If you have followed the academic conversation on AI companions for any length of time, you have heard about the MIT Media Lab’s research on chatbots, emotional outcomes, and loneliness. It is, alongside the Stanford-affiliated Replika work, the lab name that gets cited most often by journalists trying to ground a story in something more solid than vibes.

This piece is the careful overview of that research program. What we can responsibly say it covers, what its methods look like in published work, what it has reported so far, and what it has not yet established. We are deliberately conservative about specifics. The full project is ongoing, the research output is spread across several papers and collaborations, and overconfident summary is exactly the failure mode this site exists to avoid.

If you only have a paragraph: the MIT Media Lab’s Affective Computing group and adjacent collaborators (in particular work associated with Pat Pataranutaporn, Pattie Maes, and others) have produced a body of work on how interaction with conversational AI relates to loneliness, emotional dependence, social functioning, and well-being. The most prominent recent output was a 2025 study run jointly with OpenAI on heavy ChatGPT users. The lab has also flagged a longitudinal program on companion chatbots specifically. The picture so far is mixed and careful: certain patterns of use look benign or beneficial, certain patterns look concerning, and the team has been more measured in framing than the headlines that summarize them.

What the project is

The MIT Media Lab is a research lab at MIT that publishes across many disciplines. The work most relevant to AI companions has come out of the Affective Computing group (which has a long history studying emotion-aware computing and human-AI interaction) and the broader “Advancing Humans with AI” research direction.

Within that broader program, several published papers and ongoing studies look at chatbot use and emotional outcomes. The most notable in 2025 was a collaboration with OpenAI examining heavy ChatGPT users and emotional outcomes; this was a population-level analysis on a general assistant, not a companion app, but the findings translate well enough to be discussed in companion-app coverage. Other Media Lab work has examined how the framing and design of an AI agent shape user perceptions, dependence, and loneliness, and how voice-based versus text-based interaction differs.

What the lab has signaled is a longer-running research direction on companion chatbots specifically, with longitudinal work meant to address the gap that most AI companion research is cross-sectional and survey-based. The Stanford-Replika study is a snapshot. The MIT program aims at outcomes over time. We are deliberately not citing a specific paper title or a specific dataset size for the longitudinal work because the full output is still rolling out and we want to avoid stating specifics we cannot verify against a primary source.

What the methods look like, in the published work

A few patterns worth flagging from work the team has actually published.

Mixed methods. The Media Lab’s published AI companion-adjacent work tends to combine controlled experimental design (randomly assigning users to one chatbot configuration versus another) with survey instruments measuring loneliness, attachment, and well-being. This is a step up from pure self-report studies but still has the limitations that any short-duration experiment has when applied to questions about long-term emotional outcomes.

Population-level analytical work. The 2025 ChatGPT collaboration with OpenAI took a different approach: large-scale analysis of usage patterns and self-reported emotional outcomes across a population of heavy users. This kind of work can identify patterns at scale that controlled experiments cannot. It also has its own limits, since it is observational and the self-selecting population matters.

Framing and design experiments. Several Media Lab papers have looked at how the framing of an AI agent (told it has emotions, told it is a tool) changes user perception and use. This is an underrated literature for understanding companion apps, since framing choices are exactly what differentiates Replika and Kindroid and Pi.

The general orientation across this body of work is that the Media Lab is more interested in mechanism than in advocacy. The publications read as careful empirical work rather than verdicts.

What the work has and has not yet established

A short version of where things stand.

What the work, taken together, supports: people use these tools for emotional purposes; certain patterns of use correlate with reduced reported loneliness; certain patterns of use correlate with increased reported emotional dependence; voice-based and text-based interaction differ in measurable ways; framing and design choices matter to outcomes.

What the work does not establish: that any specific app treats any clinical condition; that companion app use causes reduced loneliness in the general population; that companion app use is, on net, beneficial or harmful at the population level; long-term outcomes for sustained use over years; effects in clinical populations, in older adults, or in adolescents.

Like the Stanford-affiliated work, the careful Media Lab framing is “for some users in some situations, real effects in both directions, the longer-run picture is open.”

What this means for users

If you are using an AI companion app and trying to make sense of the research:

The empirical case that companion apps can do real emotional work for some users is reasonably well-established. The empirical case that they do net good across the population, or for any specific person, is not.

If you find your use is increasing your social functioning (you talk to friends more, you feel less stuck), that is consistent with the better readings of the literature. If you find your use is decreasing your social functioning (you talk to friends less, you cancel plans more, you feel more stuck), that is consistent with the harm patterns the same literature documents.

The Media Lab work, in particular, suggests that how you use the app matters more than whether you use it. Treating the app as a tool to think with looks different from treating it as a primary relationship.

We covered the fuller research backdrop in AI Companions and Mental Health and the most-cited specific paper in The Stanford Replika Study.

The broader research context

The Media Lab program sits alongside several other lines of work worth knowing about.

The Stanford-affiliated Replika study (Maples et al.) is the most-cited single paper. Skjuve and colleagues in Norway have done extensive qualitative interview research on Replika user experiences over multiple papers. Pentina, Xie, and others have approached the topic through consumer research and marketing research. De Freitas at Harvard has published industry-skeptical work documenting specific harm patterns. Each of these uses different methods and answers different questions; the Media Lab’s contribution is the closest thing the field has to a programmatic, mechanism-oriented research direction.

For practical guidance derived from this literature, see our AI Companions for Loneliness guide.

Where to read it

The Media Lab publishes openly. The Affective Computing group’s publications are listed at media.mit.edu, and most papers are available as preprints or open-access. The 2025 ChatGPT collaboration is documented in OpenAI’s research posts and in subsequent peer-reviewed write-ups; you can find it through standard search on the lead authors’ names. We strongly recommend reading the original papers, not summaries, when the stakes warrant it.

If a specific claim in this piece does not match what the primary sources actually say, please write us at the contact form. We correct quickly.

FAQ

Is the MIT Media Lab project a single study?

No. It is a research program covering multiple papers and collaborations across several years. Talking about it as one study oversimplifies. The most prominent single output to date was a 2025 collaboration with OpenAI on heavy ChatGPT users and emotional outcomes.

Does this work say AI companions are good for mental health?

No, and the team’s published framing is careful not to. The work documents that some patterns of use correlate with reduced loneliness, that some patterns correlate with increased dependence, and that the design and framing of the AI matter. It does not endorse companion apps as a treatment for any clinical condition.

Has the Media Lab work been peer-reviewed?

Most of the published output has been, yes. Some appears as conference papers, some as journal articles. Preprints are usually available alongside peer-reviewed versions.

What is the longitudinal piece?

The lab has signaled a longer-running research direction following companion chatbot users over time, which would address the limitation that most current research is cross-sectional. We are being deliberately vague about specifics because the full output is still rolling out; we will update this article once the work publishes.

Should I trust this work over the Stanford Replika study?

Both are useful and they answer different questions. The Stanford-affiliated study is a large-population snapshot of Replika users specifically. The Media Lab program is more mechanism-focused, more methodologically varied, and more cautious in its framing. Read both.

AI Companions and Mental Health for the broader research backdrop.

The Stanford Replika Study: What It Actually Found for the most-cited single piece of research in the field.

AI Companions for Loneliness for the practical implications of this body of work.

If you are a researcher in this area and we got something wrong, please write us at the contact form. Corrections are made quickly; reviews are not.