Human-Level Artificial Intelligence Must Be an Extraordinary Science

Aiming to create a cognitive system with human-level intelligence is as different from normal scientific objectives as reaching artificial immortality is to the goals of modern medicine. Most researchers in artificial intelligence, along with the institutions that support them, advocate that AI research should conform to normal scientific standards and methods. I argue that these are often incidental and even antithetical to achieving human-level intelligence and that a different approach is required. In this essay, I propose some principles on which to base such an approach. 1. Why Human-Level Artificial Intelligence? Developing cognitive systems with human-level intelligence would have tremendous scientific and technological ramifications. However, this objective is significantly different along many dimensions from the objectives of most fields in science and technology. Unless this is recognized, and unless methods and institutions in these fields are changed significantly, we are unlikely achieve the benefits of human-level artificial intelligence in our lifetimes. For the purposes of this paper, we will say that a system has human-level intelligence if it can accomplish the same cognitive tasks as a human. For example, a system with human-level intelligence would be able to have human language conversations as well as people and be able to identify objects and interpret visual scenes as well as people. Deciding specifically what counts as a cognitive act and what it means to perform it at the same level of a human is not important for this paper, since the field is far from the point at which such fine distinctions will be important. Understanding and creating human-level intelligence are tremendously important scientific and technological objectives. The key obstacle to achieving these, which I call the intelligence puzzle, is to understand how a system composed of unintelligent parts (such as neurons or transistors) can behave as intelligently as people. Without understanding this, we cannot explain how intelligence emerges from natural scientific laws and thus the project of natural science is incomplete. This is also a serious problem because so many important areas of study (e.g., economies, culture, organizational behavior) concern individual humans and how they interact, and thus involve the exercise of human intelligence. In this paper, I will refer to the endeavor to create systems that exhibit human-level intelligence as “human-level artificial intelligence” (“HLAI”) and the fields interested in creating intelligence at any level (such as artificial intelligence, machine learning, and computational linguistics) as “artificial intelligence” or “AI”. The most convincing evidence that the intelligence puzzle has been solved would be to build a system with human-level intelligence. The technological implications of this would be at least