WHAT ARTIFICIAL INTELLIGENCE (AI) IS AND IS NOT

NARROW AI VERSUS GENERAL AI Researchers often describe two broad kinds of AI. Narrow AI, sometimes called weak AI, is a system built to do one kind of task, or a group of closely related tasks, and which cannot do much outside that area. A program that recognizes stop signs in camera images is narrow AI. A program that translates English into Spanish is narrow AI. Even a chatbot that can discuss thousands of subjects is, by the usual way of counting, still narrow AI, because it learned to work with language and does not possess a general ability to learn and do anything a person can do. General AI, often called Artificial General Intelligence (AGI), is the idea of a system that could learn and perform almost any intellectual task that a human can, moving flexibly between unrelated problems, using common sense, and setting its own goals. General AI is a goal that some researchers pursue and a subject of lively debate. At the time this book was prepared, no deployed system is accepted by the research community as general AI. Today's most capable systems are impressive and broad, but they still make surprising mistakes, they cannot reliably carry out long projects without support, and they do not set goals for themselves in a trustworthy way. Why does this distinction matter for a harness engineer? It matters because it tells you what to promise and what to build. If you assume the system has general, human-like judgment, you will trust it with tasks it cannot safely do. If you treat it as a powerful but narrow tool, you will surround it with checks, limits, and verification, which is exactly what a harness is. The responsible stance is this. Treat today's AI as narrow, and design the surrounding system so that when the AI is wrong, which it sometimes will be, the damage stays small and is caught quickly.