AI Ethics & Bias
Every technique covered so far in this course learns patterns from data. That fact has a direct, serious consequence: if the data reflects real-world inequities, the model learns those inequities too, and can apply them at a scale and speed no individual human decision-maker ever could.
Where bias actually comes from
A model has no independent judgment about fairness — everything from the machine learning basics lesson still applies here. It only knows what it can infer from the statistical patterns present in its training data. If historical hiring data reflects decades of a company disproportionately hiring one demographic, a model trained to predict "who gets hired" from that data will learn that pattern, too — not because anyone told it to discriminate, but because that's the pattern actually present in the examples it was shown. This is why "the model is just doing math, it can't be biased" is a misleading claim: the math is applied faithfully to biased inputs, and faithfully reproducing a bias is still reproducing it.
A real, well-documented example
A widely cited 2018 study (Joy Buolamwini and Timnit Gebru's "Gender Shades" research) tested several commercial facial-analysis systems and found they were consistently far less accurate at classifying the gender of darker-skinned women than lighter-skinned men — error rates differed by tens of percentage points between the best- and worst-performing groups. The cause traced back to training data: these systems had been trained and evaluated overwhelmingly on lighter-skinned faces, so the patterns they'd learned simply didn't generalize as well to faces underrepresented in that data. This is a direct, measured consequence of the same "models fail on what they haven't seen" idea from the computer vision lesson, with real consequences when such systems are deployed in things like security or hiring tools.
Fairness as a design constraint, not an afterthought
The practical lesson from cases like this is that fairness has to be actively designed for and measured throughout a system's development — checking a model's error rates across different demographic groups, not just its overall average accuracy, and deliberately building training data that represents the population the system will actually be used on. Bolting a fairness check onto a finished model as a last step, after the training data and objectives are already locked in, tends to catch problems too late to fix cheaply.
The limits of "fixing" bias
It's worth being honest that "bias-free AI" isn't a real, achievable end state — there's no neutral dataset that reflects no historical patterns at all, and different definitions of "fair" can mathematically conflict with each other (a model can't always simultaneously satisfy every reasonable fairness criterion at once). The realistic goal is measuring specific, known biases, being transparent about a system's limitations, and deciding deliberately which tradeoffs are acceptable for a given use case — not claiming a system has achieved perfect neutrality.