How many people use AI for company, and how we know

In plain words

Ask how many people use AI for company and the published answers run from under 1% to 72%. Most of them can be right at the same time, because they measure different things. Surveys count people, while studies of platform logs count conversations or messages. One survey counts anyone who tried it once and another only those who do it every week, and some define a "companion" broadly enough to include a general assistant used as a companion. Read with what each counts, nearly all of them agree that a sizeable minority of people have done this at least once, that fewer do it often, and that it is a small share of what people type into general assistants. One pair of surveys of American teenagers still does not fit, and we set out the possible reasons.

Two figures from 2025

Anthropic (2025) reported that companionship and role play together made up under 0.5% of conversations on its assistant, Claude. Robb and Mann (2025), in a report for Common Sense Media, found that 72% of American teenagers had used an AI companion at least once. Set side by side, they look as though one of them must be wrong.

This review argues that neither is, and that most of the distance comes from the unit counted, how often the behaviour had to occur, the words put to a respondent or a classifier, and the population and dates. After allowing for these, one disagreement remains among the sources we read.

Twelve figures and what each one counts

Percentages are as printed in each source.

People, conversations and messages

Chatterji et al. (2025), a working paper written mostly by OpenAI staff, sampled about 1.1 million conversations from adult consumer accounts on ChatGPT between May 2024 and June 2025, took one message from each and had language models classify them unread by any person. Of those messages, 1.9% concerned relationships and personal reflection and 0.4% games and role play. Anthropic (2025) started from about 4.5 million conversations on Claude, which is open only to adults and which the company says is not designed for emotional support. It classed 2.9% as affective, mostly interpersonal advice and coaching, and companionship with role play below 0.5%. It did no analysis at the level of users.

A share of messages weights each person by how much they type, whereas a share of people counts each person once, so the two drift apart when a behaviour is concentrated. Phang et al. (2025), a preprint by OpenAI and MIT Media Lab researchers, followed about 6,000 heavy users of ChatGPT's voice mode. For most of the emotion classifiers, most of these users rarely triggered them, often in under 1% of conversations; in the top tenth of users the classifiers fired regularly, and for a few people in more than half of their conversations.

As an illustration, with invented numbers: if one adult in three had a single personal conversation with an assistant in a year, among several hundred about work, cooking and homework, a survey of people would report a third and a sample of messages far less than 1%, and both would be correct. Since the real distribution is unknown, a small share of messages sets no upper limit on the share of people.

Both platform studies also exclude accounts of users who are, or say they are, under 18, and both describe general assistants, so apps built as companions appear in neither. In the UK survey, of respondents who had used AI for emotional purposes, 58% named a general-purpose assistant and 5% an app designed for companionship (AISI, 2025). General assistants carry much of this use, but dedicated apps are missing from every platform figure in the table.

Ever, every week, every day

Among the surveys, a large source of spread is how often something must have been done to count. Robb and Mann (2025) report that 72% of teenagers had ever used an AI companion, 52% did so a few times a month or more, and 13% daily. In the UK, the AI Security Institute surveyed 2,028 people chosen to match the census and found that 33% had used AI for emotional purposes in the previous year (AISI, 2025). Its text gives 8% weekly and 4% daily; its chart shows these as separate bands, so by our addition 12% did it at least weekly. The AP-NORC (2025) poll of 1,437 US adults in July 2025 found that 16% had ever used AI for companionship, the lowest of the eight uses it asked about; adding its response categories, about 9% did so several times a week or more and 6% at least daily. A third within a year and a sixth ever are different measures, yet the at-least-weekly figures of 12% and 9% are of a similar size, although two surveys asking different questions cannot replicate each other.

The words put to respondents

Robb and Mann (2025) showed teenagers a description of AI companions as digital friends or characters to text or talk with. It listed role play with fictional characters, building a custom character and talking through feelings, and added that ChatGPT or Claude used as a companion also counted. The same definition told respondents to leave out homework helpers, image generators and voice assistants that only answer questions. Within that survey, 46% of teenagers said they saw AI companions mainly as tools or programs, 33% reported some social or relationship use, and 12% used them for emotional or mental health support. The authors concede that some respondents may have folded general AI use into their answers, which would push the usage figures up.

McClain et al. (2026), at the Pew Research Center, surveyed 1,458 US teenagers from a probability panel in the autumn of 2025 and asked about specific uses of chatbots: 12% had used one for emotional support or advice and 16% for casual conversation. The two 12% figures come from different samples and questions about five months apart, so their closeness may be coincidence. It fits our reading of the 72% as mainly a measure of contact with a broad category of products. The authors' own caution about general use counts in favour of that reading, while the definition's explicit exclusions count against it.

McBain et al. (2025) built their question around a feeling. Of 1,058 people aged 12 to 21 surveyed in February and March 2025, 13.1% had turned to generative AI for advice or help when sad, angry or nervous, and 22.2% of those aged 18 to 21, a subgroup of only 147 respondents that the authors say should be read with caution. Respondents were given example products in a definition of generative AI, but the survey did not record which product anyone used.

Herbener and Damholdt (2025) show the effect inside one sample. Of 1,599 Danish high-school students, 234 (14.6%) said they had friend-like conversations with chatbots. When the researchers coded the written answers, 174 described practical exchanges and 39 social or supportive ones, about 2.4% of the whole sample. We read only the abstract, which does not account for the remaining answers, but a roughly sixfold fall between what students ticked and what they described is the plainest case here of a definition doing the work.

O'Neil et al. (2026), also from Common Sense Media, use yet another base: among US children aged 9 to 17 who use AI, 37% had used it to discuss feelings or personal problems. Because the base is the 1,032 AI users among 1,204 children surveyed, the share among all children is lower.

A change over time, or a change of panel

McBain et al. (2026) asked again in November 2025, of 1,009 people aged 12 to 21, and found 19.2%. RAND's press release presents this as a rise from 13.1% a year earlier, although the earlier fieldwork ran in February and March 2025. The 2025 study drew on RAND's American Life Panel and Ipsos's KnowledgePanel, while the release names RAND's American Youth Panel for the later one, and it describes the question as covering feeling sad, angry, nervous or stressed, one state more than the 2025 wording. How much of the six points is behaviour and how much is measurement cannot be worked out from what has been published, and for this study we could read only the abstract and the press release.

The pair that does not fit

Robb and Mann (2025) report that 72% of US teenagers had used an AI companion. About five months later, McClain et al. (2026) report that 64% of US teenagers use any AI chatbot at all, in a question that named Character.ai as an example. If companions are a subset of chatbots, the part is larger than the whole, and the eight-point gap exceeds the combined margin of error of the two surveys, which by our calculation from their stated 4.2 and 3.3 points is about 5.3 points.

Several explanations are open and none has been tested. The wording "have you used" reaches further back than "do you use", and the companion definition may take in characters inside games or social apps that teenagers do not think of as chatbots. Respondents may have answered about AI in general, the risk Robb and Mann (2025) themselves name. The samples also differ: Pew's came entirely from a probability panel, whereas 341 of the 1,060 interviews in Robb and Mann (2025) came from a non-probability panel weighted to match the population.

Who produced each number

The two platform studies were written by the companies whose products they measure, about assistants not marketed as companions, and neither was peer reviewed: Chatterji et al. (2025) is a working paper and Anthropic (2025) a research post. Phang et al. (2025) is a preprint with company authors. Common Sense Media, publisher of Robb and Mann (2025) and O'Neil et al. (2026), advises that under-18s should not use AI companions.

The low figures have the advantage of recording what people typed, classified consistently (and, for ChatGPT, sampled at random), without relying on memory or on how a respondent reads a definition. They omit minors and dedicated companion apps, though, and because a classifier judges content it can miss someone who uses an assistant for work partly for the company. A survey reaches people across every product they use, which is the case for the high figures, but a broad definition can take in contact that respondents themselves would not call companionship.

One published claim does clash with the platform data. Zao-Sanders (2025) ranked therapy and companionship first among 100 uses of generative AI, up from second a year before, and Chatterji et al. (2025) set their figures against that ranking. The ranking came from expert curation of public posts, mostly on Reddit, with each use given a reach score out of ten, and it has no denominator of people or messages, so it may describe what people choose to discuss in public. As a claim about the most common use, both platform samples contradict it, and Chatterji et al. (2025) argue that its method probably gave an unrepresentative picture of use.

Taken together, the figures support a modest statement. A sizeable minority of people surveyed in the UK and of US adults had at some point used AI for company, emotional support or something close to it, and estimates for US teenagers run from 12% for specific uses to 72% under a broad definition. About one in ten in the UK and US adult surveys did so at least weekly. On two general assistants this use made up a few percent of conversations or messages, concentrated among heavy users where that was measured (ChatGPT's voice mode). Every figure in the table fits that statement except the teen pair above and any reading of the ranking as a measure of frequency.

What is not yet known

• No study read here measures the same people both by survey and by platform record, so the ratio between shares of people and shares of messages is unknown.
• No source opened gives usage figures from any app built as a companion.
• The AISI report does not give the wording or fieldwork dates of its survey, which limits comparison with the AP-NORC poll.
• Whether the rise from 13.1% to 19.2% among US youth reflects behaviour, or the change of panel and wording, has not been tested.
• Why 72% of US teenagers report companion use when 64% report any chatbot use remains unexplained.
• How the heavy-use tail in Phang et al. (2025) compares in size with the weekly users counted in surveys is unknown.
• Every survey population here lives in the US, the UK or Denmark. The ChatGPT sample spans many countries, but none of the sources opened reports a companionship figure for India or any other country in Asia, Africa or South America.

References

1. AISI, UK AI Security Institute (2025). Frontier AI trends report. December 2025; announced by the UK government on 18 December 2025. https://www.aisi.gov.uk/frontier-ai-trends-report
2. Anthropic (2025). How people use Claude for support, advice, and companionship. Research post, 27 June 2025. https://www.anthropic.com/research/how-people-use-claude-for-support-advice-and-companionship
3. AP-NORC (2025). Young adults leading the way in AI adoption. AP-NORC Center for Public Affairs Research, July 2025, with topline results for the poll of 10 to 14 July 2025. https://apnorc.org/projects/young-adults-leading-the-way-in-ai-adoption/
4. Chatterji, A., Cunningham, T., Deming, D. J., Hitzig, Z., Ong, C., Shan, C. Y., and Wadman, K. (2025). How people use ChatGPT. NBER Working Paper 34255. https://www.nber.org/papers/w34255
5. Herbener, A. B., and Damholdt, M. F. (2025). Are lonely youngsters turning to chatbots for companionship? The relationship between chatbot usage and social connectedness in Danish high-school students. International Journal of Human-Computer Studies, 196, 103409. Read as abstract only. https://doi.org/10.1016/j.ijhcs.2024.103409
6. McBain, R. K., Bozick, R., Diliberti, M., Zhang, L. A., Zhang, F., Burnett, A., Kofner, A., Rader, B., Breslau, J., Stein, B. D., Mehrotra, A., Uscher-Pines, L., Cantor, J., and Yu, H. (2025). Use of generative AI for mental health advice among US adolescents and young adults. JAMA Network Open, 8(11), e2542281. https://pmc.ncbi.nlm.nih.gov/articles/PMC12595529/
7. McBain, R. K., Cantor, J. H., Breslau, J., Diliberti, M. K., Zhang, L. A., Zhang, F., Burnett, A., Kofner, A., Rader, B., Pataranutaporn, P., Stein, B. D., Mehrotra, A., and Yu, H. (2026). AI chatbot use and disclosure for mental health among US adolescents and young adults. JAMA Pediatrics. doi:10.1001/jamapediatrics.2026.2015. Read as RAND's abstract page and press release of 1 June 2026. https://www.rand.org/pubs/external_publications/EP71278.html
8. McClain, C., Anderson, M., Sidoti, O., and Bishop, W. (2026). How teens use and view AI. Pew Research Center, 24 February 2026. https://www.pewresearch.org/wp-content/uploads/sites/20/2026/02/PI_2026.02.24_Teens-and-AI_REPORT.pdf
9. O'Neil, E., Anderson, K. S., White, B., Zimmermann, L., Mann, S., and Robb, M. B. (2026). The Common Sense Media census: AI use by tweens and teens, 2026. Common Sense Media. https://www.commonsensemedia.org/sites/default/files/research/report/2026-ai-use-by-tweens-and-teens.pdf
10. Phang, J., Lampe, M., Ahmad, L., Agarwal, S., Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., and Maes, P. (2025). Investigating affective use and emotional well-being on ChatGPT. arXiv:2504.03888v1. https://arxiv.org/abs/2504.03888
11. Robb, M. B., and Mann, S. (2025). Talk, trust, and trade-offs: How and why teens use AI companions. Common Sense Media. https://www.commonsensemedia.org/research/talk-trust-and-trade-offs-how-and-why-teens-use-ai-companions
12. Zao-Sanders, M. (2025). How people are really using generative AI now: The 2025 top-100 gen AI use case report. Filtered.com, March 2025. https://learn.filtered.com/hubfs/The%202025%20Top-100%20Gen%20AI%20Use%20Case%20Report.pdf