How Many People Use AI Companions? 2026 Survey Data

What the survey actually measured

The report, The Rise of AI Companions, is a collaboration between Elon University’s Imagining the Digital Future Center and the Elon University Poll, fielded online by YouGov. YouGov interviewed 4,268 US internet-using adults aged 18 and over from 18 to 22 May 2026, then screened for people with emotional or social use of AI, producing a qualifying sample of 1,000 AI companion users. Results were weighted to census benchmarks and 2024 presidential vote, with a reported margin of error of approximately ±3.68 percentage points on the companion-user sample.

Two things about that design are worth pausing on before reading any percentage.

Why the definition matters

First, the study did not ask “do you use a companion app.” It asked about behaviour — emotional and social interaction with an AI — and the systems it names are general assistants: ChatGPT, Gemini, Claude and Copilot. So the 27% figure is not a market share number for dedicated companion products. It is a count of people doing companion-shaped things, wherever they happen to be doing them.

Second, everything after the headline describes the 1,000 companion users, not all Americans. When you read “31% consider their chatbot a friend,” that is 31% of the group that already qualified, which works out to roughly 8% of US internet users overall. Percentages of a screened subgroup get quoted as if they were population figures constantly, and it is the single easiest way to misread this report.

The headline numbers

Lee Rainie, who directs the Imagining the Digital Future Center, framed the release around how quickly a speculative question became a measurable one: “Futurists and cultural commentators have wondered what would happen once artificial systems became connected and companionable with humans.”

Most of this is happening on general assistants

The practical implication for anyone comparing companion apps is that the default entry point is no longer a companion app at all. A large share of the people in this sample built the habit inside a general-purpose assistant they already had open for work or study, and only some of them ever go looking for a purpose-built product.

That reframes the question this site usually gets asked. It is rarely “should I start talking to an AI” — for a quarter of US internet users that already happened. It is “what does a dedicated app add over the assistant I am already using.” The honest answer is a short list: durable persona memory, a character you author rather than prompt from scratch each session, story and roleplay modes, and content rules that permit adult conversation. If none of those matter to you, a general assistant is a reasonable place to stay.

Nearly half rotate two or more bots

The 47% who regularly use two or more different bots or personalities is the finding that best matches how people describe their own habits: one assistant for practical questions, one character for a running story, sometimes a third for a specific mood. It also means single-app loyalty is not the norm, so a comparison built around picking one overall winner misses what most users actually do. Our use-case guide is organised the same way for that reason.

What people bring to the conversation

The topic breakdown is broader than the romance framing that dominates press coverage. Among companion users, 67% sought health, fitness or lifestyle advice, 41% explored financial questions, 39% discussed career or education decisions, 31% raised family issues, 30% asked about legal matters and 26% discussed spiritual, religious or philosophical beliefs. Relationship talk is there — 29% discussed problems in personal relationships and 26% discussed work relationships — but so is the 39% who used AI simply to pass the time when bored.

Sexual chat or roleplay, at 9% of companion users, is a real but minority behaviour. Reporting that treats companion AI as primarily a romantic or sexual product is describing a slice of this, not the whole.

How close the bond gets

The attachment questions are where the report is most uncomfortable. Alongside the 31% who call the bot a friend, another 15% said the relationship is “more complicated” than that. About 39% said AI understands them better than most people do, roughly 36% agreed the AI cares about their wellbeing, and approximately four in ten said they would feel a personal loss if they could no longer use it.

Set against that, 59% still said they prefer conversations with friends and family, and 27% value AI conversation about equally with human conversation. The picture is not mass substitution. It is a minority forming genuine attachment inside a much larger group using AI as a convenient sounding board — which is close to what the longitudinal wellbeing research has been reporting, and consistent with the congressional research overview published in August.

What this changes if you are choosing an app

1. Start from what you already do. If your existing assistant already handles half of the topics above, the gap you are shopping for is memory, character and mode — not raw conversation quality.
2. Expect to run more than one. Nearly half of surveyed users do. Check export and portability before you invest hours in a single character.
3. Test disclosure early. 39% tell AI things they would not tell anyone else. That makes retention and deletion settings a first-week question, not a later one — see our privacy checklist.
4. Notice the agreement reflex. Support is the top-rated benefit at 59%, and agreeable systems are exactly what separate 2026 research has flagged as a trust problem. An app that can disagree with you is worth more than one that cannot.
5. Watch the boredom use case. 39% open AI to pass the time. If that is your pattern, notification design matters more than persona depth.

Four limits worth keeping in mind

• Self-report. Everything here is what people say about their own behaviour and feelings, not observed logs. Attachment questions in particular are sensitive to how they are worded.
• US only. The sample is US internet-using adults. Adoption and stigma differ sharply by country, so these shares should not be read as global.
• May data, September release. Fieldwork ran in May 2026 and the report appeared in September, a long gap in a category that ships monthly.
• Screened subgroup. Nearly every eye-catching percentage is based on 1,000 qualifying users, not on all adults. Multiply by 0.27 before comparing to any population figure.

Frequently asked questions