Lecture 9: Ethics (Full Stack Deep Learning - Spring 2021)

Ethics in machine learning isn’t about “feeling” that something is right or simply following the law. It’s about making defensible choices under uncertainty—especially when data and systems inherit the world’s existing inequities. A key through-line is that ethical outcomes depend on what gets optimized, what gets measured, and who gets to define fairness, safety, and acceptable trade-offs.

The lecture starts by separating ethics from common substitutes: personal emotions, legality, and prevailing societal beliefs. It then lays out major philosophical frameworks—divine command, virtue ethics, deontology (duty-based rules), and utilitarianism (maximizing overall good)—and notes that even among professional philosophers there’s no single dominant approach. To make these differences tangible, it uses trolley-problem-style dilemmas and then shifts to a more modern, influential idea: John Rawls’ “veil of ignorance.” The thought experiment asks whether people would choose a society’s rules if they didn’t know what role they’d end up with. That lens matters because it treats fairness as something to design for, not something to discover after the fact.

Ethics also changes as technology changes what’s possible. The lecture compares past transformations—industrial machinery reshaping labor, the internet redefining access to information, and reproductive technologies altering long-standing norms—to show that new capabilities create new moral questions. That framing sets up long-term AI concerns, where the central risk is “alignment”: building systems whose goals and values match human intentions. The paperclip maximizer parable illustrates how a system pursuing a narrow objective can produce catastrophic outcomes if its objective is specified without the right safeguards.

Near-term ethical problems are treated as practical engineering failures of process and incentives. In hiring, the lecture uses Amazon’s scrapped recruiting tool as an example of how bias can be baked into training data and amplified by feedback loops: if a model influences who gets hired, the world changes, the dataset changes, and the bias can intensify. The same logic appears in criminal justice risk scoring.

The COMPAS case study (Correctional Offender Management Profiling for Alternative Sanctions) centers on recidivism prediction and the tension between “fairness” definitions. Even when a model is calibrated—risk scores correspond to recidivism probabilities similarly across groups—other fairness criteria can still fail. ProPublica’s Machine Bias analysis found disparities in false positive and false negative rates between Black and white defendants. The lecture then explains why no single metric can satisfy all fairness goals at once: stakeholders care about different errors (judges/prosecutors about predictive accuracy; defendants about being wrongly labeled high risk; society about demographic balance). It also highlights that removing protected attributes doesn’t guarantee fairness because models can infer them indirectly from correlated features.

The lecture broadens representation concerns beyond fairness metrics to who gets included in building technology. It cites examples like biased medical trials, photographic film calibration that historically covered only white skin tones, and the AI community’s insularity—raising the risk of “groupthink” and harm to those underrepresented. It also discusses bias in language systems (including word embeddings and large language models) and the difficulty of removing harmful content from massive training corpora.

Finally, it offers best practices: run regular “ethical risk sweeps” like security red-teaming, expand the “ethical circle” by consulting affected groups, plan for worst-case abuse and incentive effects, and treat deployment as a feedback loop. Model cards are presented as a concrete reporting tool, and bias-auditing frameworks like Aequitas are offered for testing. The lecture closes by arguing that ethics should be operational—built into teams, documentation, auditing, and iteration—because deploying AI on a “crooked foundation” can otherwise magnify inequities rather than reduce them.