Definition
AI bias arises from skewed data, labeling practices, objective choices, or deployment context — producing unequal error rates or harmful stereotypes.
For language AI, underrepresenting languages like Somali is a form of allocation and representation harm.
In simple terms
A map drawn only from one neighborhood will mislead you about the whole city — models trained on narrow data inherit that blind spot.
Where you see it
- Face analysis failing more on darker skin tones in early systems.
- English-centric LLMs underperforming on African languages.
- Hiring tools reflecting historical workforce imbalances.
How it works
1.Audit data
Who is represented? Who labeled it?
2.Measure disparities
Slice metrics by group or language.
3.Mitigate
Rebalance data, change objectives, add human oversight.
Why it matters
- Bias is not just ethics theater — it is product quality and trust, especially for global and multilingual AI.
Often confused
Removing protected attributes removes bias.
Proxy variables often recreate the same disparities.