دانلود کتاب انصاف در هوش مصنوعی مولد: راهنمای عملی برای معیارسنجی، کاهش سوگیری و مهندسی مسئولانه هوش مصنوعی

As AI becomes more deeply integrated into the structures and institutions that shape people’s lives, the fairness of these systems becomes an increasingly urgent concern. The obligation to account for and mitigate the harm caused by AI systems is shared by everyone who designs, deploys, maintains, or governs them. This book provides readers with a practical guide to meeting that responsibility. I met Prasanna Vijayanathan through our work on the IEEE P3462 working group, Safety by Design in Generative Models to Prioritize Child Safety, where he led the group responsible for developing guidelines for responsible model deployment. Since then, I have had the opportunity to see him speak and write about AI safety and fairness in different contexts. He consistently impresses me with his ability to bring rigor to difficult questions without losing sight of the practical constraints involved in building tools with an emerging technology. Safety and fairness are overlapping but distinct concerns. Safety asks how a system can operate without creating unacceptable risk. Fairness asks how its benefits and burdens are distributed, whose perspectives inform its design, and whether the people affected by its decisions can meaningfully challenge or correct its harms. A system may avoid dangerous outputs while still systematically advantaging some groups over others. However, an outcome is not acceptable merely because its harms are distributed uniformly. An ethical approach to AI requires attention to both. This distinction is one reason Prasanna’s combination of rigor and pragmatism is valuable. Fairness is easy to endorse in principle but difficult to define, implement, and sustain in practice. Meaningful improvement begins by making disparities visible. When used thoughtfully, data and benchmarks can illuminate patterns of harm previously obscured, but no single score can capture fairness. Fairness should not be addressed through a one-time review or delegated solely to a separate ethics or compliance team. It requires explicit design choices built into the systems about whose experiences matter, which harms are evaluated, how competing values should be weighed, and what signals should inform deployment, updates, and retirement. Generative systems make this work especially challenging because their open-ended outputs can cause harm through stereotyping, erasure, and unequal quality of service. Fairness in Generative AI helps turn those questions into a plan of action. It connects the moral and ethical foundations of fairness with the engineering work of creating scenarios, metrics, scorecards, monitoring systems, and incident-response processes. The book’s central insight is that a benchmark can serve as a practical and enforceable agreement among engineers, product leaders, policymakers, external reviewers, and affected communities: a transparent statement of what matters and how to account for it. By incorporating it into the design of the system, fairness becomes an intrinsic governance principle. The result does not guarantee that every question of fairness is addressed. However, this approach provides something more credible: an extensible framework for bringing assumptions and patterns of harm into view, testing systems rigorously, recognizing limitations, and improving fairness over time. This book is for anyone who wants to move from good intentions to good practices.