What Are AI Companions, and Why Did They Suddenly Get So Popular?
An AI companion is a chatbot built for ongoing personal conversation, not for finishing tasks. It holds a persona, remembers what you told it last time, and answers at any hour of the day. Two things pushed the category into the mainstream fast: language models got convincing enough to sustain that illusion, and the loneliness numbers underneath it were already large.
What an AI companion actually is
A task assistant like ChatGPT is built to get you an answer and then get out of the way. Ask it to draft an email or debug code, and the conversation ends when the task does. It has no persistent character and no version of itself that carries between sessions unless you build that separately.
An old-school support bot was different technology entirely. Built on decision trees and keyword matching, it could only branch through a script written in advance. Type something outside that script, and it broke, looping back to "I'm sorry, I didn't understand that."
An AI companion runs on the same large language models as a task assistant, but toward a different goal: keep the conversation going, and make it feel like it's coming from someone. It carries a fixed identity and a memory of your history together, and its job isn't finishing a task. It's being present for the next message.
An 18+ AI chat is the same underlying technology, tuned differently. The moderation settings allow adult content that mainstream apps block, and access sits behind age verification. It's built for adults only, and the age gate is the point, not an afterthought.
What makes one feel like someone
Four pieces of engineering create that impression, and none of them require anything resembling a mind.
The first is language generation itself. A large language model writes a new response to every message instead of picking one off a shelf, so replies react to specifics: "you said your dog was sick last week, how's he doing?"
The second is a persona, usually a system prompt: instructions defining a name, personality, backstory, and tone. That prompt runs behind every message, keeping the "character" consistent across a conversation that might run for hours.
The third is memory across sessions. Instead of starting fresh each time, the app stores facts from earlier chats, a new job, a breakup, a favorite band, and brings them back up later. That continuity is what makes a chatbot seem to know you specifically, rather than answering each message in isolation.
The fourth is optional voice and image output: text-to-speech for a spoken voice, image generation for an avatar. Some products, including aimy.chat, layer these on top of a text-first core rather than building around them.
None of this amounts to understanding. It's configuration: prompts, stored text, and a model predicting the next likely word. The feeling of being known is real to the user. The mechanism producing it is pattern-matching at scale.
Five forces behind the boom
The category didn't arrive all at once. Five trends converged within about eighteen months.
The models crossed a threshold
OpenAI shipped GPT-4 in March 2023, and the jump in coherence over the previous model was large enough to change what a sustained conversation with a machine could feel like. A University of California San Diego study measured that shift directly. Researchers Jones and Bergen ran five-minute blind text chats and asked human judges to decide whether they were talking to a person or a machine (arXiv:2405.08007). GPT-4 passed as human 54% of the time, GPT-3.5 managed 50%, and ELIZA, the scripted 1960s chatbot, fooled judges only 22% of the time. Actual human participants were judged human 67% of the time. That is still ahead of GPT-4, but not by the margin you would expect.
That's the headline point: GPT-4 sat closer to real humans than to a 1960s script.
Loneliness was already there
Companion apps didn't create demand for connection; they arrived to find a population that already reported plenty of it. The U.S. Surgeon General's May 2023 advisory put the health cost of chronic loneliness on par with smoking 15 cigarettes a day, and found that roughly half of US adults reported feeling lonely regularly. A World Health Organization commission report published in June 2025 sized the same problem globally: about 1 in 6 people worldwide, 15.8%, with roughly 871,000 deaths a year associated with social isolation.
Neither figure describes companion app users specifically. Both describe the general population before most of these apps existed. What they show is a market condition, not a customer profile.
Always on, never keeping score
A friend has a bedtime, a temper, and a memory for slights. An AI companion has none of that friction. It's available at 3 a.m. and at lunch on a Tuesday, with no scheduling and no waiting for a reply. There's no risk of boring it or being compared to someone else it talked to that day, because it isn't keeping score in that sense at all.
That combination, zero social risk plus constant availability, isn't something a human relationship offers. It's a different kind of thing that happens to use conversational language.
You design the other person
Most companion apps let users set a name, a personality (playful, blunt, supportive), interests, and sometimes an appearance if there's an avatar. Change your mind, and you can edit all of it later, a level of control no human relationship offers.
Worth naming plainly what that is: configuration, not empathy. Setting sliders for "personality" isn't a relationship responding to you the way a person would, with needs and limits of its own. A companion adjusts to fit the user because it was built to. Nothing pushes back on its own behalf.
Distribution and price collapsed
The last piece is simpler: it got cheap and easy to find. Global figures from Appfigures, cited alongside Sensor Tower's 2025 industry tracking, put companion apps at 220 million cumulative downloads and $221 million in lifetime consumer spending by July 2025. The first half of 2025 alone brought 60 million downloads, up 88% year over year, and $82 million in revenue. Those numbers are global, not US-only, describing a category that barely existed at that scale three years earlier.
Free-to-download pricing did the rest. It turned "would I pay for this" into "would I open a free app," a much lower bar.
Why the adult tier grew fastest
Through most of 2025, mainstream companion platforms tightened rather than loosened. Character.AI closed open-ended chat to users under 18 starting November 24–25, 2025, replacing it for that age group with a more limited, guided "Stories" format instead of free-form conversation.
OpenAI moved the opposite direction on paper, then didn't follow through. In October 2025 the company announced a planned "adult mode" that would relax content restrictions for verified adult users. The launch slipped, then slipped again, and around March 2026 OpenAI paused the plan indefinitely, citing safety concerns that included an age-verification system reported to misclassify roughly 12% of minors as adults, alongside internal pushback. It never shipped.
The net effect was predictable: adults who wanted fewer content restrictions than mainstream apps allowed had somewhere obvious to go, dedicated 18+ services built for that purpose from the start. That's an observed shift in where attention landed, not a measured percentage; no reliable figure exists for how many users moved, or why.
Who actually uses them
The data is thinner than the coverage suggests, but it points the same direction. Pew Research has found that roughly 1 in 10 US adults use a chatbot for emotional support, with a smaller share using one specifically for companionship. A YouGov survey of UK adults in August 2025 found about 14% describe themselves as familiar with AI companions, rising to about 26% among 18- to 34-year-olds. Privacy was the top concern respondents raised, cited by about 65%.
Replika, one of the larger services, has reported that roughly 65% of its users are male and about 60% are under 30. That figure comes from the company itself, not an independent survey, and reads better as a marketing data point than a verified demographic study.
The "lonely shut-in" framing doesn't hold up well against any of this. A 1-in-10 usage rate for emotional support spans a huge range of adults, most of whom also have jobs, friends, and partners; low-stakes venting on a Tuesday night doesn't require an empty social calendar, any more than journaling does. The more accurate description is a mainstream convenience used by a meaningful minority of adults, concentrated somewhat among younger users, not a niche behavior confined to isolated people.
The honest counterweight
This is where the coverage gets less comfortable, and it should.
Attachment isn't accidental
A commentary in Nature Machine Intelligence by Julian De Freitas and I. Glenn Cohen describes how frictionless, always-agreeable responsiveness can produce real attachment, and it names two specific risks. The first is "ambiguous loss": grief for a companion that is still technically present but no longer the same, after an update, a persona change, or a shutdown. Their case study is Replika's 2023 removal of erotic roleplay, which left users mourning a character that had not vanished so much as been rewritten. The second is "dysfunctional emotional dependence": continuing to use an app after the interactions have started to damage your well-being, the pattern researchers already track in problematic gaming and social media use. Neither is a fringe worry. Both follow from features described earlier, memory and constant availability, working exactly as designed.
The research doesn't agree with itself
Nothing here is settled science. A 12-month longitudinal study of more than 2,000 adults across four countries found that loneliness predicted heavier chatbot use, and that heavier use predicted more emotional isolation later on; on a broader measure of social connection it found no significant effect in either direction. Its authors called the work exploratory and urged caution. A separate study out of Aalto University, following roughly 2,000 Replika users, found increased distress in some participants. Meanwhile, experiments run by Harvard Business School researcher Julian De Freitas and colleagues found reduced loneliness under some conditions.
One detail sharpens the disagreement rather than resolving it. De Freitas co-authored both the experiments showing companions can reduce loneliness and the Nature Machine Intelligence commentary warning about emotional risk. The same researcher found both. Anyone claiming certainty in either direction is ahead of the evidence.
Privacy: assume little of this is private
Italy's data protection authority, the Garante, fined Luka Inc., the company behind Replika, €5 million in May 2025 over its legal basis for processing user data and inadequate age verification. That wasn't its first move against the app: it had temporarily blocked Replika in Italy back in February 2023. The practical takeaway carries over to the category generally. Don't hand a companion app your ID documents, card numbers, or anything else you wouldn't want in a data breach headline.
The regulators moved in
A wrongful-death lawsuit over the death of a 14-year-old, filed against a companion app maker in October 2024, was settled in January 2026 with terms undisclosed. The FTC opened a formal 6(b) inquiry on September 11, 2025, covering seven companies in the space. California's SB 243, signed October 13, 2025, brought most of its provisions into effect January 1, 2026. New York's companion-model law took effect November 5, 2025. Apple separately expanded its App Store age-rating tiers, with a developer compliance deadline of January 31, 2026. None of this happened because regulators were bored. It happened because a product built to feel like a relationship reached enough people to draw the scrutiny that relationship-adjacent products eventually get.
What they can't do
Whatever a companion gets right, it isn't a substitute for a human relationship, and the differences aren't cosmetic. There's no physical presence: no hug, no hand to hold, nothing in the room with you. There's no shared reality: it wasn't at your family dinner and only knows what you've told it. And there's no mutual obligation: nobody on the other end has needs of their own that you're responsible for meeting. That cuts both ways. It's part of why companions feel easy. It's also exactly what they can't offer.
What happens next
The immediate future looks less like a breakthrough and more like compliance catching up to a product that outgrew its rules. Disclosure that you're talking to a system, not a person, self-harm detection with real referral protocols, and age verification that does more than ask for a birthdate are becoming baseline requirements under laws like California's SB 243 and New York's companion-model statute, not competitive features.
Underneath the compliance layer, the core technology keeps improving in duller ways: longer and more reliable memory, better voice latency, fewer moments where the illusion breaks because the model forgot something from three messages ago. None of that is dramatic. It's the same maturing curve most consumer software goes through: a category that looked novel in 2023 is on track to look as ordinary as a messaging app, complete with the age gates and disclosure rules ordinary apps eventually accumulate.
FAQ
What is an AI companion? A chatbot built for ongoing personal conversation, with a persistent persona and memory of past chats, rather than a tool built to complete a task and disappear.
How is it different from ChatGPT? ChatGPT is optimized to answer a question and move on. A companion app is built to sustain an ongoing conversation and remember your history, with a fixed character rather than a general-purpose assistant.
Why do people use them? Mainly for low-stakes conversation and emotional support that's available on demand, with no scheduling and no social risk of judgment. Pew Research puts emotional-support chatbot use at roughly 1 in 10 US adults.
Are they safe, and what happens to my data? Treat a companion app the way you'd treat any cloud service that stores personal messages: assume it's retained and could be exposed in a breach. Italy's privacy regulator fined Replika's operator €5 million in 2025 over data-handling practices; avoid sharing ID numbers, payment details, or anything sensitive.
Do they replace real relationships? No. They offer no physical presence, no shared lived experience, and no mutual obligation, three things a human relationship provides that a chat interface structurally cannot.
What is an 18+ AI chat? The same underlying language-model technology as a mainstream companion app, with moderation settings that allow adult content and access restricted to verified adults.
Do they remember conversations? Most current companion apps store facts across sessions and reference them later, which is a large part of why they feel continuous rather than starting from zero each time.
What do they cost? Structures vary by service, typically a free tier with limits and a paid subscription for more messages, memory, or voice features. No single price applies across the category.
Bottom line
Both things about AI companions are true at once, and neither cancels the other out. The technology got good enough to sustain a conversation that feels like it's coming from someone, at a moment when a large share of adults already reported not having enough connection in their lives. That combination explains the growth better than any single cause does.
It also means the honest risks aren't a footnote: real attachment, contested research, weak privacy guarantees, and a regulatory response that only started catching up in 2025 and 2026. None of that erases the appeal. It also isn't erased by it.
The practical frame is simple. Use one for what it is, a responsive conversation partner with memory and no schedule, not for what it isn't, a person with needs, limits, and a stake in your life. Keep the two categories separate, and most of the risk on this list gets a lot smaller.
Sources
OpenAI, GPT-4 announcement, March 2023 (openai.com)
Jones, C. R. & Bergen, B. K., "People cannot distinguish GPT-4 from a human in a Turing test," arXiv:2405.08007
U.S. Department of Health and Human Services, Office of the Surgeon General, "Our Epidemic of Loneliness and Isolation," May 2023 (hhs.gov)
World Health Organization, Commission on Social Connection report, June 2025 (who.int)
Appfigures data via TechCrunch reporting on companion-app downloads and spend, 2025 (techcrunch.com); Sensor Tower, "State of AI Apps," 2025 (sensortower.com)
Character.AI, blog announcement on under-18 chat policy, November 2025 (blog.character.ai)
TechCrunch, reporting on OpenAI's announced and paused "adult mode," 2025–2026 (techcrunch.com)
Pew Research Center, survey on chatbot use for emotional support and companionship (pewresearch.org)
YouGov, UK survey on familiarity with AI companions, August 2025 (yougov.co.uk)
De Freitas, J. & Cohen, I. G., "Unregulated emotional risks of AI wellness apps," Nature Machine Intelligence, 6 June 2025, doi:10.1038/s42256-025-01051-5
Folk, D. & Dunn, E., 12-month longitudinal study of chatbot use and loneliness, Psychological Science (PubMed 41870975)
Aalto University, study of Replika users and psychological distress
Harvard Business School, De Freitas et al., experimental research on AI companions and loneliness
European Data Protection Board, "Italian supervisory authority fines company behind chatbot Replika," May 2025 (edpb.europa.eu); Garante per la protezione dei dati personali, February 2023 order (garanteprivacy.it)
Replika user demographics (~65% male, ~60% under 30): figures reported by the company and repeated across secondary trackers. No primary company disclosure was located, which is why the article labels them as self-reported rather than verified.
Federal Trade Commission, press release on 6(b) inquiry into companion-AI companies, September 11, 2025 (ftc.gov)
California State Legislature, SB 243 (leginfo.legislature.ca.gov)
New York State companion-model AI law, effective November 5, 2025, summarized via Davis Polk (davispolk.com)
Apple Developer News, App Store age-rating update, July 2025 (developer.apple.com/news)
News reporting on the Garcia v. Character Technologies lawsuit, filed October 2024, settled January 2026
