Facialabuse-gaia-3 Patched
Facialabuse‑GAIA‑3: An Exploratory Essay on the Concept, Context, and Consequences
4‑3. Bias & Fairness
- Training Dataset: The original GAIA‑2 was trained on the AffectNet‑EU corpus (≈5 M faces from 30 European countries). Subsequent audits uncovered under‑representation of darker skin tones and older adults, leading to higher false‑negative rates for “sadness” in those groups.
- Mitigation: GAIA‑3 includes a bias‑aware calibration step that re‑weights AU detection based on skin reflectance and age metadata. Independent labs (e.g., the Algorithmic Justice League) have given it a C‑grade, noting residual disparities of up to 8 % in recall for certain demographics.
