AI Ethics & Safety

Ethics is an engineering discipline here

AI systems increasingly decide who gets loans, jobs, bail, medical attention and information. Their failure modes are not hypothetical, and "the model decided" excuses nothing. This lesson treats ethics the way the ML path treated evaluation: as concrete practice, not sentiment.

Bias: the pipeline imports the world

Models learn the patterns in their data — including unjust ones (correlation machines, remember). Landmark cases every practitioner should know: hiring tools that penalized women's résumés (trained on a male-dominated history), face recognition with error rates orders of magnitude higher for darker-skinned women (Gender Shades study), recidivism scores with racially skewed false-positive rates (COMPAS).

Traps and their (partial) countermeasures:

• Proxies — deleting the "race" column achieves nothing when zip code, names and shopping patterns encode it. Fairness requires measuring group outcomes, not blinding yourself to groups.
• Feedback loops — biased predictions → biased actions (more patrols in flagged areas) → data "confirming" the bias. Monitor deployed impact, not just offline metrics.
• Metric plurality — several formal fairness definitions (equal accuracy across groups, equal false-positive rates, calibration…) are mathematically incompatible; choosing is a values decision that belongs in the open, documented, with affected stakeholders — not implicit in a loss function.

Practice: slice every evaluation by demographic segments; document data provenance and known gaps (datasheets, model cards); red-team before launch and audit after.

Privacy, provenance, accountability

• Models can memorize training data (extractable personal text, regenerated faces); scraping vs. consent is under active litigation, and privacy techniques (differential privacy, federated learning) trade utility for guarantees.
• Generative provenance: deepfakes and synthetic text erode "seeing is believing" — detection, watermarking and content-credential standards are the countermeasures racing dissemination.
• Accountability demands explanation and recourse: EU regulation (GDPR, the AI Act's risk tiers) increasingly mandates transparency for high-stakes automated decisions, and interpretability tools (feature attribution; the interpretable-models preference from the ML path) plus human-in-the-loop review are how teams comply in practice.

Safety and alignment

Near-term safety is largely specification and robustness: systems optimizing the measurable proxy instead of the intent (reward hacking, engagement-maximizing recommenders amplifying outrage), models confidently wrong out-of-distribution, agents amplifying small errors across long action chains. Longer-term, the alignment problem scales up: how do we guarantee that increasingly capable, increasingly autonomous optimizers pursue what we actually value — including values we struggle to specify? Serious labs treat this as a technical research field (interpretability, scalable oversight, evaluations for dangerous capabilities), not science fiction — while regulators, standards bodies and open research communities form the institutional layer.

The practitioner's creed: you are responsible for what your system does, not what you intended it to do.

Key takeaways

• Bias enters through data, proxies and feedback loops; counter with sliced evaluation, documentation and audits.
• Privacy, provenance and accountability are legal + technical requirements, not add-ons.
• Alignment = making optimizers pursue intended values; today's reward hacking is its small-scale preview.