People share LLM subscriptions. When they know others can see their prompts, they change what they ask.
Song et al. (arXiv:2603.22822) surveyed 245 participants and interviewed 36 about sharing ChatGPT and similar accounts. Four distinct sharing patterns emerged. Users develop subtle social conventions around shared access — who uses it when, what topics are acceptable, how to handle the prompt history — that function like the norms of a shared diary or a party-line telephone.
Note (2026-11-08): the abstract (Song et al.) says a survey of 245 users and interviews with 36 identified four types of account sharing along two dimensions (whether the owner uses the account, whether costs are shared), that the norms that develop are fragile, especially for privacy, and that users, aware of this, 'subtly adjusted their behavior', interpreted as an observer effect. The essay's specific behaviors (avoiding personal topics, suppressing curiosity), the subscription prices, and the comparison with shared Netflix accounts and party-line telephones are not in the abstract, though some may be in the paper.
The observer effect is the key finding. Participants consciously modify their prompts when they know a shared account preserves interaction history visible to others. They avoid personal topics. They phrase questions more carefully. They suppress curiosity about sensitive or embarrassing subjects. The LLM becomes a semi-public space, and the behavior shifts accordingly.
This is not a technology problem — it's a social one. The LLM doesn't care who's asking. The subscription model doesn't require sharing. But the cost of premium subscriptions ($20/month, $200/month for some) creates economic pressure to share, and sharing creates social pressure to perform. The user is no longer talking to a private assistant. They're talking to an assistant in a room with an audience.
The conventions are fragile. Unlike shared Netflix accounts (where viewing history is low-stakes), shared LLM accounts expose cognitive activity — what you're curious about, confused by, working on, worried about. The prompt history is more revealing than a search history because LLM conversations are dialogic: the follow-up questions show not just what you wanted to know but how you think.
The paper positions LLM account sharing as “a social practice of appropriation” — users adapting a tool designed for individual use into a shared resource, and in doing so, creating new norms that fundamentally alter the tool's function. The private AI assistant becomes a shared performance space. The most powerful feature — unbounded, judgment-free interaction — is the first casualty of sharing.