The Elusive Definition of Artificial Intelligence

True AI doesn’t work, but surprises us when it does

Artificial Intelligence (AI) seems almost impossible to define or achieve. Like in a dream, as soon as you think you have a good sense of what it is, it dissipates and morphs into something quite different. Once you’ve achieved something in AI, the goal posts are moved farther away.

John McCarthy, who coined the term in 1956, complained that “as soon as it works, no one calls it AI anymore.” Calling something AI was to label it as impractical or experimental. And until recently, most AI was just that.

AI has always existed on the edge of what was possible. By its very nature, when it works it surprises. It challenges our expectation of what computers can (or should) do. Multiplication was thought only possible by people a century ago, but now we don’t think of it as artificial intelligence. Capabilities like map directions and playing chess games were hard AI problems of the past decades, but now are so ubiquitous and useful that we consider them trivial enough to reside on our phones.

But why is that? Why are the standards for what constitutes artificial intelligence ever rising? Humans are not really getting inherently any smarter (at least our brains are not evolving that quickly). In part we are in denial, clinging onto an identity that we are the only beings capable of intelligence. We also recall the many AI failures and broken promises along the way. For this reason we repeat the expectations cycle, swinging between despair and hype about AI, while being repeatedly surprised by its capabilities, and leaving confusion about the term in its wake.

AI may be defined by its distinctive approach to solving problems

From one perspective, AI is a collection of techniques and technologies for solving problems is a very different way.

One technique particular to AI is machine learning, a process in which generic algorithms use large amounts of data to derive relationships and detect patterns. We’ll discuss machine learning more in subsequent articles.

Also, some AI technologies are modeled on the human brain or other organic systems. By mathematically imitating the behavior of individual components of biology, like the neurons in our brain, AI can assemble massive collections of these relatively simple elements to simulate the complexity of the thought process.

Complexity of this sort seems somehow tied to intelligence. Even mechanical adding machines of the 19th century were able to surpass the capabilities of most humans by combining many small tasks (turns of a gear) into a relatively complex action and intelligent result. Electronic computers only accelerated this complexity by many orders of magnitude.

Thus intelligence, whether artificial or biological, is more than the sum of its parts. It is an emergent behavior of a complex system.

AI as a mirror to humanity

More often AI is defined by which functions it replaces.

When the steam engine replaced functions typically fulfilled by horses (e.g., transportation and work), we called it artificial horsepower. We may one day describe AI-enabled computers in terms of what quantity of “human-power” they possess. AI is when machines mimic cognitive functions that we associate with humans, such as problem solving, or when it performs a task we believe requires intelligence, such as driving a car, writing an article or identifying objects.

One famous indicator of a successful artificial intelligence is when it is indistinguishable from human intelligence. This is the famous Turing Test (or Imitation Game) originally developed by Alan Turing in 1950. In order to pass this test, an AI must interact with a person in a way that the person is unable to tell whether it is human or machine.

In this sense, AI is imitating human intelligence, or more precisely human behavior (which is often not that smart). Like with Dorian Gray, the portrait of humanity painted by AI may grant us some measure of immortality and insights about ourselves, but the image may capture some of our ugliness as well.

Intelligences are not measurable, or even comparable

Artificial Intelligence might just be different intelligence. Saying we are more intelligent than apes, means that we can do some things considered to be intelligent that they can. But a calculator is already more intelligent than we are by some measure.

For this reason it is better to think of intelligence as domain specific. In the same sense that one person may be street-smart and another math-smart. The path to increasing intelligence is lumpy. Besides just thinking faster (which is one measure of intelligence), thinking better or smarter is more qualitative.

Two Kinds of AI, Two Time-Horizons

This leads us to a distinction within AI between narrow and general intelligence. Artificial Narrow Intelligence (ANI) is the ability to do just one thing very well, perhaps better than humans. Narrow AI is here today. It is all around us , and getting better and more pervasive every day. There are lots of challenges and ethical issues related to this technology, but its power and usefulness are undeniable. Most companies and governments are working on narrow AI. Narrow AI is a servant to humanity, a specific tool in our toolkit.

Artificial General Intelligence (AGI) is the ability to do anything that most humans can do, and perhaps indistinguishable from humans. This means passing the Turing Test. But as Nick Bostrom points out, there is no reason to believe a growing AGI will stop there. Once an AGI becomes as intelligent as humans, it would likely quickly evolve to become a superintelligence, eventually smarter than all humanity together. In essence, a god to us!

Many researchers are probably exploring building an AGI, but we may never know exactly who is doing so. Some argue that it is possible that a collection of narrow AIs may merge to become an AGI. However, such a Hollywood version of this story in which someone spills coffee on a computer somehow creating a sentient AGI hellbent on our destruction seems highly unlikely. Developing an AGI is not going to be easy and may not even be possible. In any case, an AGI is likely decades away if not a century.

Most of this article series will focus on the promise and perils associated with narrow AI.

Sources include Artificial Intelligence: A Modern Approach (Stuart Russell, Peter Norvig); Superintelligence: Paths, Dangers, Strategies (Nick Bostrom). Also, see this repository of continually updated resources.

Note: this article is based on a lecture given by me at UC Berkeley, Haas School of Business, Global Access Program, “What Real People Need To Know About Artificial Intelligence” February 15, 2019