Happy Young Women, Grumpy Old Men? Emotion-Driven Demographic Biases in Synthetic Face Generation

Researchers audited eight text-to-image models and found that emotionally conditioned prompts systematically amplify demographic biases, with negatively valenced emotions consistently shifting outputs toward White, middle-aged, male-coded faces while underrepresenting younger women and Black individuals. The study reveals that intersectional demographic combinations face near-erasure in synthetic face generation, highlighting critical gaps in current bias evaluation practices.

This research exposes a critical vulnerability in generative AI systems that extends beyond simple demographic representation into the intersection of emotion, identity, and algorithmic output. The finding that negative emotions correlate with shifts toward White, older, male-coded faces suggests that emotional conditioning inadvertently reinforces culturally embedded stereotypes—a phenomenon absent from traditional single-attribute audits. This matters because synthetic faces increasingly populate digital media, training datasets, and AI systems, potentially encoding these biases at scale.

The cross-ecosystem comparison between Western and Chinese model families reveals whether bias patterns reflect fundamental architectural flaws or culturally specific training data choices. The compound underrepresentation of intersectional groups, particularly young Black women, demonstrates that demographic bias cannot be assessed through isolated demographic categories. An audit showing balanced gender representation might simultaneously mask the near-total absence of specific gender-race-age combinations.

For AI developers and deployers, these findings underscore that pre-deployment evaluation must move beyond aggregate demographic metrics. Emotion-conditioned prompts represent a deployment reality—users naturally express emotional intent when generating content. The valence-driven bias mapping suggests that systems trained or fine-tuned on emotionally diverse data may perpetuate harmful associations between negative emotions and specific demographic profiles.

Standardizing intersectional, emotion-conditioned audits as baseline practice would require significant operational changes in model evaluation. This research establishes that current deployment standards are insufficient, creating liability exposure for organizations using these models in sensitive applications like hiring, dating, or media generation where demographic representation directly impacts real-world outcomes.

• →All audited T2I models significantly overrepresent young faces, with compound underrepresentation of intersectional demographic groups including young Black women.
• →Negative emotion prompts consistently shift generated faces toward White, middle-aged, male-coded individuals, creating a valence-driven demographic bias pattern.
• →Single-attribute demographic audits fail to capture intersectional erasure patterns, requiring more granular evaluation methodologies.
• →Cross-ecosystem comparison shows bias patterns exist in both Western and Chinese model families, though specific patterns may diverge.
• →Current pre-deployment evaluation practices are insufficient and should incorporate emotion-conditioned and intersectional audits as standard requirements.