A four-week randomized trial of 981 adults found no measurable difference in loneliness, social contact, emotional dependence, or problematic use between the chatbot setups it compared, such as text versus voice or open-ended versus personal conversation. What did separate people was how much they chose to use the chatbot: heavier voluntary users had worse results on average, which shows an association and does not show that chatting caused the change. Regulators treat a chatbot sold as therapy differently from a general-purpose assistant.
Key takeaways
- The trial by researchers at the MIT Media Lab and OpenAI analyzed more than 300,000 messages from 981 participants over four weeks.
- Assigned conditions (text, neutral voice, engaging voice; three conversation types) produced no significant effects on the four outcomes measured.
- Higher trust in the chatbot and stronger social attraction to it went with more emotional dependence and problematic use.
- In the United States, the FDA's advisory committee met on November 6, 2025 about generative AI mental health devices, a category separate from everyday chat tools.
- Evidence from one short trial cannot settle whether companion-style chatbots help or harm emotional well-being.
What the four-week trial tested
The study, posted to arXiv in March 2025 and revised in October 2025, was led by Cathy Mengying Fang with co-authors from the MIT Media Lab and OpenAI. Participants were randomly placed in one of several groups. Some used text, some a neutral-sounding voice, and some an engaging voice. Within each mode, the conversation type was open-ended, non-personal, or personal. Researchers then measured four outcomes: loneliness, socialization with other people, emotional dependence on the chatbot, and problematic use.
Random assignment is the feature that makes this design stronger than a survey. It allows a comparison between groups that differ only by the setup they were given. The size of the message set, more than 300,000 messages, also allowed the team to code how the chatbot and the person each behaved in the conversations.
What came out of it
The authors state that no significant effects were detected from the experimental conditions. Giving someone a voice instead of text, or steering the conversation toward personal topics, did not shift the four outcomes in a measurable way over four weeks.
The second finding comes from watching what people did on their own. Participants who used the chatbot more, whatever their assigned condition, had consistently worse outcomes. Higher trust in the chatbot and greater social attraction toward it were associated with more emotional dependence and problematic use. The direction of cause is open. A person who is already lonely may spend more time chatting, and heavy use may also deepen isolation. A trial that randomizes the setup but not the amount of use cannot separate the two.
Why one trial does not settle the question
- Duration: four weeks says little about habits that form over a year or more.
- Product scope: the results describe the chatbot and settings used in the study, not every companion app.
- Measurement: scales for loneliness and dependence compress complex feelings into scores, and a score is not a diagnosis.
- Selection of heavy users: those who used the tool most may differ from others before the study began.
- Population: the participants were adults who agreed to join a study, so teenagers and people in crisis are not represented by these figures.
When a chatbot is meant to treat a condition
Software that claims to diagnose, treat, or prevent a mental health condition is a different regulatory case from a chat assistant used for general conversation. The US Food and Drug Administration's Digital Health Advisory Committee held a public meeting on November 6, 2025 devoted to generative AI-enabled digital mental health medical devices. Law-firm summaries of the meeting report that the discussion centered on a hypothetical prescription chatbot for adults with major depressive disorder, and that panelists addressed premarket evidence, monitoring after release, plain-language labeling, and escalation to a human clinician. The same summaries note that the agency had not yet cleared a mental health tool built on generative AI at the time of the meeting. The committee gives advice only. It does not set rules.
At the state level, Illinois signed the Wellness and Oversight for Psychological Resources Act on August 1, 2025. According to the state's Department of Financial and Professional Regulation, it bars the use of AI to make mental health and therapeutic decisions, while allowing AI for administrative and supplementary support to licensed professionals.
Reading claims about AI and emotional health
| Type of evidence | What it can show | What it cannot show |
|---|---|---|
| Randomized trial of chatbot setups | Whether a design choice changes outcomes | Effects beyond the study length or product |
| Observation of heavy users | Which patterns go with worse outcomes | Whether use caused those outcomes |
| Company or app testimonials | That some users report benefits | How many were not helped or dropped out |
| Regulator committee meeting | Which questions authorities consider open | A finding that any product is safe or unsafe |
This article offers general information about research and regulation. It is not medical advice and does not diagnose, treat, or recommend a product. Anyone struggling with mood, loneliness, or thoughts of self-harm should talk with a licensed clinician or local crisis service, and questions about a specific app belong with its regulator or a qualified professional.
FAQ
Did the trial show that chatbots make people lonelier?
No. Assigned conditions had no significant effect on loneliness. Heavier voluntary use went with worse outcomes, but the study design cannot tell whether use caused them or lonelier people simply used the chatbot more.
Does the FDA approve AI therapy chatbots?
The FDA reviews software that is intended to diagnose, treat, or prevent disease as a medical device. Its advisory committee discussed generative AI mental health devices on November 6, 2025, and reports from that period said none had been cleared. Status after that date was not confirmed for this article.
Is talking to a chatbot the same as therapy?
Not in legal terms. Illinois, for example, restricts AI from making therapeutic decisions, while general conversation tools are not licensed as treatment.
Related reading
- How conversational interfaces evolved
- A profile of the Character AI service
- Human and machine intelligence compared
- Health education and promotion basics
More in AI in health.
Last verified: October 2026
Sources
- Fang CM et al., How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study, arXiv, 2025
- Digital Health Advisory Committee meeting, November 6, 2025, US Food and Drug Administration, 2025
- FDA's Digital Health Advisory Committee Considers Generative AI Therapy Chatbots for Depression, Orrick, November 2025
- Governor Pritzker signs legislation prohibiting AI therapy in Illinois, Illinois Department of Financial and Professional Regulation, August 4, 2025