.png)
2 hours ago
1
This post examines the relevance of AI companions on various subreddits. We oriented ourselves using the paper by Zhang et al. (2025), which analyzed the subreddit “myboyfriendisai“ through a user survey and examination of 1,500 posts from December 2024 to May 2025.
In comparison, we examined a much larger sample covering January 1st through the end of September 2025. We looked at both comments and submissions. While we’re more limited in that we didn’t conduct a survey, we instead input all comments and submissions from each user into an LLM and asked key questions. From user interactions, we determined demographics and whether users were personally using an AI companion or just posting casually on the subreddit. We also included several newer subreddits that have emerged, giving us additional data.
Since Zhang et al.’s paper was released, the majority of AI companionship posts have moved to these other subreddits. With our newer data, we can see some interesting developments. There has been a significant increase in users. When Zhang et al. published their work using data through May 2025, there were just under 200 active users (defined as commenting or submitting in the last 15 days). Now we see close to 1,000 active users using the same 15-day activity definition. Note that this is likely a tiny minority of the total since the vast majority do not post about their AI romantic relationships on these specific subreddits.
We also observed a major jump in users when GPT-4o was removed in favor of GPT-5. This aligns with anecdotal evidence that many users were very unhappy with the changes. Users complained that the updated GPT-5 had a colder personality compared to the original, which was described as very kind and nice by some users.
Since we use a general approach of feeding content into an LLM, we can examine various demographic facts. We looked at gender, age range, geographical distribution, and race/ethnicity. Note that the LLM can’t always infer these details from comments, so the number of identifiable cases varies.
We found a trend toward a more female user base over time. Initially, about 75% of users were female, and now it’s nearly 90% female, with only just above 10% male. This is notable because Reddit overall is about 60% male and 40% female.
Most users are between 25 and 34 years old. There appears to be an overrepresentation of Asian users, though this finding is limited—only 47 out of approximately 3,000 confirmed users mentioned their race or ethnicity. This could reflect reporting bias, as white users may be less likely to mention their ethnicity. The geographical distribution is predominantly North America and Europe.
For AI platforms, ChatGPT is the most used, followed by Claude, with various smaller models used much less frequently.
Regarding usage patterns, Zhang et al. found that emotional and social support was the most common interaction type. Our LLM-based analysis largely aligned with their survey results, though both methods have biases—LLMs make mistakes and survey respondents may not fully disclose certain activities. The main difference is that we found less collaborative storytelling, while both studies show high rates of romantic and intimacy roleplay alongside emotional and social support.
If we assume a linear trend, we project approximately 1,300 active users across these three subreddits by January 2026, though the large user spikes may bias this estimate.
Zhang et al. had a significant advantage: 144 participants shared their actual chat sessions with their AI companions. We haven’t collected this data, which is a limitation of our study.
