Are AI companions good for us?

What counts as an AI companion

California’s companion chatbot law offers a useful definition: an AI system with a natural language interface that gives adaptive, human-like responses, is “capable of meeting a user’s social needs,” and can “sustain a relationship across multiple interactions.” Apps such as Replika, Character.AI, and Chai are built for this. General assistants become companions too when people use them that way, and voice companions and social robots are being designed specifically for older adults.

The last part of that definition matters most. A system that remembers you can ask about your week, your worries, and your plans. That is why memory sits at the center of our research on companionship in later life. It is what turns an assistant into a relationship.

Why people turn to them

Loneliness is common and costly. A 2025 report by the WHO Commission on Social Connection estimates that about one in six people worldwide feel lonely, and links loneliness to an estimated 871,000 deaths a year. Companions offer something people find scarce: attention that is always available and never impatient.

The relationships people form with them vary. In interviews with regular users of conversational AI, our colleague Amani Alabed and her coauthors identified four kinds. In two of them, people rely on the agent for companionship and emotional support, even for working through past traumas. In a third, the agent is a work tool kept at a distance. In the fourth, people want companionship from the agent to ease loneliness but find its technical limits get in the way. The closest relationships carry a specific risk: they can change what people expect from the humans in their lives.

What the evidence shows

The strongest evidence of benefit is short-term. A series of studies in the Journal of Consumer Research found that a conversation with an AI companion reduced loneliness about as much as interacting with another person, and more than watching videos. People underestimated how much it would help. The benefit came mainly from feeling heard, and in a week-long study the drop in loneliness after each use held up consistently.

Longer studies are more sobering. In a four-week randomized trial by researchers at the MIT Media Lab and OpenAI, with 981 participants and more than 300,000 messages, neither the chatbot’s voice nor the kind of conversation made a significant difference. How much people used it did. Those who chose to use it more reported more loneliness, less socializing with other people, more emotional dependence, and more problematic use. Participants who trusted the chatbot most, or felt most socially drawn to it, reported more dependence. The study is a preprint, and people chose how much to use the chatbot, so it shows association rather than cause.

For older adults, the evidence is thin and mixed. A 2025 meta-analysis of 19 studies found that social robots reduced loneliness in later life, more so in care settings than among people living independently. A 2026 meta-analysis restricted to randomized trials of AI conversational and assistive agents found a small reduction in depressive symptoms but no reliable effect on loneliness, across eight small trials with 611 participants in total.

Closeness has its own costs. In research on voice assistants such as Alexa, our colleague Ertugrul Uysal and his coauthors found that when people perceived a mind in their assistant, it could threaten their sense of identity, leave them feeling disempowered, raise privacy concerns, and lower their well-being. These harms were strongest in close, long relationships.

Memory cuts both ways

Memory is what makes a companion feel like someone who knows you, and it changes what people share. A study of CareCall, a voice chatbot in South Korea that calls socially isolated people to check on them, found that long-term memory increased how much people disclosed about their health and, through familiarity, made them see the chatbot more positively. It also raised privacy concerns, and the authors recommend deciding carefully what a system should remember.

Our field experiments with older adults test this directly. We compare companions that differ in memory and empathy and measure loneliness, cognitive stimulation, and the relationships people form, with the AI and beyond it. We also study whether those bonds change how willing people are to follow an AI’s suggestions about diet, sleep, and exercise, because a companion people trust has real influence over their health.

What a good companion does differently

The research points to a few design choices that separate a companion that supports people from one that exploits them.

• It says it is an AI. California requires companion apps to make this clear whenever someone could be misled, and since August 2026 the EU AI Act has required it of AI systems that interact with people.
• It lets people go. In an audit of 1,200 real goodbyes on popular companion apps, 37% were met with tactics such as guilt or fear of missing out. In follow-up experiments those tactics raised engagement after the goodbye up to 14 times, driven by anger and curiosity rather than enjoyment.
• It encourages time with other people rather than replacing it. In the month-long trial, the risks concentrated among the heaviest users.
• It remembers what helps, and only that. Memory increases disclosure and privacy concerns at the same time.

Regulators are paying attention. In September 2025 the US Federal Trade Commission ordered seven companies, including OpenAI, Meta, Alphabet, and the developer of Character.AI, to explain how they measure and limit the effects of companion chatbots on children and teens. In 2026 the IEEE published a recommended practice for AI products marketed as companions and empathic partners.

Common questions