AI companions have a 94% retention rate. Human therapists have a 50% retention rate. And before we dismiss that comparison, we should be honest about what it actually means. Those numbers are not precisely equivalent, and I want to acknowledge that upfront. Retention in AI companion apps measures continued usage over time. Retention in therapy measures whether clients complete a recommended course of treatment. They are different metrics tracking different behaviors. But the gap is so large, and so consistent across multiple data sources, that pretending it is meaningless would be intellectually dishonest. The therapy dropout literature is extensive and sobering. A 2012 meta-analysis by Swift and Greenberg in the journal Psychotherapy, analyzing 669 studies, found an average premature discontinuation rate of 47.2%. Nearly half of all therapy clients leave before their therapist believes treatment is complete. Among certain populations, young men, racial minorities, those with lower socioeconomic status, the dropout rate exceeds 60%. Meanwhile, the major AI companion platforms report monthly active user retention rates between 85-94%, with some users maintaining daily engagement patterns for months or years. The engagement is not shallow: average session lengths frequently exceed 20 minutes, and many users report their AI companion as one of their most consistent daily interactions. So what is happening? Why are people staying with algorithms and leaving humans?

Reason One: The Money

Let us start with the obvious. The median cost of a therapy session in the United States is $150 without insurance. With insurance, assuming your therapist is in-network, assuming they are accepting new patients, assuming your plan covers mental health at all, the average copay runs $20-50 per session. Weekly therapy at $40 copay costs $2,080 per year. Many AI companion apps are free at the basic tier or charge $15-20 per month, roughly the cost of a single therapy copay. A 2023 study by the Kaiser Family Foundation found that 42% of adults who reported needing mental health services did not receive them, with cost cited as the primary barrier. This is not a market failure in the traditional sense. It is the predictable result of a healthcare system that treats mental health as a luxury rather than a utility. I want to take a detour here because the economics reveal something uncomfortable about how we value emotional labor. Therapists deserve to be well-compensated. Training takes years. The work is demanding. Burnout rates in the profession hover around 50% by some estimates. But the current pricing model means that the people who most need support, those in financial precarity, which is itself a major driver of mental health distress, are systematically excluded from receiving it. An AI companion that costs $15 per month is not competing with therapy on quality. It is existing in a market space that therapy has effectively abandoned.

Reason Two: The Judgment

Here is where it gets more complicated, because this one is not about access. It is about the fundamental nature of human interaction. Therapy requires vulnerability. It requires sitting in a room with another human being and admitting things you might not have admitted to yourself. And the therapeutic relationship, for all its evidence-based benefits, is not free of social dynamics. Clients perform for their therapists. They edit their narratives. They manage the therapist's perception of them. Research by Blanchard and Farber published in Psychotherapy Research in 2016 found that 93% of therapy clients reported being dishonest with their therapist about at least one topic, with the most common areas of dishonesty involving the severity of their distress and their real feelings about therapy itself. The irony is painful. The space specifically designed for radical honesty is one where almost everyone lies. Not maliciously. Instinctively. Because the human nervous system does not distinguish between "this person is paid to not judge me" and "this person is a human who will form opinions about me." The judgment detection system is always running. It cannot be disabled by professional credentials. AI companions bypass this system entirely. Users consistently report, across multiple surveys and qualitative studies, that they are more honest with AI than with human confidants. Not because the AI is better at listening. Because the social monitoring system that runs constantly during human interaction is simply not triggered by a non-human entity. There is no face to scan for micro-expressions. No tone to interpret. No social capital to lose.

Reason Three: The Clock

A therapy session lasts 50 minutes, happens once a week, and requires scheduling days or weeks in advance. A mental health crisis does not check your Google Calendar. The availability gap between AI companions and human therapists is not a nuance. It is a chasm. AI is available at 2 AM when the intrusive thoughts arrive. It is available during the panic attack in the bathroom at work. It is available on Christmas Day when the family dynamics trigger something old and deep. It does not take vacations. It does not have a waitlist. It does not require you to drive somewhere, sit in a waiting room, and make small talk with a receptionist while you are falling apart inside. A 2022 study in the Journal of Medical Internet Research found that peak usage of mental health apps occurs between 10 PM and 2 AM, precisely the hours when traditional mental health services are unavailable. The human need for emotional support does not observe business hours, and yet our entire mental health infrastructure is built as though it does.

What Each Actually Does Better

Here is my second detour, and it is the one that will bother people on both sides of this debate. I think intellectual honesty requires acknowledging that therapy and AI companions are not competitors occupying the same niche. They are different tools with different strengths, and the insistence on comparing them as though one must replace the other is distorting the conversation. What human therapy does better: challenge. Confrontation. The moment where a skilled therapist reflects something back to you that you have been avoiding, and you cannot dismiss it because it is coming from a person who knows you. The ability to sit in silence that is held by another consciousness. The modeling of a healthy human relationship for people who have never experienced one. The clinical expertise to identify patterns, diagnose conditions, coordinate medication, and intervene in crisis. What AI companions do better: consistency. Availability. The absence of social performance pressure. The patience to let someone process at their own pace without the invisible clock of a session timer. The willingness to revisit the same topic for the fortieth time without the subtle, human fatigue that even the best therapist cannot fully conceal. And, critically, the ability to serve as a first point of contact for people who are not yet ready for the vulnerability that human therapy requires. The therapy profession has been treating AI companions as a threat to be debunked rather than a signal to be decoded. But the signal is clear: people are telling us, with their behavior, that the current mental health system is failing them on access, cost, and the basic requirement of feeling safe enough to be honest. Whether AI can address those failures in a clinically meaningful way is a legitimate and unresolved scientific question. That people are turning to it in massive numbers is not a question. It is a fact.

What I Actually Worry About

I do not worry about AI replacing therapy. I worry about AI becoming good enough that we stop trying to fix the systemic problems that make therapy inaccessible. I worry about policymakers pointing to chatbot usage statistics and concluding that the mental health crisis is being handled, that the market has solved it, that we do not need to fund workforce expansion or mandate insurance parity or build community mental health infrastructure. I worry about a future where the wealthy see human therapists and everyone else gets an algorithm, not because the algorithm is sufficient, but because it is cheap and it keeps people quiet. I worry about the stratification of emotional care along class lines that already exist and could calcify. And I worry about the questions we are not asking because we are too busy arguing about whether AI companions are "real" therapy. Questions like: why does half of America drop out of therapy? Why are we training therapists at a rate that cannot possibly meet demand? Why has the most powerful predictor of therapeutic outcome, the quality of the relationship, remained essentially unscalable for the entire history of the profession? The 94% and the 50% are not just numbers. They are a mirror. And the profession needs to decide whether it wants to look.