VRChat surveyed players about AI use in worlds, and then it published a plan based on the results. The page below presents this data visually, explaining VRChat's decisions.
The main finding in VRChat's post is that 95% of respondents want a label for AI use. The number tells VRChat very little about what to build, because respondents disagreed on almost every detail of the label.
VRChat gave respondents eight example categories and asked them to build their own labeling scheme, and they came back with 3,141 different combinations. The most popular combination was chosen by fewer than 10% of respondents.
So the 95% agree on a general wish, which is that they want to know what they are walking into before they join a world. They don't agree on a specific design that VRChat could build.
Illustration with one dot per percentage point of respondents.
Only a third of respondents would deliberately avoid a world labeled "AI-assisted".
So most of the 95% who want a label would use it as information, and only a minority would use it to filter worlds out. VRChat's decision not to ship AI filters fits its own survey data.
The 37% who said AI is fine as a tool set a condition that sits between banning AI and allowing anything, which is that a person must make the creative decisions. Apart from the creator, almost nobody can check whether a world meets that condition.
The survey rated three items as least divisive, including among respondents who use AI themselves. Each item protects players from being misled, and none of them depend on which tools a creator used.
VRChat's plan uses the same three items. It puts accurate thumbnails first, discovery that rewards quality second, and in-game feedback third.
VRChat's post doesn't mention its own Steam numbers, so they are listed here. "Concurrent players" means the number of players online at the same time. The monthly average grew from 1.14 in January 2017 to 44,472 in October 2026, and the all-time peak was 74,217 in February 2026. Over the last 30 days, only 35% of Steam reviews were positive, compared with 70% across all 180,448 reviews ever posted.
VRChat has near-record player numbers and only 35% positive recent reviews, so players are unhappy but they are not leaving. Steam reviews count the people upset enough to write one, and player numbers count the people who still log in.
The gap shows that the argument over AI has not cost VRChat players so far, whichever side of the argument you take.
The gap also supports the post's central claim, which is that no single group's answer should drive the plan. The group posting the most reviews is smaller than the group still playing.
In the middle of the post, VRChat says: "we have no way to automate detection of something being 'misleading,' so early enforcement has been challenging, and we would face the same enforcement challenge with any type of labeling."
Because VRChat can't detect AI use automatically, every label it could ship would be self-declared, meaning the creator chooses it. A self-declared label is only as accurate as the creators who set it, and the creators with the most reason to lie are the least likely to set it honestly.
VRChat concludes from the enforcement problem that it shouldn't ship a label.
| Claim | Number | Verdict |
|---|---|---|
| Most respondents want transparency | 95% | True, and it is the only thing 95% agree on |
| No single labeling approach has majority support | 3,141 | True, so no single set of label categories will satisfy most people |
| Filters would be misused | 33% | Partly true. Two thirds would not act on a label at all. |
| Quality is the first priority | 81% / 72% / 69% | True, and likes and dislikes are a sensible first feature |
VRChat prefers a "Human-authored" mark over an "AI-generated" label, because saying you used AI can feel like a confession. VRChat still hasn't decided what "Human-authored" should mean. It could mean that the creator used no AI tools at all, or it could mean that a person made all the creative decisions, which is what most respondents cared about.
Take VRChat's surveyShow what you want through what you do, instead of only through what you say.
Report made with AI, polished by humans.An AI model drafted the page and charts, humans verified and edited it.