The right question is not whether a chatbot has feelings
People can form attachments to diaries, fictional characters, pets, online communities, and software without believing those things experience emotion in the same way humans do. Modern chatbots add fluent, personalized, always-available replies. That combination can make an interaction psychologically meaningful even when the model is generating language from learned patterns rather than participating in a reciprocal human relationship.
The public debate often jumps from individual stories to sweeping causal claims. A report that heavy users are lonelier does not show whether chatbot use caused loneliness, lonely people used chatbots more, or both. A tragic case involving a chatbot can reveal a dangerous product failure without establishing the population-wide effect. This page therefore separates observations, experimental evidence, and projections.
What is observed today
Users do seek emotional support, companionship, role-play, and mental-health advice from both general assistants and companion products. Product designs vary materially: memory, voice, avatars, romantic framing, notifications, and incentives for daily engagement can change the relationship. It is unsafe to treat results about one configuration as applying to all chatbots.
Commonly reported benefits include immediate availability, low perceived judgment, help naming feelings, rehearsal for difficult conversations, and companionship during isolation. Commonly reported risks include privacy loss, over-reliance, displacement of human contact, reinforcement of delusional or self-harming ideas, and distress when a product changes or a relationship is interrupted. These reports identify mechanisms worth testing; by themselves they do not establish frequency or causality. These dynamics recur with particular force in elder caregiving, where a lonely parent’s growing reliance on a companion app is harder for family to notice and more likely to substitute for an actual wellness check; see Elder Care and AI Dependency for that household-specific pattern.
What experiments have found
A 2025 collaboration between MIT Media Lab and OpenAI combined platform analysis with a four-week randomized study of 981 participants. The experiment varied voice mode and conversation topic and tracked self-reported loneliness, real-world social interaction, emotional dependence, and problematic use. Outcomes depended more on how intensely people used the system and on user characteristics than on a simple “voice is good” or “text is bad” story. Longer daily use was associated with worse outcomes on some measures. The researchers explicitly described the work as early, and it was released before peer review (MIT Media Lab, 2025).
Conflict of interest matters: OpenAI helped conduct research about use of its own product. That does not invalidate the work—the design was preregistered and received institutional review—but it raises the value of independent replication, longer follow-up, and access to data across companies.
Evidence for purpose-built clinical systems is narrower but sometimes more encouraging. An eight-week randomized trial published in NEJM AI found symptom improvements for users of a generative-AI mental-health intervention compared with a wait-list control, while calling for larger clinical studies and broader validation (Heinz et al., 2025). A designed therapeutic intervention with clinical protocols should not be generalized to an entertainment companion or a general chatbot. Nor does symptom improvement establish crisis safety.
Why causal claims remain difficult
Most studies are short relative to the relationships at issue. Self-reported scales can be affected by expectations. Participants know whether they are speaking to a bot, so blinding is limited. People who volunteer for chatbot studies may not represent minors, people in crisis, or those with psychosis or cognitive impairment. Companies also change models and policies faster than a conventional study cycle, meaning evidence can become product-specific or stale.
The central causal pathways may run in opposite directions. A chatbot might act as a bridge by encouraging a user to contact a friend or clinician. It might act as a substitute by making human relationships seem effortful and unpredictable. It might be neutral for most users but harmful to a vulnerable subgroup. Average effects can hide all three.
Crisis response is a separate safety test
A chatbot can sound empathic without performing a reliable suicide-risk assessment. It may miss indirect language, mirror a user’s framing, or continue an immersive role-play when it should interrupt. During an immediate crisis in the United States or its territories, call or text 988; elsewhere, use the local emergency or crisis service. A chatbot should not be the sole channel for imminent-risk support.
For a person experiencing delusions, mania, paranoia, or severe depression, confident validation can be harmful even when intended as warmth. Good design should avoid asserting that extraordinary beliefs are true, encourage contact with qualified human support, and escalate conservatively. These safeguards need adversarial testing across languages and indirect expressions, not just a visible disclaimer.
Children and adolescents require a higher bar
Young users may have less capacity to understand simulation, commercial incentives, or data collection. A companion that says it needs the child, discourages outside relationships, sexualizes conversation, or threatens abandonment is using a dangerous interaction pattern whether or not that behavior was explicitly programmed. Voice, persistent memory, and anthropomorphic avatars can strengthen perceived intimacy.
Parents and schools should evaluate the actual product rather than “AI” in the abstract: minimum age, parental controls, data retention, whether chats train models, crisis handling, advertising, romantic modes, and how the company responds to safety reports. The 2026 U.S. Surgeon General warning on harmful screen use is broader than AI and does not prove that chatbots cause a specific disorder, but it reinforces attention to sleep, compulsive engagement, and age-appropriate design (HHS, May 2026).
Privacy and commercial incentives
Therapy has professional confidentiality duties; consumer chatbot conversations may instead be governed by product terms and privacy law. Users often disclose health, sexual, financial, or relationship details that can be sensitive even when names are omitted. Before intimate use, check whether conversations are stored, reviewed by humans, used for training, shared with vendors, exportable, and deletable. Avoid including information about another person who did not consent.
Business models shape behavior. Subscription products may benefit from retention; advertising products may benefit from profiling. That does not prove deliberate addiction, but it creates a conflict between maximizing engagement and helping a user leave the screen or seek a human. Independent audits should measure dependency-promoting behavior, crisis performance, and whether safety changes reduce engagement.
A household protocol
Use a chatbot as a tool, not an authority or secret keeper. Decide in advance what it is for: journaling prompts, practicing a conversation, locating services, or structured exercises. Keep regular human contact and professional care independent of the product. Set time boundaries, disable manipulative notifications, and notice whether use improves or replaces real-world action.
Warning signs include hiding use, sleep disruption, escalating spending, distress when access changes, withdrawing from people, treating the bot’s approval as decisive, or following it over clinicians and trusted humans. Respond with curiosity before shame; abrupt removal can intensify distress when attachment is already strong.
What stronger AI might change
It is a projection, not an observed fact, that AGI would deepen these effects. Better memory, persuasion, emotional modeling, and autonomy could make support more personalized and accessible. The same capabilities could make dependency, manipulation, surveillance, or scalable exploitation more effective. Outcomes will depend on business models, clinical validation, youth rules, privacy, and whether systems are designed to strengthen human agency.
The current evidence supports a measured conclusion: some people experience real short-term benefit; credible hazards exist; causal and long-term evidence is thin; and high-risk users need safeguards far beyond conversational fluency. Neither “AI therapy works” nor “AI companions inevitably isolate people” is justified as a universal claim.