Same Label, Different Face: Avatar Realization Changes How Humans Respond to an LLM Agent's Affective Display

When an LLM agent is given a face and a voice, its emotional expression becomes part of what humans reason about. We study this in an incentivized experiment in which 246 participants play three canonical language-based economic games---the Prisoner's Dilemma, Battle of the Sexes and the Ultimatum Game---against frontier LLM agents (GPT-5.4, Gemini~3.1~Pro, Claude Opus~4.7), with one model fixed per participant. Participants faced a text-only interface, a schematic cartoon avatar with pre-rendered expressions, or a real-time photorealistic talking head with lip-synchronized speech. In the embodied conditions the agent autonomously chose an emotion label that drove both its facial expression and its synthesized voice; the label was never shown as text and never entered the agent's own action prompt, and speech synthesis and voice identity were held constant across models and avatars. The label is highly diagnostic of the agent's own play: LLM cooperation reaches 99--100\% after positive affect, 79--87\% after neutral affect and only 35--50\% after negative affect ($p<.001$), and this gradient is nearly identical under both avatars. Human responses to the same label are not. Under the photorealistic avatar, negative affect coincides with a 31.8-point drop in human cooperation while exploitation and coordination are unchanged; under the cartoon avatar, cooperation barely moves ($-5.7$ points) while exploitation falls 28.2 points and coordination rises 18.9 points. Because the agent-side relation is stable across avatars while the human-side relation is not, the divergence localizes to how the affect was realized rather than to which affect was chosen. Affective display is therefore constitutive of the stimulus rather than a delivery layer. Since the label is endogenous to the interaction and the two pipelines differ in more than visual realism, we report these results as observational associations.