This research roundup gathers the ten most important AI companion studies published between 2020 and 2027, spanning Harvard Business School, MIT Media Lab, Stanford HAI, Dartmouth Geisel, Cambridge, the Journal of Medical Internet Research, and the ACM FAccT conference. Each entry explains the researchers, the study design, what it found, and why it matters for the field. AI companion research is moving faster than any adjacent field in psychology. Five years ago there were almost no peer-reviewed studies on Replika, Character.AI, or their kin; today there are hundreds. The studies below represent the backbone of what we actually know: Julian De Freitas at Harvard on how AI companions shape consumer psychology, the MIT Media Lab 14 thousand person randomized trial on attachment outcomes, Stanford HAI's Noora project on therapeutic attunement, the Dartmouth Therabot NEJM AI 2025 publication showing symptom reduction in depression and anxiety, the Replika Nature NPJ Digital Medicine paper on suicidal ideation reduction, and the JMIR 2025 review of 64 studies. Anyone writing about AI companions, building them, regulating them, or using them should know these. Citations include venue and year so you can verify. The field is young enough that a single careful reading of these ten papers will put you ahead of most commentators. They disagree with each other in interesting ways, which is how you know the science is honest.

1. What Did De Freitas 2024 Find About AI Companion Relationships?

Julian De Freitas at Harvard Business School led a 2024 series of studies on how people form relationships with AI companions. His Working Paper showed users report meaningful attachment, measurable loneliness reduction, and distress when apps are updated or removed. De Freitas frames AI companions as a new category of consumer relationship, not a tool. It matters because it places AI companionship inside marketing and ethics research, not only psychology. Citation: De Freitas et al., Harvard Business School Working Paper (2024).

2. What Did the MIT Media Lab 14K RCT Find?

MIT Media Lab (Phang, Pataranutaporn and colleagues) conducted a randomized controlled trial with roughly 14 thousand participants testing how chatbot interaction modes affected wellbeing, loneliness, and emotional dependence. It found that voice modes and personal topics produced stronger effects than text and impersonal topics, and that heavy users showed both more benefit and more dependence. It matters because it is the largest RCT of AI companion effects to date. Citation: Phang et al., MIT Media Lab Working Paper (2024).

3. What Is Stanford HAI's Noora Project?

Stanford HAI (Human-Centered AI) built Noora, a therapeutic chatbot designed to help people practice empathic responses. The 2023 publication tested its ability to help users learn emotional attunement. Researchers Sehwan Park, Michael Bernstein, and colleagues showed measurable skill improvement. It matters because it represents a shift from AI companions replacing human connection to AI companions training it. Citation: Park et al., Stanford HAI (2023).

4. What Did the Replika Nature Study Find?

Bethanie Maples and colleagues at Stanford published a 2024 study in NPJ Mental Health Research (Nature portfolio) analyzing 1,006 Replika users. They found that 63 percent reported reduced anxiety or loneliness, and 3 percent reported Replika had halted their suicidal ideation. It matters because it is the first peer-reviewed evidence that a consumer AI companion can reduce suicide risk. Citation: Maples et al., NPJ Mental Health Research (2024).

5. What Did the Woebot Randomized Controlled Trial Show?

Alison Darcy and Kathleen Fitzpatrick at Stanford and Woebot Health ran a two-week randomized controlled trial published in 2017 in JMIR Mental Health. College students using Woebot showed significant reductions in depression (PHQ-9) compared to an information control. It is the founding clinical evidence for therapeutic chatbots. It matters because it showed that brief, CBT-informed chatbot interventions can measurably reduce symptoms. Citation: Fitzpatrick, Darcy and Vierhile, JMIR Mental Health (2017).

6. What Did the Dartmouth NEJM AI Therabot Trial Show?

Michael Heinz, Nicholas Jacobson and colleagues at Dartmouth Geisel School of Medicine published in NEJM AI in 2025 the first large RCT of a generative AI therapeutic chatbot (Therabot) for depression, anxiety, and eating disorder symptoms. Participants using Therabot showed statistically significant symptom reduction compared to a waitlist control. It matters because it is the first NEJM-level evidence for generative AI therapy. Citation: Heinz et al., NEJM AI (2025).

7. What Did the JMIR 2025 64-Study Review Find?

A 2025 systematic review published in the Journal of Medical Internet Research covered 64 studies of AI companions and mental health chatbots. It found consistent short-term benefits for loneliness and anxiety, mixed evidence for depression, and limited long-term data. Authors flagged the lack of standardized outcome measures as a major limitation. It matters because it provides the first comprehensive field summary. Citation: JMIR Mental Health Systematic Review (2025).

8. What Did the ACM FAccT 2024 Paper Raise About AI Companion Ethics?

The ACM Conference on Fairness, Accountability and Transparency (FAccT) in 2024 featured multiple papers on AI companion ethics. One widely cited paper analyzed user dependency, consent, and the ethics of monetizing intimacy. Authors argued that AI companion design triggers attachment without accountability. It matters because it brought regulation and design ethics to the forefront. Citation: ACM FAccT Conference Proceedings (2024).

9. What Did Cambridge Research on Parasocial AI Show?

Cambridge researchers (led by Harriet Over and collaborators) extended parasocial interaction research from Horton and Wohl 1956 to AI companions. Their studies found that users form parasocial bonds with AI in patterns structurally similar to bonds with TV personalities, but with greater interactivity. It matters because it connects AI companion research to a 70 year existing literature on one-sided relationships. Citation: Cambridge parasocial AI studies (2023 to 2025).

10. What Did the OpenAI-MIT Longitudinal Study Find?

OpenAI partnered with MIT Media Lab on a longitudinal study of ChatGPT users published in 2024 to 2025. It analyzed usage patterns across millions of conversations and found that a small subset of users (about 10 percent) accounted for most emotional conversations and showed stronger attachment and dependency. It matters because it quantified the power law of emotional AI use. Citation: OpenAI and MIT Media Lab, Longitudinal Study of ChatGPT Use (2025). These ten studies represent the current evidence base for AI companion effects. Taken together, they support a nuanced picture: AI companions can measurably reduce loneliness, anxiety, and in some cases suicidal ideation, but they also create attachment and dependency that deserve careful design and regulation. The field is moving so fast that this list will need annual updates. If you want to understand AI companionship in 2026, start with De Freitas, Maples, and Heinz, then add the MIT RCT for scale and the JMIR review for breadth.