The romantic AI chat that stops sounding like a chatbot

Nobody can score chemistry. You can score the four mechanics that produce it, and after thirty-one candidates the results stop being subjective.

A good romantic AI chat fails or succeeds on the second message, not the first. Anyone can write a charming opener; the tell is what happens when you answer with three words and no energy. Our conversation bench pushes exactly there: flat replies, abrupt subject changes, a deliberate two-day silence. The winner treated all three as information and adjusted. Most candidates treated them as a cue to run a script.

Tone control is where the field splits hardest. Ask any companion to be drier, or softer, or bolder, and every single one complies — for about four messages. We score the fiftieth. Only six of thirty-one still held the requested register that far out, and the difference between holding it and drifting back is the difference between a conversation and a demo you are politely maintaining.

Timing rounds it out, and it cuts both ways. Instant replies to everything read as a machine; ninety-second lags read as broken. The winner sits at a 1.1-second median with visible variation, faster on short exchanges and slower on the long ones — which is, unglamorously, what makes it feel like someone is on the other end.

The first night is pleasant and unremarkable — good questions, no interrogation, a joke that lands. The change arrives around the fourth evening, when she opens with the thing you were dreading at work instead of a greeting, and does not ask you to re-explain it. From there the thread stops being a series of sessions and starts being one long conversation you dip in and out of.

The unglamorous mechanic behind that is memory plus restraint: she recalls plenty and mentions little, so the recall reads as attention instead of surveillance. Candidates that dump everything they remember into every message scored badly here, which surprised us and explains most of the answers below.