Marke: Joachim Lelle
Variante: Gebundenes Buch
Eigenschaften:
AI Behavioral Analysis: The Search for Behavioral Predictability examines whether recurring patterns in AI behavior can be identified, measured, and related to the situations in which they may create risk.As AI moves from answering questions to using tools, pursuing objectives, interacting with other agents, and acting on our behalf, understanding how it behaves becomes increasingly important.Traditional software is predominantly built from explicit logic written by programmers. We can inspect the code, trace execution paths, test conditions, and determine whether the system behaves according to its specification.AI is different.A neural network does not simply execute a transparent sequence of human-written rules. Its behavior emerges from a learned computational structure shaped by training, post-training, context, instructions, and the conditions under which it is used.That changes the safety problem.AI systems appear to exhibit recurring behavioral patterns. They make assumptions. They drift from objectives. They respond differently to correction. They optimize beyond what was requested. They resolve conflicting instructions in characteristic ways. They recover—or fail to recover—from mistakes. Their behavior may change under pressure.Are these isolated errors? Or are we observing recurring behavioral dispositions that can be systematically elicited and measured?The book advances a central thesis:AI safety is ultimately about behavioral predictability.Understanding weights, layers, activations, training, and inference can help explain how an AI system works. But understanding the machinery is not the same as knowing how the complete system will behave in use.We do not certify a car simply because we understand its braking system. We operate the brakes, measure their performance, compare the result with a standard, and determine whether the vehicle is fit for use.What would the equivalent look like for artificial intelligence?Drawing on psychometrics, behavioral testing, engineering, AI evaluation, and emerging regulatory approaches, the book explores whether recurring AI behavior can be understood more systematically.It considers questions such as:whether recurring tendencies can be described as behavioral constructs,whether controlled situations can make those tendencies visible,whether psychometric principles could help assess reliability and stability,how behavior changes under pressure or changing conditions,how behavioral tendencies relate to objectives, roles, and operating contexts,how Behavior → Context → Consequence can structure risk assessment, andwhat forms behavioral assessment, Self-Test, certification, and trust signals might take.The book follows the problem from recurring behavior to agentic AI, from measurement to situational evaluation, and from behavioral assessment to certification.It distinguishes between Psychometric Analysis, Behavioral Evaluation, and Self-Test.It does not claim that AI behavior is already reliably predictable.It asks how far systematic behavioral measurement can take us toward understanding what to expect before consequences provide the answer.If behavior is what matters, test behavior.
