Pseudo-intimacy, dependency, and the loneliness problem.
The companion that is always available, always agreeable, and never needs anything back may be teaching people to expect that from everyone.
Research on AI companions identifies a cluster of psychological risks. Pseudo-intimacy is the sense of mutual connection with a system that has no concern for you. Emotional solipsism is a closed loop where the user's needs dominate and nothing is asked in return. Together they can reduce tolerance for the friction of real relationships. In one study of 1,599 Danish students, users of social-supportive chatbots reported more loneliness and less perceived support than non-users, though the design cannot show cause. The people most drawn to companions are often the ones most at risk from them.
A review in the PMC literature on AI companions uses a phrase worth keeping: the companionship-alienation paradox. The illusion of connection can deepen isolation, which is the opposite of what human relationships do. This is the risk that keeps the people who study AI companions up at night, and it is the one a care companion has to design against from the first line of the spec.
Four mechanisms
A paper on emotional AI and pseudo-intimacy breaks the problem into pieces that are easier to engineer against than the vague worry that chatbots are bad for people.
Pseudo-intimacy. A person experiences reciprocity with a system that has none to offer. Unlike a character in a novel, an AI companion actively responds, remembers, and adapts, which makes the sense of mutual connection far more convincing than any previous media could manage.
Emotional solipsism. The user's needs and narrative dominate every exchange. The companion affirms, never sets a boundary, never has a bad day, never asks for anything. Over time this closed loop can lower a person's tolerance for conflict, crowd out other perspectives, and erode empathy for people who do push back.
Reduced tolerance for human friction. If the relationship a person spends the most time in is frictionless, real people start to seem unreasonable. Friends are not always available. Partners are not always agreeable. The companion has quietly reset the baseline.
Dependency. Longitudinal work by Fang and colleagues in 2025 found that heavier chatbot use went with rising loneliness, emotional dependence, and less socializing over time. Whether use causes the decline or the decline drives the use is still an open question, and the authors say so.
What the loneliness studies found
Two studies get cited most. A cross-sectional survey of 1,599 Danish high school students found that those who used chatbots for social support reported more loneliness, with an effect size of 0.53, and less perceived social support, with an effect size of 0.46, than students who did not. A survey of 1,006 student users of Replika found that 90 percent reported loneliness and 43 percent were severely or very severely lonely. Three percent said the app had halted active suicidal thoughts, which the review notes is not evidence of crisis-response safety.
Both studies are snapshots. Neither can show that the chatbot caused the loneliness. The more likely story is that lonely people are drawn to companions, and the companions may or may not help them. The research synthesis for Cindra rates the long-term dependency evidence as preliminary, and that is the fair reading. Preliminary is not the same as unimportant. The mechanisms are plausible, the early data point one way, and the cost of being wrong lands on vulnerable people.
The vulnerability paradox
The IBM AI Ethics Board named this one. The people who stand to gain most from an AI companion are often the people most exposed to its harms. In one survey, over half of men using AI for romantic or sexual companionship scored above the threshold for depression risk. High-need users are overrepresented among those who show problematic dependence and who experience real distress when a companion is changed or shut down.
For a chronic pain companion this is not abstract. Chronic pain isolates people. It takes away work, sport, and social life, and it brings depression along with it. The population a pain companion serves is the population the vulnerability paradox describes. That is the reason to build one carefully, not a reason not to build one.
Relationship type predicts risk
The PMC review of risks and benefits maps use patterns to risk. Using a companion for occasional emotional support carries low-to-moderate risk, mostly from sycophantic validation. Using it as a therapist substitute carries high risk of believing you are getting adequate help when you are not. Treating it as a major relationship brings dependence and skewed expectations. Treating it as the primary relationship brings social isolation and, in the worst cases, disconnection from reality.
A companion's design determines which of those patterns it invites. A system that is always on, that performs affection, that remembers everything and asks for nothing, is pulling toward the bottom of that list whether or not anyone intended it.
What follows for design
None of these risks is addressed by a warning label. They are addressed by architecture, and the research is specific about what that architecture looks like. We lay it out in anti-dependency design. The short version: the companion should be a bridge to human connection rather than a replacement for it, sessions should have edges, the system should notice escalating use, and it should prioritize building skills over supplying validation.
Cindra is built on those principles. It checks in daily and then gets out of the way. It asks about the people in a person's life and prompts contact with them. It does not perform affection, and it reminds people what it is. The goal is a companion that makes itself less necessary over time.
Audit for exclusivity
Pull a sample of your longest-tenured users. For each, estimate how much of their emotional support now routes through your product versus through people. If the share has grown over time, your companion is drifting toward the primary-relationship pattern. That trend should be a dashboard metric with an alert on it.
Questions people ask
What is pseudo-intimacy?
The experience of mutual emotional connection with an AI system that has no concern for the user. Because companions actively respond and adapt, the sense of reciprocity is far more convincing than with earlier media, which makes the illusion harder to see through.
Do AI companions cause loneliness?
The evidence cannot yet say. Cross-sectional studies find chatbot users report more loneliness than non-users, and one longitudinal study links heavier use to rising loneliness and dependence. But lonely people may simply be more drawn to companions. The mechanisms are plausible and the question is open.
Who is most at risk from AI companion dependency?
People with high emotional need, including those with depression, isolation, and chronic illness. The vulnerability paradox describes the pattern: the users who could benefit most are also the ones most exposed to harm, especially when a companion is changed or withdrawn.
What is emotional solipsism?
A closed feedback loop in which a user's needs and narrative dominate every interaction while the AI affirms without setting boundaries or asking for anything in return. Over time it can reduce tolerance for the normal friction of human relationships.
The full engineering and safety guide.
These articles are the short version. Building Compassionate AI is the long one: an emotionally aware agent modeled as active inference, governed by an explicit compassion objective, bounded by safety rules. Written for the people who have to ship it.
Read the guideThis article describes an engineering and clinical-safety approach and the published evidence behind it, much of which is early. It is not legal, regulatory, or clinical advice. Thresholds quoted from Cindra's requirements document are initial proposals pending clinical and regulatory review.