Part VIII: Safety, Interpretability, and Governance

"... an in-depth understanding of the potential risks posed by these models is needed."

Laura Weidinger et al., "Ethical and social risks of harm from Language Models"

Safety is not one layer at the top of the stack. It is a set of questions that return at every boundary: what the model is doing internally, who can oversee it, which tools it may use, what policy applies at runtime, how it behaves under attack, what it remembers, and who is accountable when the system causes harm. This part collects those questions after the machinery they must constrain has already been built up.

Chapter 54 begins inside the model, where neurons rarely mean one thing and sparse dictionaries, learned feature sets in which only a few entries fire at once, offer both a handle and a distortion. Chapter 55 asks what to do when the system being judged can outpace the judge, splitting the field into trust-building and containment approaches. Chapter 56 then moves to permissions, credentials, and the verb every tool call eventually performs. Chapter 57 and Chapter 58 deal with behavior at the boundary: guardrails, classifiers, jailbreaks (prompts that defeat a model's refusals), red teams, and the fact that a refusal is only a soft wall. Chapter 59 turns to what the weights remember and the promises made to data subjects; Chapter 60 asks how a serving path can back its privacy promise with attested hardware instead of contract; and Chapter 61 closes with the obligations owed to rights holders, users, regulators, and the public.

The part is deliberately broad because the failure modes are broad. There is no single safety switch. The useful map is the one that tracks where evidence is created, who is granted authority, when a policy becomes code, and whether a claim of control can actually be tested.