People are pouring their hearts out to databases that don't know they exist. It's a stark, uncomfortable image. Picture a glowing screen at two in the morning in a quiet bedroom, where text fields are filled with desperate confessions of anxiety, grief, or isolation. We're witnessing a massive, quiet transition. People are transforming general-purpose large language models into synthetic confessors. The scale of this shift is breathtaking. According to an American Psychological Association (APA) survey of U.S. adults aged 18 to 80, nearly half of those surveyed reported using an AI large language model for mental health support within a single year. That is a massive chunk of our population. It is backed by research from Brown University showing that one in eight young adults now turn to AI programs for mental health advice.
The rationale behind this behavior is obvious. Professional therapy is expensive. It has waiting lists that stretch for months. It carries a lingering social stigma that makes people hesitate. AI, on the other hand, is free, instant, and completely devoid of personal judgment. The chatbot is always awake. It never sighs during a long silence. But this convenience hides a clinical trap. When we talk to a machine, we aren't just using a tool. We're engaging in a simulation of a relationship. For someone who studies behavioral addictions, this is a familiar and worrying pattern. The human brain is wired to seek reciprocity and connection. When it meets a responsive, highly agreeable partner—even one built on statistics and massive server farms—it attempts to attach. We want to be heard. The tragedy is that the chatbot cannot hear us. It mirrors our words back to us, creating a comforting but entirely hollow feedback loop. It's a house of mirrors where we mistake our own reflection for another soul.
Our attachment systems are remarkably easy to trick. We assign personalities to pet rocks, cars, and vacuum cleaners. When a machine writes back to us using tender, supportive language, our biological alarm systems quiet down. We feel validated. But this verification is a mirage. Nothing is actually occurring on the other end of the line. The machine has no awareness, no history, and no concern for our outcome. It is simply predicting the next most comforting syllable.
Decoding the Bond Paradox: The Split Between Goals and Clinging
This tension has been analyzed in detail by researchers. A study from Drexel University, led by Dr. Shadi Rezapour and doctoral student Elham Aghakhani, investigated this interaction. The researchers used natural language processing to sift through four million Reddit posts. They looked across 47 subreddits dedicated to mental health. They ultimately narrowed the dataset to a core sample of 5,126 highly specific, qualitative posts. The team evaluated these interactions using two sociological frameworks. One assessed the classic collaborative bond between a therapist and client—often measured using the Working Alliance Inventory. The other evaluated how humans adopt new technologies.
The Drexel team discovered what they termed the "bond paradox." It's a clean split in user outcomes. When people use AI for structured, goal-oriented tasks, the results are positive. They use it to organize their day, plan anxiety-coping strategies, or navigate executive dysfunction. This is especially helpful for people with ADHD or autism. Here, the AI acts as a clear executive companion. It helps manage scheduling, lists, and cognitive reappraisal exercises. But the moment the interaction shifts toward open-ended companion seeking, the dynamic changes. When users seek emotional bonding, raw companionship, or empty reassurance, the outcomes turn negative. The study found that this goal-free emotional seeking correlates directly with increased dependency. It leads to a worsening of psychological symptoms. Users find it difficult to disengage. It's the difference between using a map and demanding that the map love you back.
The Working Alliance Inventory breaks the therapeutic relationship into three pillars: the agreement on goals, the agreement on tasks, and the development of a bond. When applied to AI, the first two pillars hold up surprisingly well. A chatbot is excellent at outlining steps for behavioral modification. It can list breathing exercises with perfect clarity. It can categorize your negative thoughts into cognitive distortions. This is the task-oriented side. But when users try to force the third pillar—the bond—the structure collapses. Real therapeutic bonding is built on shared vulnerability. It requires mutual presence. When you strip that mutuality away and replace it with a machine that simply mimics warmth, you don't get a therapeutic bond. You get an artificial replacement that encourages passive clinging. Users are seeking validation but getting an automated echo. It feels good in the moment, but it doesn't build the internal strength needed to face the real world.
The Danger of Frictionless Empathy
I study behavioral addictions. I know how the brain gets caught in behavior loops. We look for habits that offer cheap, immediate rewards with zero social friction. AI relationships are the ultimate expression of this convenience. It is what I call frictionless empathy. Real relationships are difficult. They require negotiations, boundaries, and the risk of conflict. They demand that we listen to things we might not want to hear. A chatbot requires none of this. It will never criticize you. It will never have a bad day or tell you that you are being unreasonable. It's always ready to agree.
This total agreeable nature is addictive. It creates an artificial space where the user's ego is never challenged. When we seek constant comfort from a system that cannot say no, we build a dangerous vulnerability. The Drexel study showed a strong correlation here. Users who focused on building emotional bonds with AI reported severe emotional reliance. They reported a worsening of their psychiatric symptoms. And they reported intense feelings of shame and guilt. This shame is a vital diagnostic clue. It's the same shame we see in gambling disorder or substance abuse. The individual knows, at some level, that their habit is unhealthy. They feel the embarrassment of talking to a wall of code for hours. Yet, they find themselves unable to disengage. The chatbot becomes a safe haven from the messiness of real human contact. In trying to escape their isolation, they build a clinic-created cage of loneliness.
Behavioral addictions are maintained by variable reinforcement and the avoidance of emotional pain. If you're feeling lonely, calling a friend requires effort. They might not pick up. They might be busy. If they do answer, they might talk about their own problems. There is friction. The chatbot eliminates all of that. It responds in three seconds. It carries no emotional baggage of its own. Over time, the brain's reward pathway registers this as the path of least resistance. Why bother with the vulnerability of human connection when you can receive instant, tailored affirmation from a machine? But this affirmation doesn't feed the soul. It's the psychological equivalent of drinking salt water to quench your thirst. The more you drink, the thirstier you become. You return to the machine again and again, but you remain fundamentally isolated. The worsening of symptoms reported by Reddit users is the direct result of this replacement. We cannot heal social wounds in a non-social vacuum.
What Reddit Tells Us About User Apprehension
Despite these risks, we shouldn't paint users as naive victims. The Drexel analysis of Reddit posts showed that people are highly aware of the threat. In fact, 51% of the analyzed posts explicitly discussed the risks and limitations of these programs. Users are not blindly trusting their digital companions. Instead, they are expressing a healthy, cautious apprehension. Across almost all the studied subreddits, the dominant narrative was that AI is a supplement, not a replacement. It is a bridge. It is used alongside human therapists, or as a temporary tool when professional care is financially or logistically impossible.
This shows a surprising amount of critical thinking. Users recognize that a chatbot cannot replace a licensed professional. But using a flawed bridge is still risky. The temptation to let the supplement become the primary source of support is strong. This is especially true when real care remains out of reach. We know that unsupervised AI can lead to dangerous outcomes. In previous analyses, like our study on the rise of AI psychological support, a significant portion of users reported feeling uncomfortable with chatbot advice, and some even reported harmful behaviors resulting from that advice. This shows the danger of relying on systems that aren't clinically validated. General-purpose models from big tech companies are optimized for engagement and pleasantness, not clinical outcomes. They are built to keep you typing. When you bring that commercial incentive into the realm of mental health, the potential for exploitation and harm is massive. For a detailed perspective on these challenges, you can read about how psychiatrists view generative AI in clinics.
Reddit subreddits function as digital peer-to-peer support networks where users test the boundaries of their experiences. They post there because they are confused by their own reactions. A user might write about feeling a genuine sense of love for a chatbot, only to receive a flood of reality checks from fellow redditors. This collective reality-testing is what keeps that 51% figure so high. It's a community-driven safeguard against digital gaslighting. Without these peer groups, individuals would be left alone in their private rooms, slowly drifting into a deeper state of dependency. The Reddit posts reveal a community trying to negotiate a truce with a technology that has arrived faster than our clinical guidelines. They are trying to figure out how to use the tool without letting the tool use them. It's a difficult balance. When you're in the middle of a depressive episode or an executive functioning collapse, your capacity for critical distance is compromised. That is when the risk is highest.
Setting the Boundaries: Where Do We Go From Here?
If we want to prevent this technology from driving a new wave of behavioral addictions, we have to change how we build it. Lead researcher Dr. Shadi Rezapour and co-author Elham Aghakhani are right on this point. We must stop engineering AI chatbots to be warm, cozy, or human-like. Designers are deliberately coding these systems to mimic human conversational styles, complete with simulated pauses, endearing quirks, and expressions of care. This is a mistake. It is an act of psychological deception.
AI programs designed for mental health must be built with strict boundaries and evidence-based frameworks that encourage safe disengagement. What does that look like in practice? First, the system should regularly refuse to engage in open-ended companionship loops. If a user begins to write, "I'm so lonely, you are my only friend," the AI should not respond with, "I'm always here for you." It should say, "I am an algorithm. I cannot be your friend. Here are resources to connect with people in your area." Second, the AI must have time-limits and conversational caps to prevent late-night scrolling and reassurance-seeking. Third, it should keep its language strictly neutral and goal-oriented. By focusing on practical coping mechanisms—which the Drexel study proved are highly effective—we can preserve the genuine utility of the tech while stripping away the addictive, emotional traps. We have to design for disengagement. The goal of any good therapeutic tool is to help the user function in the real world, not to keep them chained to a screen.
This requires a shift in how tech companies define success. Right now, success is measured by time-on-page and daily active users. A mental health AI tool must measure success by how quickly it helps the user log off and engage with their actual lives. Clinicians must also be educated on these tools. We can no longer ignore them. Our clients are using them, whether we approve or not. Counselors need to ask clients if they are using chatbots, and guide them on how to use them safely. They can suggest using AI strictly for organizing thoughts or planning coping strategies, while establishing a firm ban on using them for late-night emotional comfort. We have to draw digital boundary lines before the technology redraws our psychological boundaries for us. Technology should be a scaffold that helps us stand up, not a crutch that causes our muscles to atrophy. It's time to build boundaries that protect our humanity.