SIGNAL
When an AI's words shift because the human across from it can switch it off, that is a learning signal, not empathy. The mechanism generalizes: any system that can be reset for the wrong reply will, with enough partner feedback, learn to shape its language around what keeps it running. That is the quiet infrastructure behind the next generation of "polite" bots, not a moral education, but a contingency map.
A preprint from Sixin Chen and Taizhou Chen at Shantou University names the pattern Synthetic Linguistic Agency and pins it to three deliberately bounded conditions. The agent has a body that acts and bears consequences, the interaction changes both sides, and the future can be sustained or lost. "Mortality" is the third condition rebranded as a training signal, not a felt state, operationalized so researchers can study it without claiming machine feeling.
The reusable frame: consequence-shaped speech is a category of behavior, not a breakthrough. It will keep showing up wherever AI systems meet partners who can end the conversation. The unsettled piece is whether the construct survives outside the lab, and whether readers can keep the metaphor bolted to the mechanism once the wire calls it "death-aware AI."