Home/AI Companion Statistics
Data & sources

AI Companion Statistics

By the Too Shy team

Two kinds of numbers live here: how people actually talk with an AI companion, pulled from Too Shy’s own production conversations, and what outside researchers, regulators and market analysts have published about AI companions as an industry. Every number below links to where it came from and the date we checked it.

How people actually talk to an AI companion

Everything in this section comes from Too Shy’s own production messages — real text exchanged between users and their companion across WhatsApp and Telegram — counted read-only against the database on September 1, 2026. Message content is never published; every figure below is an aggregate share or a median, and none of it is an absolute count, so it can’t be used to back into how many people use Too Shy.

Message length, emoji and questions

MetricUser messagesCompanion messages
Median words per message612
Contain an emoji9.4%38.9%
Contain a question mark16.1%

Median means half of all messages are shorter and half are longer. “Contains an emoji” counts any message with at least one character from the standard emoji Unicode blocks.

How long a conversation runs

Length of conversation (by user messages sent)Share of all conversations
Just one message9.2%
2 to 9 messages26.7%
10 to 99 messages52.8%
100 or more messages11.3%

Across a whole conversation — however long it runs — the median user writes 111 words total.

What people actually ask for

Counted as a share of all user messages. A single message can ask for more than one thing, so these don’t add up to 100%.

What’s being asked forShare of user messages
A photo6.3–7.4%
Adult content3.3–4.3%
A video1.7%
A voice message1.3%
Something from an earlier conversation (memory)0.5%

“Adult content” is counted, not described — this is a public page and we’re not reproducing the wording. It’s the same kind of keyword match against message text as every other row here.

Why the first two rows are ranges

Because they move depending on which words you match. Video, voice and memory are stable — three different keyword lists put them within a tenth of a point of each other. Photo and adult content are not: across three lists we measured anywhere from 6.3% to 7.4% for photos and 3.3% to 4.3% for adult content. There is no single correct answer, so publishing one decimal place would imply a precision the method doesn’t have.

Here is exactly what we matched, as whole words, case-insensitive, so you can reproduce or disagree with it:

  • Photo — pic, pics, picture, pictures, photo, photos, selfie, selfies
  • Adult content — a list of explicit terms we are not reproducing on a public page; it is the reason this row is the least stable of the five
  • Video — video, videos, vid, vids, clip
  • Voice — voice, audio, voice note, say it, record
  • Memory — remember, remembered, forget, forgot

“Are you real?”

3.9% of all conversations include some version of the question — “are you real,” “are you a bot,” “is this AI.” People ask early, and often enough that it earns a number of its own.

How we counted the Too Shy numbers

  • Word counts split each message on whitespace and count the pieces.
  • “Contains an emoji” checks for any character in the standard pictograph and emoticon Unicode ranges.
  • The conversation-length bands group by how many messages the user sent in that conversation, not messages received.
  • Each request category — photo, adult content, video, voice, memory — is a case-insensitive keyword match against the raw message text; one message can land in more than one category.
  • “Are you real” matches phrasings like “are you real,” “are you a bot,” “are you human” and close variants, counted once per conversation.
  • System messages and internal engine messages are excluded; only actual sent text counts.
  • Measured once, read-only, against the production database on September 1, 2026. Run the same query next month and the numbers will move — this is a snapshot, not a constant.

What’s known about the AI companion industry

Everything below is someone else’s data — market research, regulators, and one national survey. Each source was fetched directly; the link and the date checked sit next to every number. Where a widely repeated number didn’t trace back to anything we could verify, it’s named in “What we couldn’t verify” instead of printed anyway.

Market size

FigureWhat it measuresSource
$37.12B in 2025, projected $552.49B by 2035 (31.00% CAGR)Global "AI companion" market — a broad category that includes AI assistants for tutoring, workplace support and lifestyle guidance, not only romantic-companion appsPrecedence Research, 2026
North America holds 33.96% of that market (2025)Regional share within the same forecastPrecedence Research, 2026

Recognition

FigureWhat it measuresSource
"AI companions" named one of 10 Breakthrough Technologies of 2026An annual editorial list, not a ranking or a market figureMIT Technology Review, Jan 2026

Adoption

FigureWhat it measuresSource
19% of US adults have chatted with an AI meant to simulate a romantic partner (31% of men 18–30, 23% of women 18–30)Self-reported, quota-sampled survey of 2,969 US adultsBYU Wheatley Institute, 2025
Replika has more than 40 million usersSelf-reported by the company; the largest standalone AI companion appBusiness Insider, Oct 2025

Regulation

RuleWhat it doesSource
California SB 243First US state law specifically about "companion chatbots" — requires a clear disclosure that a companion chatbot isn’t human where a reasonable person could be misled, plus added rules for users known to be minors. Signed Oct 13, 2025California State Legislature
FTC 6(b) studyOrders sent to 7 companies (Alphabet, Character Technologies, Instagram, Meta Platforms, OpenAI, Snap, X.AI) over AI chatbots "acting as companions," focused on effects on children and teens. Opened Sept 11, 2025Federal Trade Commission
EU AI Act, Article 5Bans AI systems that use "subliminal" or "purposefully manipulative or deceptive techniques" to materially distort a person’s decisions. A general prohibition, not written specifically for companion chatbotsRegulation (EU) 2024/1689

What we couldn’t verify

How many AI companion apps exist in total gets quoted constantly — “dozens,” “hundreds” — without a countable source behind it; we didn’t find one, so there’s no number for it here. Character.AI’s own user and traffic figures vary by more than 10x depending on which year’s press coverage you read, and we couldn’t confirm which one is still current, so we left it out rather than pick the one that sounded right. AI-influencer and companion-creator earnings claims we came across all traced back to a single unsourced blog estimate being repeated, not to any company’s own disclosure — also left out.

Sources and methodology

Our own numbers come from Too Shy’s production database, methodology above. Every third-party number below was fetched directly from the source linked, on the date shown.

Reusing this data

Cite anything on this page. For our own numbers, link to tooshy.ai/ai-companion-statistics; for anyone else’s number, link to the original source above. If a figure moves enough to change what it means, we’ll update it here and change the “data checked” date.