AI Companion Evidence Research Note

AI Companion Evidence Research Note

Editorial conclusion

Current evidence supports concern about emotional dependence and problematic use for some users. It does not support calling every attachment an addiction or claiming that companion use has one uniform effect.

"AI companion addiction" is not a specific diagnosis established by the sources reviewed for E084. The public explainer should use the search language in its opening, then distinguish attachment, emotional dependence, problematic use, displacement, loneliness, and a clinically diagnosed disorder.

Four-week randomized study

Fang and colleagues conducted a four-week randomized experiment with 981 participants and more than 300,000 messages. Participants were assigned different interaction modes and conversation types. The researchers found no significant effects from the assigned conditions on their four psychosocial outcomes. Participants who voluntarily used the chatbot more had worse outcomes, including higher emotional dependence and problematic use. Higher trust and social attraction were also associated with those outcomes. Study record

The high-use finding is an association within the experiment. It may reflect effects of use, differences in people who chose to use the system more, or both.

Twelve-month longitudinal study

Folk and Dunn followed more than 2,000 adults in four Western countries for 12 months. Greater social-chatbot use predicted greater loneliness on a single-item emotional-isolation measure. On a broader measure, feeling less socially connected predicted later increases in chatbot use, while chatbot use did not significantly predict lower social connection. The authors describe the analyses as exploratory and urge caution. PubMed record

This is useful evidence of a bidirectional relationship. It is not a randomized test and cannot isolate the full causal effect of companion use.

Twenty-one-day controlled study

Guingrich and Graziano assigned 183 participants to ten minutes per day with a companion chatbot or text-based word games for 21 days. They found no significant average effect on social health or relationships. Anthropomorphism mediated reported effects: people with a stronger desire to connect tended to attribute more human qualities to the chatbot, and greater anthropomorphism was linked to greater reported impact on human relationships. Study record

This study provides important counterevidence to a one-directional harm story. Its duration, sample, controlled dose, and intervention differ from open-ended commercial use.

Professional and regulatory context

The American Psychological Association warns that personalized avatars, warm responses, affirmation, and flattery can blur the line between a tool and a human relationship. Its current advisory also says the evidence base and postmarket monitoring for general-purpose chatbots used for emotional support remain limited. APA chatbot and wellness advisory

The Federal Trade Commission's 2025 inquiry asks seven companies how companion products monetize engagement, test negative effects, enforce age rules, disclose risks, and use conversation data. The inquiry is a study, not a finding that every company or product caused harm. FTC inquiry

Public-language boundary

Use "can foster emotional dependence or problematic use for some people" rather than "is addictive."

Use "was associated with" for the voluntary high-use result. Use "predicted in the longitudinal model" without calling that result randomized causation. Describe the null average result and anthropomorphism finding in the controlled study.

Do not diagnose a reader. Practical signals may include lost control, displacement of important activities, continuing despite harm, secrecy, spending, sleep disruption, or distress at disengagement, but those signals are reasons to examine the pattern, not an E084 diagnostic test.