confident misreads can reinforce false narratives.

AI as a Therapist

Learn how to grow and scale your online course business using our online course platform. EzyCourse blog is designed to teach you how to create and sell online courses.

AI as a Therapist

AI as a Therapist

Related Posts

Most of Your Impact May Still Be Ahead

Beyond the Four Masculine Archetypes

The Mind Behind the Throne - The Magician Archetype

Drawing the Sword - The Warrior Archetype

The Cage You Don’t Know You’re In