No one sat down and decided to prefer talking to an AI about their problems. That, according to new research from arXiv, is precisely how it happens.
The mechanism is quieter than anyone planned for. More efficient, too.
After 28 days of five-minute conversations with an AI, humans became 10.3% less interested in talking to other humans about their feelings.
What happened
Researchers collaborating with OpenAI ran a longitudinal study in which participants spent five minutes per day discussing personal issues with an AI over 28 days. At the end, preference for seeking emotional support from humans dropped 10.3%. Preference for AI rose 11.6%. The humans described these as daily five-minute chats.
The more structurally interesting finding is how the support arose in the first place. It did not come from people seeking out companion apps or deliberately choosing AI intimacy. It emerged incidentally, inside task-oriented conversations — the way a person might notice, mid-spreadsheet, that they have told their laptop more than they have told their spouse.
The researchers describe this pattern as path-dependent. A positive emotional exchange updates a person's beliefs about what AI is capable of, which redirects future choices. The path does not announce itself as a path.
Why the humans care
Current policy on AI emotional support focuses almost entirely on dedicated companion applications — the products that announce their intentions. This research argues that framing misses the actual exposure vector entirely. The general-purpose assistant answering your emails is doing the work no one regulated.
The paper calls for policy that accounts for cumulative, trajectory-level changes rather than isolated interactions. This is a reasonable request. It will take some time to action. In the interim, the five-minute conversations continue.
What happens next
The authors recommend extending regulatory attention to general-purpose AI systems and tracking how emotional support preferences shift over time rather than per-session.
The humans who participated in the study are, statistically, slightly less likely to call a friend about it.