Bias in AI — Beginner-Friendly Notes
1. What Bias Means Here
Bias = systematic unfair deviation where AI performs worse or harms for some groups (not random error, but patterned). Often measured as gap in accuracy or in outcomes (hiring, loans).
2. Where Bias Creeps In (lifecycle — exam must list)
| Stage | How Bias Enters | Example |
|---|---|---|
| Data collection | Under-representation, sampling skewed, labels biased by history | Face data mostly light-skinned → dark-skinned error higher |
| Annotation | Labellers' stereotypes encoded | Toxicity labelled more for one dialect |
| Feature / Model choice | Proxy features (pincode → caste/race), model picks easiest correlation | Resume filter learns gender-correlated words |
| Evaluation | Aggregate accuracy hides group gap | 95% overall but 70% for minority |
| Deployment / Feedback | Biased outputs shape future data → loop | Policing tool → more patrols → more data → loop |
Important: bias can be data, algorithmic, or human — all three.
3. Types to Name
- Historical bias: data reflects past injustice (hiring history).
- Representation bias: some groups barely in dataset.
- Measurement bias: proxy mismeasures (e.g., arrest = crime).
- Aggregation bias: one model for all groups harms minorities (need disaggregated evaluation).
- Evaluation bias: wrong metric hides harm.
4. How to Detect & Mitigate (concept)
- Detect: disaggregate metrics by group (accuracy by gender/region/language), audit datasets, use fairness toolkits.
- Mitigate:
- Data: re-sample, augment under-represented, clean labels.
- Model: constrain for fairness (equal opportunity, demographic parity — trade accuracy), remove proxies, use adversarial debiasing.
- Process: diverse teams, stakeholder review, model cards, continuous monitoring.
5. India-Specific Note (good for viva)
Indian AI risks: language bias (Hindi-heavy, low-resource languages neglected), region/caste bias in historical records, gender bias in hiring models — SDGs require inclusion.
6. One-Line Summary
Bias is systematic unfair performance from skewed data, modelling and deployment, detectable by disaggregated evaluation and mitigated through data, model and process interventions with particular attention to representation in India's diverse context.
7. Quick Self-Check
- Why does aggregate accuracy hide bias?
- Name one bias source at data vs model stage.