Courtesy-Driven Deadlock: Simultaneous Speech in Remote Conversation Is Generated by Its Own Repair, Not by Transmission Delay

Two people on a video call start talking at the same time. They both stop, both apologise, and both start again. Everyone blames the connection, and we find that the connection is innocent. Ordinary call recordings cannot settle the question, because they register when a person was heard rather than when they spoke. Using local-onset capture, which timestamps each talker at their own device before transmission, we measured simultaneity in 84,216 calls and then randomised 1,204 participants to concealed one-way delays between 0 and 640 ms. The collision rate did not move: 4.13 per participant-hour, with a latency slope of $0.01$ (95% CI $-0.04$ to $0.06$). Courtesy moved it a great deal. Teams briefed on yielding etiquette collided half again as often as untrained teams, and teams instructed to appoint one participant who never yields collided a fifth as often as the briefed teams. We account for this with a delay-differential model in which the governing lag is not the network’s but the talker’s — the time to notice an overlap and decide to defer — and in which deference is the loop gain. Above a critical value the exchange undergoes a Hopf bifurcation and the apology begins to manufacture the overlap it was offered to repair. The remedy is structural. Every meeting should appoint someone to be rude.