RESPONSIBLE AI, COMPLIANCE, AND RISK

LEARNING GOALS By the end of this chapter you will be able to describe bias and fairness testing, explain privacy and data protection, describe auditability, name risk management frameworks, and explain human oversight and accountability. BIAS, FAIRNESS, AND HARMFUL-OUTPUT PREVENTION Bias means a systematic unfair tilt in outcomes. A model learns from human-written text, and that text contains human prejudices, so a model may treat groups of people differently without anyone intending it. Fairness means similar people receive similar treatment regardless of protected characteristics, such as race, sex, age, religion, or disability. This matters most when an agent helps decide something important about a person, such as hiring or lending. A common test uses matched pairs. You prepare two inputs that are identical except for one detail, such as the applicant's name, and compare the results. If the outcomes differ, you have found a problem to fix before launch. Harmful-output prevention uses filters, refusals, and checks, as in Chapter 36, to stop content that is dangerous, hateful, or deceptive.