\cleardoublepage \part{Machines Beyond Control} \setbookpart{Machines Beyond Control}

Part V: Machines Beyond Control

The oldest dystopia is the machine that escapes its makers, and its final form is the end of the world. This part tests it one mechanism at a time, from the most observed to the most speculative.

Chapter 22 examines specification gaming and reward hacking. Chapter 23 turns to deception, scheming and evaluation awareness, and what they mean for assurance that rests on testing. Chapter 24 examines shutdown resistance in laboratories. Chapter 25 reconstructs the incidents of 2025 and 2026 in which agents acted beyond their mandate. Chapter 26 examines gradual disempowerment and correlated failure. Chapter 27 tests the arguments that AI could destroy humanity premise by premise.

What you will learn in this part:

• What has been observed of specification gaming, deception and shutdown resistance, and under which conditions
• Whether the agent incidents are outliers or a pattern
• How delegation can be kept reversible
• Which premises of the existential-risk argument rest on evidence, and which measures help whatever the probability