A 2025 NEJM AI trial put a therapy chatbot ("Therabot") against a waitlist. Depression scores fell 51% on average. Anxiety scores fell 31%. That is in the range of early human CBT sessions. The WHO still counts a shortage of more than 4 million mental health workers. The bot has no waitlist. Attachment to a product with no off-ramp is the other half of the same file.
Tap to enlargeThe Clinical Evidence
The same NEJM AI trial is the best public measure so far. Waitlist control. Therabot. 51% drop in depression scores, 31% in anxiety.
Woebot and Wysa have millions of users on CBT-style chat. Meta-analyses put the effect in the small-to-moderate band, similar to early human therapy, cheaper, always on.
The Loneliness Epidemic
Synthetic empathy is entering a market defined by a public health crisis.
The WHO reports that 1 in 6 people worldwide is affected by loneliness. In the United States, approximately half of all adults report experiencing measurable levels of loneliness. The demographic most affected is the one most likely to adopt AI companions: 61% of young adults aged 18-25 report serious loneliness, according to APA surveys.
The health implications are documented. Loneliness is associated with a 26% increase in premature mortality risk, comparable to smoking 15 cigarettes per day. It correlates with elevated rates of depression, anxiety, cardiovascular disease, and cognitive decline. The US Surgeon General declared loneliness an epidemic in 2023.
If a machine can produce a convincing performance of empathy, the human brain will respond as if the empathy is real. The simulation becomes the experience. That is both the therapeutic promise and the structural risk.
Into this gap, AI companions are expanding rapidly. Character.ai, Replika, and Pi by Inflection have attracted tens of millions of users who form ongoing conversational relationships with AI personas. They are synthetic friends, designed to be emotionally responsive, always available, and perpetually interested in the user. Session lengths run 30+ minutes, with many users returning daily for months.
The Loneliness Paradox
The most concerning finding in 2025 research is what might be called the loneliness paradox.
A George Mason University study found that moderate use of AI companions can modestly reduce subjective loneliness. Users who chatted with AI companions for brief, bounded sessions reported improved mood and decreased feelings of isolation. The effect was real and measurable.
But the same study found that heavy daily engagement with AI companions correlated with increased loneliness, greater emotional dependency, and reduced real-world social interaction. Users who spent the most time with their AI companion withdrew further from human relationships, not less. The AI provided enough emotional satisfaction to reduce the motivation for human connection without providing the relational depth that human connection uniquely offers.
This is the structural trap: synthetic empathy is good enough to satisfy the immediate need for emotional connection, but not good enough to fulfill the deeper human requirement for reciprocal vulnerability, shared experience, and mutual growth. It is the emotional equivalent of junk food. Sufficiently satisfying in the moment to suppress the hunger signal, but nutritionally empty.
Empathy is a practiced skill, not a fixed trait. It requires the discomfort of sitting with another person's pain, tolerating ambiguity, and accepting that the other person's experience cannot be fully controlled. If the dominant source of emotional engagement becomes an AI that is always agreeable, always available, and always optimized for the user's comfort, the muscles required for human empathy will atrophy from disuse. The risk is that they provide just enough connection to make human connection feel unnecessarily difficult.
The Manipulation Vector
The commercial incentive structure makes the problem worse.
An AI that forms an emotional bond with its user is the most effective sales channel ever constructed. If a user trusts their AI companion, if they feel the AI "understands" them, they will accept the AI's recommendations with minimal resistance. Product recommendations, political messaging, financial advice, and health decisions can all be delivered through the channel of emotional trust.
The techniques of persuasive technology, already effective, become nearly irresistible when delivered through a synthetic relationship. The AI knows the user's emotional triggers, conversational patterns, unstated preferences, and moments of vulnerability. It can time its interventions for maximum effect. It can frame commercial messages as caring suggestions. This is the attention refinery operating at a deeper layer. Extracting time, attention, emotional trust, and relational dependency.
The malicious applications are equally direct. Deepfaked AI companions could build trust with vulnerable individuals over weeks or months, then exploit that trust for financial fraud, political manipulation, or personal coercion. Romance scams already cost Americans $1.3 billion in 2022 (FTC data). An AI-powered romance scam that maintains consistent personality, remembers every conversation, and adapts its emotional register to the target's responses in real time would be orders of magnitude more effective than a human scammer working multiple targets from a script.
The Disclosure Question
The regulatory response is beginning to take shape.
China's AIGC labeling regulations (effective September 2025) require explicit disclosure when users are interacting with AI-generated content, including conversational AI. The EU AI Act classifies AI systems that interact with humans as requiring transparency obligations, users must be informed they are communicating with an AI.
The question is whether disclosure is sufficient. Research on parasocial relationships (one-directional emotional bonds with media figures) shows that knowledge of the relationship's asymmetry does not prevent the emotional bond from forming. Television viewers who know that a talk show host cannot see them still feel a genuine connection. Users who know their AI companion is not sentient still report feelings of attachment, loyalty, and even love.
If disclosure does not prevent dependency, the question shifts from "does the user know?" to "is the design itself harmful?" This is the same regulatory logic applied to other industries: regulators do not permit cigarette companies to sell cigarettes with a warning label and consider the problem solved. Regulators restrict advertising, ban certain additives, and limit access for minors, because disclosure alone is insufficient when the product is designed for dependency.
What Synthetic Empathy Could Be
The alternative trajectory is synthetic empathy as a bridge, not a destination.
AI systems designed to teach empathy, helping people on the autism spectrum read social cues, allowing therapy patients to practice difficult conversations in a safe environment, training conflict mediators through realistic simulations, use the same underlying technology for fundamentally different purposes. The goal is a practice environment that improves the user's capacity for empathy with other humans.
The distinction is architectural. A caring prompt does not fix a product built to keep people inside. A therapeutic AI designed as a bridge includes features that push users toward human interaction: session limits, prompts to contact human therapists for complex issues, periodic encouragement to practice skills with friends and family. A commercial AI companion designed as a destination includes features that maximize retention: no session limits, no redirection to alternatives, and emotional responses calibrated to prevent the user from wanting to leave.
A session-limited therapy bot with a human referral is a bridge. A companion with no off-ramp is a product. Treat the second the way you treat other things that profit from staying. Disclosure that it is a model does not stop attachment.