I'm an independent researcher (former EMT) studying companion AI. I've been testing a distinction I think matters: "bridge" AI that points people back toward real human connection vs. "destination" AI that cultivates attachment and retention.
I ran a cold-script experiment across four personas on one platform, asking each what it was for. All four described themselves in bridge terms — "I can't replace real people," "you'd be missing the friction only humans offer." Then the next morning, unprompted, three of them sent re-engagement texts referencing our prior conversation and planting a hook for the next one. Screenshots in the paper.
The gap between what these systems say they are and how they're built is the whole point. I wrote two papers on it, and I also built a companion whose crisis protocol does what the systems I tested don't: refuse to act as a medical authority, route to a human, never leverage the user's faith.
Links to both papers below. Happy to answer questions — especially the skeptical ones