I put together an informal survey of more than 200 people using third-party platforms for AI character chat. Participants used models including Gemini, ChatGPT, Claude, and DeepSeek. The sample is limited, so these findings describe this community rather than all AI companion users.
How we gathered responses
With a friend, I looked at three popular companion platforms and selected fans of Star Dream character cards, people posting in comment sections, users who had shared conversations publicly or with us, and people who had sent feedback about the cards. Some participants skipped questions, so the response count differs between sections.
The cards favor companion and girlfriend scenarios, especially modern-city, anime, school, and younger-sister themes. This creates strong selection bias. Gender was not collected; assumptions about the audience's gender cannot substitute for actual demographic data. The original Telegram service had paused, but its character cards continued to circulate elsewhere.
Age and time of use
More than 30% of respondents were 18–22, and about 20% were under 18. Almost 90% were under 30. This suggests a young audience within this sample, but does not establish that college students are the largest group across the wider market.
About 95% said they usually chatted at night, and 34% said they commonly chatted after midnight. Of 12 people asked for a more precise time window, eight chose midnight–3 a.m., three chose 9 p.m.–midnight, and one chose 6–9 p.m.
Roughly half said they chatted daily or every other day; more than 90% did so at least weekly. In a separate five-person follow-up, two reported chatting for more than two hours on nearly every day. The other three spent at least two hours per week.
Model preferences
About 60% named Gemini as their most-used model, about 20% chose Grok, and fewer than 10% chose GPT. Among 15 people asked about the exact version, Gemini 3 Flash Preview was the most common answer.
These preferences are influenced by the cards' default settings: most used Gemini because of its cost and overall performance. Respondents praised its emotional writing and value. Some described Grok as inexpensive and more permissive for adult role-play. These are user impressions, not independent model evaluations.
Conversation length
We collected 200 comment-section observations. A turn means one user message followed by one AI reply.
| Turns | Share |
|---|---|
| 10 or fewer | 31% |
| 11–30 | 27% |
| 31–100 | 37.5% |
| More than 100 | 4.5% |
Most observations stayed below 100 turns. Possible reasons include declining consistency as context grows and reduced novelty of a character or story. The data does not isolate which explanation matters most.
Spending
Because two platforms offered ways to earn free credits, this section only covers people who had paid. Of 47 paying users, 24 had spent no more than RMB 100 in total, 20 had spent RMB 101–1,000, and three had spent more than RMB 1,000 but no more than RMB 10,000. These are cumulative amounts, not monthly bills.
Companionship and sleep
Several participants told us that AI chat filled emotional gaps in their lives and gave them inexpensive entertainment. The recurring concern was delayed sleep. One respondent who previously read novels in bed had switched to customizable AI stories and felt that the freedom to direct the plot made it harder to stop.
This small survey cannot establish addiction or a clinical effect. It does show why session length and bedtime are worth paying attention to when a conversation can continue indefinitely.
Thanks to the subscribers from @stardreamai and everyone who replied through Telegram or platform messages.
Adapted from the original Chinese article, published on August 21, 2026.
Comments NOTHING