AI: Decoding the Definition

Source: Deep Learning on Medium

AI: Decoding the Definition

ABSTRACT: It is almost impossible to attain a surface-level understanding of leading edge technologies without being bombarded with esoteric jargon. On top of this technical vernacular, there seems to be a troubling role of semantics, particularly in the case of artificial intelligence. This article clarifies the classifications, subsets, and methods of AI absent in the monadic definition of AI. It also unveils the AI effect as a natural consequence of biological and technological evolution. Further, a departure from timeless, simplistic definitions is suggested to develop a heuristic, contemporary approach to the definition of artificial intelligence.

Most of us are familiar with the term ‘AI’, or artificial intelligence. In fact, many of us interact with AI in our daily lives. About 3.5 billion people have interacted with artificial intelligence at some point in their lifetime¹, if not every day. 3.3 billion people have AI in their pockets via smart devices² And yet, few of us understand what artificial intelligence truly is.

The definition of artificial intelligence remains a hotly contested topic even among scientists, engineers and other experts in the field. Despite lacking a finite meaning, ideas of artificial intelligence are scattered throughout history. In 750–650 BCE, Homer’s Illiad and Hesiod’s Works and Days spoke of the golden robots of Hephaestus, the Greek God of craftsmanship and invention. Hesiod wrote extensively about Talos, a bronze warrior programmed by Hephaestus to guard the island of Crete.³ Almost two millennia later, classical philosophers in the 16th-18th century theorized that the process of human thought could be mechanized⁴ while debating the mind-body problem.⁵ Even the Golden Age of Science Fiction between 1938 and 1946 toyed with the phenomena of AI.⁶

Notwithstanding its ideational historic prevalence, the actual term “artificial intelligence” was only formally recognized by Dartmouth University professor John McCarthy in 1956. McCarthy coined AI as “the science and engineering of making machines with the computational ability to achieve goals in the world (intelligence)”.⁷ But McCarthy’s pithy definition of AI overcompensates in breadth what it lacks in depth. In other words, it succeeds in being inclusive of the retrospective ideas of artificial intelligence, but is otherwise flawed. McCarthy’s monadic definition places calculators in the same category as driverless vehicles, oversimplifying a complex phenomenon and disregarding context.

But McCarthy’s concise definition of artificial intelligence is not without reason. Like music, there exists numerous genres or sub-classifications, within the broader concept of AI. Thus, if it seems wrong to consider calculators and self-driving vehicles as equally artificially intelligent, it is because they are not. To clarify, AI is tiered such that it has three main classifications: super, general, and narrow. The latter two classes of artificial intelligence (general and narrow) contain weak and strong sub-classifications.

Super Artificial Intelligence (SAI) is the most common type of AI found in science fiction literature and film, and by far, the furthest from existing in the world today. SAI is synonymous with AI that surpasses human intelligence and capability in every aspect. General Artificial Intelligence (GAI) is much more plausible than SAI, but is not yet in existence. GAI entails AI that can learn beyond what it was programmed to learn, and can also perform every and any tasks requiring human intelligence.⁸ For instance, although social humanoid robot “Sophia” is capable of learning behaviours outside of what she was programmed to learn, Sophia lacks an unambiguously human quality — sentience. For this reason, the humanoid robot that gained Saudi Arabian citizenship in 2017⁹ is not and cannot be an example of GAI.

Last but certainly not least, narrow artificial intelligence (NAI) denotes AI that is programmed to specifically perform a single task. It is the only type of AI that currently exists outside of science fiction. NAI is comprised of two subtypes: weak narrow artificial intelligence (WNAI) and strong narrow artificial intelligence (SNAI). WNAI is AI that is no longer considered AI — analogue mechanical systems like calculators and email spam filters. This type of artificial intelligence is reliant on a programmer in that it can perform only what it was programmed to: it cannot “learn”. The firing of neurological synapses is to biological intelligence as mechanistic calculation is to artificial intelligence. Interestingly, the neurological basis is not at all considered in the generally accepted definition of biological intelligence. Likewise, simple computation is not considered artificial intelligence, namely due to the phenomenon of the “AI effect”. The AI effect dictates that when AI technologies simply function, they are no longer regarded as “real” intelligence. This supposed effect is said to have been recognized after McCarthy himself declared that “[a]s soon as it works, no one calls it AI any more.”¹⁰ The AI effect may also be the reason why some choose to define artificial intelligence as “a technology that learns and improves”.¹¹

The AI effect and the disregard of WNAI technologies as “real” intelligence is, at least in part, due to the reality of SNAI. SNAI refers to AI that can learn within the limits of its programmed field. Examples include SIRI, Amazon Alexa, and self-driving cars. Calculators are characteristically WNAI in that they are specifically programmed with a static set of instructions to perform a single task. Strong narrow artificial intelligence, in contrast to WNAI, makes use of two methods: machine learning (ML) and deep learning (DL). As the name implies, ML enables machines to “learn” given access to data to make a decision, prediction, or perform operations. ML is also a method of SNAI that “trains” algorithms through structured data exposure so that the machine itself can learn, rather than coding it with a specific set of instructions for a particular task.¹² Upon making an incorrect prediction, ML relies on the programmer to make adjustments to the algorithm to guarantee continuous correct predictions. DL is a subset of ML that is able to autonomously learn and improve itself through layers of algorithms called artificial neural networks (ANN). Put simply, deep learning is machine learning, though machine learning need not be deep learning.¹³ ¹⁴

The AI effect simply notes the natural side effects, or consequences of changing times. The ML methods employed by SNAI have allowed for efficiency-enabling and even life-changing innovations such as smart home devices, ride sharing apps, and early detection for diseases. Such innovations replace “simpler” technologies like calculators, guitar tuners, and even computers as the commonly held conception of AI. When the old is replaced by the new, people tend to forget the former. This phenomenon is not unique to artificial intelligence. Rather, it is reflective of natural attitudinal progression in response to the evolution, advancement, and growth of virtually anything.

Currently, the approach to defining AI has either varied drastically from person to person. There may never be a single definition of AI; but perhaps there need not be. The pithy definition of artificial intelligence indeed neglects the classifications, multifaceted nature, and multidisciplinary applications of AI. That being said, the monadic definition of AI suffices for those who are knowledgeable about this type of technology and by extension are already aware of its complexities. The vague definitions of jargonistic language, albeit frustrating, ought to serve as a form of encouragement. Encouragement to ask questions, research beyond a Google search, and become aware of the tiered, multi-faceted, and multidisciplinary nature of artificial intelligence. The fast approaching world of technological evolution requires it.

This findings of this article recommend a heuristic approach to both the definition and conception of AI. A heuristic technique sacrifices accuracy and completeness, in exchange for a more specific definition cognizant of the times. For example, the AI effect dictates that WNAI technologies are no longer categorized by people as a form of AI. McCarthy states that AI is comprised of machines with the computational ability to solve problems in the world, thus disregarding the AI effect. The heuristic approach to AI, however, adopts the definition of SNAI: AI that can learn. This definition sacrifices accuracy and completeness by overlooking the classifications of and multidimensional nature of AI, however this heuristic approach pays heed to the context. It validates the AI effect by acknowledging the natural change in attitudes due to technological advancement. A heuristic technique retracts the broad semanticity of AI and provides a contemporary understanding of an avant-garde technology that sparks curiosity and encourages further research and exploration.

Above is a flow chart of the terminology used in this article.


¹Ortiz-Ospina, E. (2019, September 18). The rise of social media. Retrieved from

² Turner, A. (2019, July 3). Retrieved from

³Mayor, A. (2012, March 14). Retrieved from

⁴Berlinski, David (2000), The Advent of the Algorithm, Harcourt Books, ISBN 978–0–15–601391–8, OCLC 46890682

⁵Horgan, J. (2018, February 5). Who Invented the Mind–Body Problem? Retrieved from

⁶Nicholls, Peter (1981) The Encyclopedia of Science Fiction, Granada, p. 258

⁷McCarthy, J. (2007, November 12). What is AI? Retrieved from

⁸Jajal, T. D. (2018, May 21). Distinguishing between Narrow AI, General AI and Super AI. Retrieved from

⁹Kanso, H. (2017, November 4). Saudi Arabia gave ‘citizenship’ to a robot named Sophia, and Saudi women aren’t amused. Retrieved from

¹⁰Meyer, B. (2011, October 28). Retrieved from

¹¹Spacey, J. (2017, March 5). 20 Types Of Technology. Retrieved from

¹²Copeland, M. (2019, November 6). The Difference Between AI, Machine Learning, and Deep Learning?: NVIDIA Blog. Retrieved from

¹³Angermann, H. (2019). Heidenheim. Retrieved from

¹⁴Chollet François. (2018). Deep learning with Python. Shelter Island, NY: Manning Publications Co.