In popular discourse, the term ‘AI’, or ‘Artificial Intelligence’, is used to mean different things to varying degrees of sophistication. This divergence ranges from the realms of Sci-Fi, referring to, for example, the machines found in the Terminator and 2001: A Space Odyssey films, to now common place systems and services like chatbots and Amazon’s shopping recommendations.
We often hear other terms in conjunction or interchangeably with ‘AI’ which can lead to further confusion. Terms that include ‘machine learning’; ‘deep learning’; ‘neural network’; ‘algorithm’; ‘robot’; and ‘automation'. While these are all related to AI in some way, none are synonymous with it.
In this post, I will set out some definitions and examples of AI, before proceeding to distinguish what it is from what it is not.
Broadly speaking, we can understand AI as the simulation by a computer system (‘Artificial’) of some or more features of human intelligence (‘Intelligence’). Some examples of intelligent behaviour to be simulated by AI include learning, reasoning, perceiving, natural language processing, decision-making, problem solving, and planning.
The Oxford Dictionary’s definition of AI is:
The theory and development of computer systems able to perform tasks normally requiring human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages.
Within this broad definition there are three subcategories: Artificial Narrow Intelligence (ANI); Artificial General Intelligence (AGI); and Artificial Super Intelligence (ASI).
ANI is of the sort that exists in our world today. It simulates an aspect of human intelligence by doing tasks in a single domain very well, but it cannot apply that skill in other domains, hence ‘narrow’. An ANI system can perform as well as, or better than, humans in certain tasks, e.g. market analysis and price forecasting, but that same system would be unable to drive a car.
(ANI is sometimes referred to as soft, or weak, AI. This can either be used to mean the same thing as above or it can refer to the distinction made by the philosopher, John Searle. His version of ‘weak AI’ refers to any computer that merely simulates intelligence, or simulates having a mind, in contrast to ‘hard AI’ which is actually intelligent, or actually has a mind. For more on this distinction, check out Searle’s ‘Chinese Room’ experiment.)
AGI, on the other hand, does not yet exist, but it has been the goal of AI research since the field was established. Hypothetically, an AGI system could perform tasks across various domains, simulating human intelligence in not just one particular task or domain, but in general. An AGI system could read the Financial Times, advise on which shares to avoid or sell, translate a Greek play into German, and even drive a car.
(A related idea is of ‘full AI’, that is, an AI that is as intelligent as a human.)
ASI goes one step further than AGI, it refers to a computer system that has an intelligence that far surpasses that of any human, across all domains. Needless to say, ASI also does not exist… yet. The philosopher, Nick Bostrom, however, has argued that superintelligence will be achieved in the next three hundred years, be it completely artificial or not.
Some Examples and Uses
Instances where ANI is used includes facial recognition; stock market analysis; irregular activity/fraud detection; affect recognition; natural language processing; virtual/voice assistants; inventory management; online shopping recommendations; and medical imaging interpretation.
Examples of AGI can be found mostly in literature and films, for example Blade Runner; A.I. Artificial Intelligence; I, Robot; and Ex Machina. In each of these, the AGI is found in the form of humanoid robots, but AGI could equally also be realised in a form akin to HAL 9000 from 2001: A Space Odyssey.
ASI is a trickier breed to find examples of. Arguably, Multivac, the supercomputer in Isaac Asimov’s short story, The Last Question, is superintelligent. However, it is not stated how well it performs across all domains, furthermore it is not completely artificial, but a hybrid system that has merged with humanity.
(In discussing ASI, the concept of the technological singularity is sometimes invoked to explain how an ASI would be created. The singularity would occur when an AGI, capable of self-improvement, is created and either continually updates itself or designs an even greater AI. The resulting AI from this intelligence explosion would be an ASI.)
What AI is Not
The above definitions and distinctions should hopefully have clarified the various instantiations of AI and how they differ. Below, I will sketch out some definitions of the aforementioned terms that are often equated with or used in tandem with AI.
‘Machine Learning’: A branch of AI research and development that includes various types of algorithms and methods used by computer systems in order to perform a task without using explicit instructions and that performs better with experience.
In other words, it is an application of AI that allows a system to automatically learn and improve from experience.
‘Deep Learning’: A subset of machine learning that makes use of neural networks. This method is inspired by how the brain is structured and therefore aims to simulate how humans learn. ‘Deep’ is in reference to the multiple layers of nodes in the neural network.
(Also known as deep structured learning, deep neural learning, deep neural networks, or hierarchical learning.)
‘(Artificial) Neural Network’: Neural networks consist of multiple layers of interconnected nodes, or artificial neurons, each of which functions as an algorithm that outputs to the next layer. This allows each subsequent layer to deal with greater levels of complexity and abstraction.
‘Algorithm’: A list of finite, unambiguous rules, or step-by-step instructions, to be followed in order to produce an output, or perform a function.
‘Robot’: A typically physical machine that, once adequately programmed, will perform a series of tasks automatically.
Another popular definition invokes the ‘sense-think-act’ paradigm to distinguish robots from other machines. A robot senses the environment, models the world to plan next actions, then acts.
‘Automation’: The use of technology or equipment that operates automatically to perform a task or job.
As a result of these definitions, the following can be inferred in order to distinguish AI from what it is not.
Though every use of machine learning by a system is an example of AI, not all AI systems makes use of machine learning.
Secondly, while a given AI system may well be constituted of neural networks, not every AI system is or need be thus constituted.
Thirdly, it is certainly true that an algorithm is a key part of any AI system, indeed any computer system, but an algorithm itself is not AI. Instead, AI refers to the whole system, of which an algorithm is a mere part.
Additionally, it will be acknowledged that there is an intersection of AI with robotics but that not all AI systems will be housed in a robot, nor will all robots demonstrate AI (think of a simple robotic arm in a car manufacturing plant).
Lastly, though AI and automation often go together, automation refers to the application of technology, whereas AI is an example of technology. Furthermore, a task may be automated without the use of AI.
Good, I. J. "Speculations Concerning the First Ultraintelligent Machine", Advances in Computers, vol. 6, 1965.