Unmasking Bias: How Language Models Profile Authors From Lyrics

The age of artificial intelligence is unveiling an inconvenient truth about the way large language models (LLMs) perceive culture. As these models become more integrated into applications affecting society, their biases are coming under the microscope. A recent examination highlights how LLMs interpret authorship by analyzing song lyrics, revealing a tendency to default to North American ethnicity.

Probing Cultural Bias

The study evaluated various open-source LLMs on over 10,000 song lyrics, assessing their ability to conduct author profiling. The models predicted singers' gender and ethnicity without any task-specific fine-tuning, an approach known as a zero-shot setting. While the models exhibited a non-trivial level of accuracy in profiling, they also displayed distinct cultural alignments.

Most models leaned towards assuming a North American ethnic background. However, DeepSeek-1.5B stood out by aligning more closely with Asian ethnicity. This divergence wasn't just evident in the prediction outcomes, but also in the rationales generated by the models. What does this tell us about AI's understanding of cultural nuances?

Measuring Bias: New Metrics

To quantify these biases, the researchers introduced two fairness metrics: Modality Accuracy Divergence (MAD) and Recall Divergence (RD). These tools revealed that Ministral-8B had the highest level of ethnicity bias, while Gemma-12B emerged as the most balanced model. It's a clear indicator that not all models are equal handling cultural diversity.

The implications of these findings are significant. As AI continues to be woven into the fabric of societal decision-making, can we afford a technology that sees the world through a skewed lens?

Why It Matters

Understanding and addressing bias in LLMs is essential. These models, used by companies and governments alike, influence everything from hiring practices to content moderation. If they inherently favor certain cultural perspectives, the risk is perpetuating stereotypes and excluding minority voices.

The licensing race in Hong Kong is accelerating, and models like DeepSeek-1.5B could find themselves at the forefront, catering to Asian markets. Yet, the critical question remains: Will these models evolve to understand a broader spectrum of cultures, or will they continue to reflect the biases ingrained in the data they consume?

As AI technology advances, the onus is on developers to ensure models aren't only accurate but fair. This study serves as a wake-up call, urging a deeper reflection on how we teach our machines to perceive the world.