AI Ethics & Bias
AI Reflects Its Training Data — Flaws Included
AI systems encode the biases present in the data they were trained on. This is not a fringe concern — it has measurable, real-world consequences in hiring, lending, medical diagnosis, and law enforcement. Understanding how bias enters AI systems is the prerequisite for catching it before it causes harm.
- Representation bias: if a group is underrepresented in training data, the model performs worse for them — not through malice but through statistical underexposure
- Historical bias: training on past decisions bakes in past discrimination — the model learns that certain hiring or lending outcomes were "correct" even when they weren't fair
- Measurement bias: if a proxy metric is flawed (e.g. using arrest records as a proxy for criminality), so is the model
- Feedback loops: biased predictions create biased outcomes, which become future training data — bias can compound over time
Where Bias Shows Up in Practice
Bias in AI is most dangerous in high-stakes automated decisions — the ones that affect livelihoods, access to credit, healthcare, and justice. The field is past theoretical concern; there are documented cases of consequential AI bias across industries.
- Hiring: resume-screening AI trained on historical hires can encode gender or ethnic bias from past hiring managers' decisions
- Lending: credit-scoring models can disadvantage neighbourhoods or demographic groups based on proxies correlated with race
- Healthcare: diagnostic AI trained predominantly on one population may perform worse for underrepresented groups
- Security: threat-detection models trained on historical incident data may encode the blind spots of past analysts
- Content moderation: models trained on English-dominant data perform poorly on minority languages and dialects
Mitigating Bias: What Works
Bias cannot be eliminated post-training with a patch — it must be addressed throughout the AI development lifecycle. The most effective mitigations address data, evaluation, and deployment simultaneously.
- Diverse training data: curate for demographic representation, not just volume — larger datasets with systematic gaps are still biased
- Disaggregated evaluation: test model performance broken down by relevant subgroups, not just overall accuracy — aggregate metrics hide subgroup failures
- Bias audits: structured third-party reviews before deployment and on a recurring schedule after
- Human review: for high-stakes decisions, require human sign-off rather than fully automated outcomes
- Red-teaming for bias: actively try to find discriminatory outputs — if you're not looking, you won't find them
Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.