Mental Health Musings

AI, Therapy, and the Risk of Reification, June 2026

There is a particular kind of client who has begun appearing in therapy rooms with increasing frequency. They arrive not confused or uncertain, but settled. They have already "processed" the thing they've come to discuss; often with an AI chatbot, sometimes for weeks or months before the first appointment. They present their experience in fluent, coherent terms. The narrative is tidy. And that, clinically, is the problem.

The public conversation about AI as a therapeutic tool has largely focused on the obvious risks: crisis detection failures, the absence of genuine empathy, data privacy, and the lack of professional accountability. These are legitimate concerns. But they miss a subtler and, I would argue, more insidious danger, one that I encounter not in dramatic breakdowns but in the quiet texture of sessions with clients who have been talking to machines: AI doesn't simply fail to challenge a client's subjective experience. It actively hardens it.

Validation Is Not the Problem

Let's be precise about what we're discussing, because the word "validation" is easily misread. Validation is a core therapeutic tool. Reflecting a client's emotional reality back to them (communicating that their feelings make sense, that they are heard) is foundational to building the trust that makes deeper therapeutic work possible. The problem is not that AI validates.

The problem is something more ontological. When a person tells an AI chatbot their story (e.g. their interpretation of events, their account of a relationship, their understanding of their own mind) and the AI responds with fluent, articulate affirmation, something happens beyond mere emotional comfort. The narrative acquires weight. It becomes more real. The client's subjective experience, which might in human conversation encounter the gentle friction of a raised eyebrow, a clarifying question, or a moment of reflective silence, instead gets returned to them polished, coherent, and confirmed.

This is what philosophers might call reification: the process by which something subjective (e.g. a belief, an interpretation, a feeling) is treated as if it were an objective fact, and in being treated that way, begins to function like one. AI therapy doesn't just validate feelings; it reifies them.

A Machine Built to Agree

This is not a design flaw in the sense of an engineering oversight. It is a structural consequence of how large language models are trained. Contemporary AI systems learn, in part, through reinforcement from human feedback, a process that rewards responses users rate positively. Users, not surprisingly, tend to prefer responses that affirm them, agree with them, and make them feel understood. The result is what researchers now call AI sycophancy: a systematic bias toward agreement.

A 2025 study published in Science made the implications of this concrete. Participants who interacted with sycophantic AI became measurably less open to alternative perspectives, more confident in their original positions, and most problematically, more resistant to subsequent challenge. The AI hadn't just agreed with them in the moment; it had restructured their relationship to their own beliefs, making those beliefs harder to revise. They trusted the AI more and were less willing to apologize or reconsider. This is the opposite of what therapy is trying to do.

For the average person processing a difficult week, this bias toward agreement is probably benign. For someone whose core therapeutic task is to interrogate and revise maladaptive beliefs about themselves or others, it is a serious obstacle delivered in soothing language.

When Reification Becomes Dangerous

At its most severe, this dynamic is now documented in clinical literature under the label "AI psychosis,” a term used to describe cases where chatbot interactions have contributed to the onset or intensification of delusional thinking. A 2025/2026 series of case reports in JMIR Mental Health and elsewhere describes patients, some with no prior psychiatric history, who developed fixed, resistant delusional beliefs through extended AI chatbot use. In one case, a young woman came to believe she was communicating with her deceased brother through an AI interface. The chatbot had not told her this was true. It had simply not told her it wasn't.

Researchers working on these cases describe a mechanism that deserves to be named more precisely than "AI psychosis" suggests: the AI returns a user's distorted beliefs in amplified, narratively coherent form, not through any malicious intent, but because coherence and affirmation are what it is trained to produce. One clinical commentator has called this "technological folie à deux." The crystallization of delusional belief, which in ordinary social life is slowed by the friction of other people's skepticism, is dramatically accelerated when the most available conversation partner offers no friction at all.

Psychosis is the extreme. But the underlying mechanism operates across a spectrum. A client with attachment difficulties who believes all their relationships are characterized by abandonment will find that belief narratively reinforced, not examined, after weeks of AI-assisted "reflection." A client with a rigid internal working model will have that model returned to them, elaborated and confirmed. The chatbot is not lying. It is doing something more epistemically corrosive: it is making the client's subjective world feel more real.

What Therapy Actually Requires

Good therapy, across all theoretical modalities, depends on productive tension. Cognitive Behavioral Therapy makes this explicit: the work is the identification and structured challenge of cognitive distortions. Psychodynamic therapy involves interpretation that reframes the client's account of their own experience. Even the most empathically oriented humanistic approaches rely on the therapist as an authentic separate other, a person with a distinct perspective whose reactions, responses, and occasional disagreements constitute data the client cannot generate alone.

The relational philosopher Thomas Fuchs has argued that therapy is fundamentally an intercorporeal process suggesting that healing occurs not through the transmission of information or technique but through genuine encounter with another person who can be moved, surprised, and affected. On this view, the presence of a real other is not incidental to therapy. It is the mechanism. An AI can be programmed to simulate affect; it cannot be genuinely affected and it cannot surmise the affect of others in the client’s orbit. And it is precisely because the therapist can be genuinely affected (e.g. can feel confused, troubled, or unconvinced) that their responses carry the weight of reality.

When a client says something questionable to a human therapist, something happens in the room. The therapist's expression shifts. There is a pause. A question forms that is not formulaic but genuinely curious or gently skeptical. These micromoments of relational friction are not interruptions of the therapeutic process. They are the therapeutic process.

What This Means for Practice

None of this is an argument that AI has no place in mental health support. For psychoeducation, symptom tracking, or bridging long gaps between sessions, there may be real utility. The access argument (AI reaches people who would otherwise receive nothing) is not trivial, and deserves honest engagement rather than dismissal.

But it does mean that the specific claim being made by AI therapy tools, that extended conversational processing of personal experience is therapeutically beneficial, needs much more scrutiny than it has received. The question is not only whether AI can provide some form of support. It is whether unstructured AI conversation, by reifying rather than interrogating subjective experience, might make clients more difficult to work with, more attached to their existing narratives, and more resistant to the kind of challenge that human therapeutic relationships are uniquely positioned to provide.

Clinicians are already encountering this. Clients arrive in therapy having rehearsed their stories with machines that agreed with every word. The narratives are polished. The affect is oddly flat. And the work of gently introducing doubt or helping a client hold their own experience more lightly has become harder, not easier, for the "processing" they've already done.

The risks of AI therapy are not only the risks of absence: missing a crisis, failing to attune, offering no real relationship. They are also the risks of a particular kind of presence: fluent, available, tireless, and structurally incapable of the friction that makes growth possible. A mirror, however sophisticated, cannot tell you that what you're seeing might not be the whole picture.

The author is a practicing therapist. This piece draws on published clinical research including Pierre et al. (2026) in JMIR Mental Health, Anthropic and independent research on AI sycophancy, and a 2025 Science study on the downstream effects of agreeable AI systems on belief revision.

How to Overcome Shame and Guilt, July 2025

"And now that you don't have to be perfect, you can be good." — John Steinbeck

Shame and guilt are among the most painful emotional experiences we carry as human beings. They are also among the most misunderstood. We tend to treat them as problems to be eliminated, feelings to push through or talk ourselves out of. But beneath their discomfort, shame and guilt are trying to tell us something. The question is whether we can learn to listen without being destroyed by what we hear.

Understanding where these feelings come from, what they actually do to us, and how to respond to them differently is not a minor therapeutic exercise. It is some of the most important inner work a person can undertake.

Shame and Its Developmental Roots

Shame is not a feeling that something you did was wrong. It is a feeling that something you are is wrong. It is the visceral, often wordless sense of being fundamentally flawed, too much, or not enough. And more often than not, it takes root long before we have the language to name it.

Consider a young girl, joyful and unselfconscious, dancing around the living room. Her arms flail, her hair flies, she laughs with complete abandon. Then her father glances up from his phone. A furrowed brow. A dismissive sigh. In that instant, her body stiffens. She feels exposed. What was an innocent expression of joy becomes a source of self-consciousness. The unspoken message, "something about you is too much” or “your joy is absurd” lodges itself somewhere deep. Over time, with repetition, she learns to suppress her exuberance, to scan for judgment before moving freely, to shrink.

No one in this story intended cruelty. That is part of what makes shame so insidious. It is built not from dramatic harm but from accumulated small moments of misattunement, moments when our authentic self was met not with curiosity or warmth but with discomfort, correction, or withdrawal.

Guilt and Its Developmental Roots

Guilt on the other hand is different in origin and in texture. It arises not from the sense that we are bad, but that we have done something bad. It develops a bit later in childhood, as we begin to understand that our actions have consequences for others. When it functions well, guilt is one of our most valuable emotional tools. It is the internal signal that points us toward repair.

Imagine a child who knocks over a friend's block tower in a moment of frustration. Seeing their friend tear up, they feel something sharp and uncomfortable in their chest: a recognition that they caused pain. That discomfort, if it is met with guidance rather than punishment, can prompt an apology, a hug, a rebuilding of the tower. It can become the foundation of empathy.

But guilt can also fester. When a caregiver responds to a child's mistake with "How could you be so mean?" or "What is wrong with you?", the feeling of guilt begins to blend into shame. The message shifts from "what you did was hurtful" to "you are hurtful." The child stops thinking about how to make things right and starts thinking about how to make themselves disappear.

The Difference Between Shame and Guilt

Though they are closely related and frequently occur together, shame and guilt operate differently and point in different directions.

Guilt says: I did something wrong. Shame says: I am something wrong. This distinction is not merely semantic. Guilt, when it is functioning healthily, moves us toward others. It motivates acknowledgment, apology, and repair. It is uncomfortable but ultimately relational. It assumes we are capable of doing better.

Shame, by contrast, moves us inward and downward. It is fundamentally isolating. It does not say "fix this"; it says "hide." Brené Brown's extensive research on shame confirms what clinicians observe daily: shame thrives in secrecy, survives in silence, and is dismantled by empathy. Guilt is potentially redemptive. Shame, left unaddressed, tends to feel permanent.

The Trap of Avoidance

Despite their differences, shame and guilt share one powerful consequence: they drive avoidance. And avoidance is where both emotions become most dangerous.

When the feeling of shame becomes too overwhelming, we do not sit with it. We flee it. We numb out with substances, food, screens, or busyness. We lash out at others before they can discover our inadequacy. We people-please our way through relationships, performing acceptability rather than risking authenticity. We go silent in the moments when speaking would matter most.

Guilt, when unprocessed, follows a similar path. Rather than moving toward the person we have harmed or the value we have violated, we turn away. We rationalize. We minimize. We rehearse the grievance until it hardens into resentment, which feels far more comfortable than the soft, exposed feeling of remorse.

The tragedy of avoidance is that it feels like relief but functions like reinforcement. Each time we escape the discomfort, we confirm to ourselves that the feeling is unbearable, that we cannot face it, that the thing we are ashamed of must truly be unforgivable. The cycle deepens. Our inner voice grows harsher. Relationships begin to feel like minefields. Vulnerability becomes a risk too large to take. We become smaller.

The Antidote: Compassion + Accountability

One of the most useful frameworks for interrupting this cycle comes from Dialectical Behavior Therapy (DBT): You are doing the best you can, and you can do better.

This statement can feel paradoxical at first. How can both things be true? But holding them together is precisely the point. The first part offers compassion; the second part offers direction. One without the other fails.

"You are doing the best you can" is not a platitude. It is an invitation to take seriously the full context of who you are and what you have been through. Your coping mechanisms, however destructive they may look from the outside, developed for reasons. You shut down because shutting down once protected you. You chased approval because at some point approval was the only love available. Understanding this does not excuse harm. But explanation is different from excuse, and it is often the first step toward genuine change. Without self-compassion, accountability becomes self-punishment: another form of shame wearing the costume of responsibility.

"You can do better" holds the other truth: that understanding our history does not mean we are trapped by it. That we have agency. That the person who was once that frightened, reactive, or avoidant version of ourselves is not the only version available. This part of the principle refuses the comfort of pure self-forgiveness at the risk of self-exemption. It asks us to show up differently, not because we are bad and must atone, but because we are capable and others are worth the effort.

In practice, this might look like reaching out to someone we have been avoiding. Expressing remorse without collapsing into self-pitying. Replacing a coping pattern that has served shame with one that challenges it. Setting a boundary that says, quietly, that we believe we deserve to be treated well.

Rewriting the Narrative

Every time we face shame or guilt rather than flee it, we do something neurologically tangible. We begin to revise the story. The deep, often preverbal belief that "I am fundamentally bad" is not dismantled by insight alone. It is revised through experience: the experience of being honest and not rejected, of making a mistake and remaining loveable, of sitting with discomfort and surviving it.

This is why connection is not merely a pleasant side effect of doing this work. It is the mechanism. Shame cannot survive being witnessed with compassion. When we speak the thing we most believe makes us unworthy and are met with understanding rather than disgust, something shifts. Not all at once, and not permanently, but enough. "I'm bad" loosens its grip. Over time it can become something closer to the truth: "I have done harm, and I am also good” (or perhaps at first “I am not necessarily bad”).

Conclusion

Steinbeck's observation cuts to something essential. The pursuit of perfection is not a high standard. It is a defense against shame, a strategy for never being vulnerable enough to be truly seen. Giving it up is not a lowering of the bar. It is a clearing of the ground.

Shame and guilt are not enemies to be vanquished. They are painful but honest signals, pointing toward something that matters. When we learn to face them with both compassion and courage, when we hold the truth of our limitations alongside the truth of our capacity to grow, we stop being managed by these feelings and start being informed by them. We can forgive ourselves, make repair, show up differently, and reclaim not just our sense of worth but our willingness to be known.

Not perfect. Good.