Defining Intelligence

To be able to discuss the subject of intelligence, we must first define the term. People are intelligent in many different ways. Different types of intelligence[1] can be characterized by how a person perceives the world, processes that information, and subsequently acts on that perception.

The most common meaning of the word “intelligence” centers on a person’s ability to understand and communicate complex ideas, which is demonstrated by a person’s ability to perceive their environment and effectively communicate their understanding of the situation to those around them. Athletes possess a high level of intelligence along an additional dimension: “physical intelligence,” which is demonstrated by their ability to quickly respond to their environment through precise and powerful physical motions. Other people are said to be exceedingly “emotionally intelligent” when they perceive other people’s emotional states and respond in an empathetic, understanding, or otherwise socially-appropriate manner [2]. Artists are extraordinarily intelligent at painting, drawing, or sculpting, just to name a few of the media through which their heightened perception of the physical properties of the world results in their actions manifesting as art.

Those with greater intelligence in a certain area possess a more accurate, rapid, or heightened perception of the world along that dimension and are able to take actions that influence the world to a greater effect than those who are less intelligent in that area. The outcome of intelligence applied to a problem can range from winning a debate due to one’s book knowledge or mastery of logic, to quick perceptions and reflexes in a sports match, or a deeper understanding of colors and the ability to imitate a beautiful scene on canvas. Therefore, when we say “intelligence,” we refer to one or more types of intelligence.

The phrase “general intelligence,” popular in the Artificial Intelligence (AI) vernacular of today, implies high performance, or the capacity for rapid adaptation, across all types of intelligence. Adding the word “embodied” when discussing the intelligence of a physical agent conveys the fact that the knowledge must be grounded in real-world objects and the outputs are capable of effecting change in the physical environment [3]. The phrase “super intelligence” implies one or more types of intelligence to an extreme degree, potentially out-pacing humans along this dimension [4], [5]. Finally, the term “artificial intelligence” is punctuated by the term “artificial,” which refers to an entity that is made by humans, and is not natural. Because all judgments of intelligence originate from human perception, the implication is that “artificial intelligence” must aim to replicate humans’ “natural intelligence,” so that it is perceived as intelligent by the people evaluating it. The “value alignment problem” in AI aims to address a corollary of this challenge: ensuring that as artificial systems scale in these various types of intelligence, their actions and objectives remain aligned with human values and perceptions [6].

While human intelligence can peak along specific dimensions, I propose that “generality,” the underlying capacity to adapt to novel situations, is what differentiates true intelligence from mere competence at a specific task. Although “artificial general intelligence” is typically considered to contain a superset of capabilities found in narrower “artificially intelligent” systems, without general-purpose adaptability, such systems lack the capability to handle new problems and situations as a human naturally would. Relative to artificial intelligence systems today, the average human is exceedingly capable of understanding new ideas and adapting to new situations. For instance, adversarial attacks have been developed against image recognition systems[7] and large language models¹. These attacks send subtly modified pixel data to image classifiers or long, repetitive prompts to language models, causing them to make errors that a casual human observer would never make, even if presented with new, unseen information. Consider another example, the current state of self-driving cars. Waymo and Tesla, leaders in the space, operate fleets of cars that have driven billions of miles, but they both rely on strategies to avoid extreme novelty. Waymo geofences its vehicles to well-mapped areas and Tesla requires constant human supervision. Both strategies intentionally limit vehicles’ exposure to the long tail of edge cases that human drivers handle intuitively every day. Such shortfalls of AI systems indicate that scaling intelligence along one or even several dimensions is not enough to replicate baseline human intelligence. True intelligence requires a foundation of general adaptability.