[2025 AI Ethics Seminar EP.2] Beyond Bias, the Journey Toward Fair AI - LG AI Research BLOG

AI Bias refers to the phenomenon where biases in human society are reflected in training data or algorithms, resulting in disadvantages to specific groups. This goes beyond a simple ethical issue, leading to practical risks such as performance degradation for specific user groups or the reproduction of discriminatory expressions.

Key cases include differences in accuracy by race in facial recognition technology, stereotypes about occupations reflected in image generation models, and biases in language use according to gender and race. Bias has also been confirmed in real-life systems such as Amazon's hiring system and the UK's welfare fraud detection AI.

There are several practical challenges in addressing AI bias:

• Difficulty in Information Collection: Legal and institutional constraints on collecting sensitive personal information for bias assessment
• Difficulty in Measurement and Improvement: The complexity of quantifying and technically improving the subjective concept of fairness
• Prioritization Issues: Conflicts with business KPIs and limitations in verification during the later stages of development due to concerns about performance degradation

Ultimately, the core cause of AI bias lies in biased training data. Sampling bias during data collection or the lack of representation due to insufficient data for specific groups becomes a decisive factor in producing biased results, following the 'Garbage In, Garbage Out' principle.