What is the definition of artificial intelligence?

  • $\begingroup$ we don't know how to define "artificial intelligence" probably because we don't know how to define "intelligence". $\endgroup$ – k.c. sayz 'k.c sayz' Oct 20 '19 at 3:36

11 Answers 11


Over the years, many people attempted to define artificial intelligence. A lot of those definitions are summed up by Stuart Russell and Peter Norvig in their book Artificial Intelligence - A Modern Approach

The definitions of AI can be summarised as falling into the following categories:

  1. Those that address thought process and reasoning (how an AI thinks/reasons)
  2. Those that address behaviour (how an AI acts given what it knows)

Furthermore, the above 2 categories are further divided into definitions that:

I. assess the success of an AI (to do the above) based on its ability to replicate human performance

II. or an ability to replicate an ideal performance measure called 'rationality' (does it do the 'right' thing based on what it knows?)

I will cite you definitions that fit into each of the above categories:

  • 1.I. "The [automation of] activities that we associate with human thinking, activities such as decision making, problem solving, learning.." - Bellman 1978
  • 1.II. "The study of the computations that make it possible to perceive, reason, and act." - Winston, 1992
  • 2.I. "The study of how to make computers do things at which, at the moment, people do better" - Rich and Knight, 1991
  • 2.II. "The study of the design of intelligent agents" - Poole et al., 1998

In summary, AI is devoted to the creation of intelligent and rational machines that can make rational decisions and take rational actions.

I would suggest you read up on the Turing test, which Alan Turing proposed to test if a computer was intelligent. However, the Turing test has a few issues, because it is anthropomorphic.

When Aeronautical engineers created the airplane, they didn't set their goal that planes should fly exactly like birds, but rather, they started learning how lift forces were generated, based on the study of aerodynamics. Using this knowledge, they created planes.

Similarly, people in the AI world shouldn't put, IMHO, human intelligence as the standard to strive for, but, rather, we could use, say, rationality as a standard (amongst others).

  • $\begingroup$ I really like your dual explanations. Strength is not a requirement for AI imo--just has to be automated and make decisions. $\endgroup$ – DukeZhou Nov 3 '18 at 22:48
  • $\begingroup$ Most of the definitions listed are of twentieth century academic activity, not of AI itself. They are neither progress-independent nor based on measurable system capability. Bellman's is the closest to a functional definition that applies to a system, but it is grossly insufficient. A person can make a decision about which lottery ticket to buy, solve the problem of long grass by mowing it, or learn steer around their mailbox, but those are inadequate requirements for a system labelled intelligent. None of them mention incremental improvement, adaptivity, or inventiveness. $\endgroup$ – FauChristian Nov 3 '18 at 23:23
  • $\begingroup$ @DukeZhou, strength, I agree, is not a requirement for brains or simulations of them. It is a requirement for muscles and simulations of them like hydraulics and springs. Yet, what some articles call strength is really EXTENT. If someone could order books on a book shelf by the Dewey system but couldn't alphabetize their DVDs, we would wonder whether they were dumb. We wouldn't say, "How smart they are to be able to at least do one of the two." We don't want smart phones and cars to be as smart as any human. They must do things we ourselves are too tired or incapable to do. $\endgroup$ – FauChristian Nov 3 '18 at 23:51
  • 1
    $\begingroup$ @FauChristian Strength as in Strong AI comes from the works of the philosopher John Searle who introduced the Chinese room experiment to refute Alan Turing 's proposal that the Turing test implicated intelligence. Searle argued that a computer merely manipulating symbols didn't really understand, just as a man who doesn't speak Chinese is able to fool Chinese speakers that he can comprehend Chinese by matching symbols using a manual to construct replies. Searle argued that Strong AI is when a machine can be described as having a mind, when a machine not only acts intelligently, but understands $\endgroup$ – Omar K Nov 4 '18 at 5:00
  • 1
    $\begingroup$ and this is exactly the problem. Without rigorous philosophical foundations, like basically every research that uses the scientific method as its grounding, AI will always be obscured behind the problem of definitions. The early AI researchers stepped around the philosophical issues at hand just like they refused to address combinatorial explosion early on (which almost led to the end of AI research). Without a real stab at the philosophical foundations, a lot would argue that anything you say is just your feelings about AI. $\endgroup$ – Omar K Nov 6 '18 at 22:30

In the paper Universal Intelligence: A Definition of Machine Intelligence (2007), Legg and Hutter, after a quite serious study, informally define intelligence as follows

Intelligence measures an agent's ability to achieve goals in a wide range of environments

In the same paper, they also formalise this definition. You can have a look at the paper for more details, but, in a few words, in order to come up with this definition, they have looked at multiple definitions of intelligence given by people throughout the years and they have tried to summarise the key points of all these definitions. They also discuss issues like intelligence tests and their relation to the definition of intelligence: that is, is an intelligence test sufficient to define intelligence, or is an intelligence test and a definition of intelligence distinct concepts? They also point out the relation between this definition and AIXI.


In the article What is artificial intelligence? (2007), John McCarthy, one of the founders of artificial intelligence and who also coined the expression artificial intelligence, writes

Artificial intelligence is the science and engineering of making intelligent machines, especially intelligent computer programs. It is related to the similar task of using computers to understand human intelligence, but AI does not have to confine itself to methods that are biologically observable.

However, this definition is related to human intelligence, so not everyone will agree with this definition.

He further states

Intelligence is the computational part of the ability to achieve goals in the world. Varying kinds and degrees of intelligence occur in people, many animals and some machines.

The field of AI has evolved since his official conception at the Dartmouth conference in 1956, so the definition of artificial intelligence will also evolve. Previously to that conference, there were already several related fields and expressions, for example, cybernetics.


The shortest answer I could can come up with could be as follow; take it with a grain of salt though since we still do not know a lot about natural intelligence:

What natural intelligence is could be seen as the process of learning abstract concepts from limited observations with the intention to use them for solving a [new] task. This process involves using those concepts to imagine new, hypothetically correct scenarios/theories and combine them in a meaningful way to cut down the enormous hypothesis space of possibilities and enable generalization to new situations without observing any data beforehand. Artificial intelligence is to bring what natural intelligence does into machines.


There is no formal definition that most people agree on. Hence here is what I, as a data science / machine learning consultant, think:

Artificial intelligence as a research field is the study of agents which sense and act autonomously in an environment and improve their situation according to some metric with their actions.

I don't like the term, because it is too broad / vague. Instead, look at the definition of machine learning by Tom Mitchell:

A computer program is said to learn from experience 'E', with respect to some class of tasks 'T' and performance measure 'P' if its performance at tasks in 'T' as measured by 'P' improves with experience E

Machine learning is an important part of AI, but not the only one. Search algorithms, SLAM, constrained optimization, knowledge bases and automatic inference are also certainly part of AI.


Artificial intelligence researchers undoubtedly have an interest in defining the term used for their own field of discipline, and there is little ambiguity in the term artificial. The challenge is in that the word intelligence has historically been more of a qualitative description than a real number quantity.

How can the intelligence of one be compared with the intelligence of another? IQ testing, averaging college board testing categories, net worth, Chess and Go tournament victory, wrong decision rates, various intellectual speed races, evaluation boards and panels have been a painfully inadequate set of kludges for those involved in the mathematical theory behind the mental capabilities we call intelligence.

Only a century ago, intelligence was a qualitative term people associated with the ability to find solutions to problems in academia, business, and personal life. As culture began to seek quantitative treatment of things once solely qualitative, the dependency of mental capability on a person's age and and their environmental opportunities introduced a challenge. The idea of intelligence quotient (IQ) grew out of the desire to quantify mental potential independent of age and opportunity.

Some have attempted to minimize environmental factors by producing standardized testing of fundamental cognitive skills as they apply to mathematics and language.

Production system and fuzzy logic containers (rules based), deep learning (artificial network based), genetic algorithms, and other forms of AI research have not produced machines that can score well in standardized tests designed for humans. Yet the natural language capabilities, mechanical coordination, planning excellence, and the drawing of conclusions based on clear and verifiable reasoning continue to be sought in machines.

The below are categories of mental capabilities, distinguished by their methods of measurement, architecture of use, and the kinds of research that have produced early promising results and continuous improvement.

  • Dialog — measured both by Alan Turing's proposed imitation game and through the viability of answering system automation, personal assistants, and mobile chat-bots
  • Mechanical control — measured both by system stability criteria and by rate of incident cost and reduction in loss of life in the case of transportation sector use of intelligent automation
  • Business intelligence — measured primarily by increase or decrease in profitability in relation to trends prior to or concurrent with manual planning and operational control

The discovery of the most likely and optimal parameterization for a complex function based on some mathematical expression of what optimal means is deliberately not listed above. What is the central activity for machine learning devices does not fit well into the categories of what has historically been called intelligence, nor should it be. Statistical treatment of data sets for predictive purposes is not learning in the intellectual sense. It is surface fitting. Machine learning is currently a tool to be used by human intelligence, to extend its power, like other computational tools.

This constraint on machine learning may, in the future, be transcended. It is not known if and when artificial networks will demonstrate cognition, logic, the ability to recognize significance, and effective capability in the categories listed above.

One of the difficulties in defining intelligence is the lack of unanimity regarding its dimensionality. If intelligence is to be quantified, the value of $n$ where the measure of intelligence $\mathcal{I} \in \mathbb{R}^n$ is significant. The concepts of g-factor and IQ imply that $n = 1$, but several critics of this ideology, such as Howard Earl Gardner, Ph.D. and Thomas Armstrong, Ph.D. have proposed that there are multiple dimensions to intelligence.

  • Linguistic intelligence (“word smart”)
  • Logical-mathematical intelligence (“number/reasoning smart”)
  • Spatial intelligence (“picture smart”)
  • Bodily-Kinesthetic intelligence (“body smart”)
  • Musical intelligence (“music smart”)
  • Interpersonal intelligence (“people smart”)
  • Intrapersonal intelligence (“self smart”)
  • Naturalist intelligence (Armstrong's addition)
  • Existential intelligence (Armstrong's addition)
  • Moral intelligence (John Bradshaw, Ph.D., addition)

The argument that these are all manifestations of a single intelligence capability expressed in varying effectiveness due to education or other training has been systematically weakened by evidence-based discovery in the fields of cognitive science, genetics, and bioinformatics.

In genetics, at least twenty-two independent genetic components to intelligence have been identified, and that number is likely to grow. These independent switches in human DNA do not all impact the same neural controls in the brain, indicating the evidence-based weakness of the g-factor ideology.

It is likely that some of the forms of human intelligence and DNA expression map in complex ways that will be discovered over time and that this mapping will entirely replacing the g-factor simplification over time.

The term Artificial Intelligence may be better expressed as Simulations of the Forms and Expressions of Human Intelligence and merely abbreviated as AI. However that is not a definition. It is a rough description. There may never be a single precise definition for all the dimensions we loosely group under the single term. If that is the case for human intelligence, then it may also remain true for artificial intelligence.

There are some common characteristics one can list about all intelligent responses.

  • Intelligence can only be measured and obtain usefulness within the context of a particular environmental condition and some objective or set of objectives. Examples of objectives include staying alive, obtaining a degree, negotiating a truce amidst a conflict, or growing assets or a business.
  • Intelligence involves adaptation to unexpected conditions based on what is learned through experience, therefore learning without the ability to apply what is learned is not intelligence and applying a process that was learned and simply transferred to some one or some thing controlling the a process is also not considered intelligence.

Human intelligence can learn and apply in what seems to be a concurrent fashion. Beyond that, it would be amiss to discuss a working definition to intelligence without mentioning some of the key human mental abilities that have been proposed as recursion on lesser forms, but proof that recursion or composition will produce these mental features does not exist.

  • Concurrent learning and use of what has been learned
  • The ability to invent new mechanisms of incremental improvement
  • Adaptivity to unexpected conditions
  • The ability to invent structure outside currently learned domains

Future requirements for intelligent machines may include these, and there may be some wisdom to include them now.


Scripts, plans, goals, and understanding: An inquiry into human knowledge structures, Schank, Abelson, 2013, cited by 16,689 articles, T&F excerpt: In the summer of 1971, there was a workshop in an ill-defined field at the intersection of psychology, artificial intelligence, and linguistics. The fifteen participants were in various ways interested in the representation of large systems of knowledge or beliefs.

Understanding Our Craft — Wanted: A Definition of Intelligence, Michael Warner, 2002

The Concept of Intelligence and Its Role in Lifelong Learning and Success, Robert J. Sternberg, Yale University, 1997

Some Philosophical Problems From the Standpoint of AI, John McCarthy and Patrick J. Hayes, Stanford University, 1981

Understanding and Developing Emotional Intelligence, Olivier Serrat, Knowledge Solutions, pp 329-339, 2017

Frames of Mind: The Theory of Multiple Intelligences, 2011, Howard Gardner

7 (Seven) Kinds of Smart: Identifying and Developing Your Multiple Intelligences, 1999, Thomas Armstrong

Genome-wide association meta-analysis of 78,308 individuals identifies new loci and genes influencing human intelligence, Suzanne Sniekers et. al., 2017

  • $\begingroup$ This definition of Intelligence focus strongly on a scientific background which is grounded in the human DNA. What is missing is the social component of inventing magic tricks. The first example of robotics were so called Automaton made by the wrongplayer Wolfgang von Kempelen. The idea was to deceive the people. Betting on horse races (Ada Lovelace), crap games and fast calculation in the human brain can all be seen as cheating done by illusionist. $\endgroup$ – Manuel Rodriguez Nov 4 '18 at 8:31


A measure of the strength of a decision-making agent relative to other decision-making agents, in regard to a given task or set of tasks. The medium is irrelevant—intelligence is exhibited by both organic and intentionally created mechanisms. May also be the capability to solve a problem, as in the case of a solved game.


Relates to the term artifact, a thing which is intentionally created. Typically this term has been used to connote physical objects, but algorithms created by humans are also regarded as artifacts.

The etymology is derived from the Latin words ars and faciō: "To skillfully construct", or, "the art of making".

Artificial Intelligence

  • Any decision-making agent that is skillfully (intentionally) constructed.

APPENDIX: The meaning of "intelligence"

The original meaning of "intelligence" seems to be "to acquire", back to the Indo-European. See: intelligence (etymology); *leg/*leh₂w-

The OED 1st definition of intelligence is not incorrect, extending the meaning to acquisition of capability (demonstrable utility), just that the second definition is the older and fundamental: "The collection of information of [strategic] value; 2.3 (archaic) Information in general; news."

You can regard the universe as being comprised of information, whatever form that information takes (matter, energy, states, relative positions, etc.) From the standpoint of an algorithm, this makes sense since the only means they have to gauge the universe are percepts.

Take a flat text file. It may just be data, but you could try and execute. If it actually runs, it might demonstrate utility at some task. (For instance, if it is a minimax algorithm.)

"Intelligence as a measure of utility" is itself "intelligence" in the sense of information, specifically that information by which we measure intelligence, as a degree, relative to a task or to other intelligences.

  • $\begingroup$ Note, this also hews to the Russell&Norvig basic definition of intelligence, rooted in utility. Sans utility, there is no meaningful definition of Intelligence, at least not in the sense of being concrete or practical. $\endgroup$ – DukeZhou Mar 22 '19 at 20:20

A.I is basically the act of implementing human intelligence in machine. This is done through various algorithms that implement human intelligence.


AI is a field that uses computation techniques to approximate complex decisions.

  • 1
    $\begingroup$ Can you explain your use of "approximate"? (It's an interesting choice that I think is worthy of clarification!) $\endgroup$ – DukeZhou Nov 3 '18 at 22:51

More Conventional: A computer program (mostly) that can calculate outputs for arbitrary inputs it has never seen before, pre-programmed for or is not provided with an explicit relationship between inputs and outputs (i.e. domain and range). Google Search, Alexa, Siri, Cortana, IBM Watson... The definition applies to all of them; even for General Purpose AI

I am going one step further (controversial!). If you remove A non-human entity from the first definition, that is the definition for human intelligence, for me. For instance, RMBs can infer some hidden abstract meaning from data during unsupervised pre-training. We may call this intuition for us, but it seems it is not unique to humans. (Geoffrey Hinton's cat recognition experiment is a good example but couldn't find a link). RBMs can also dream. So Maybe the human intelligence, which we perceive almost like a supernatural phenomenon, can be modeled by a mathematical model no matter how complex it may be. Thus, before judging my reduction of AI to a composition of functions (roughly speaking), judge my human intelligence argument. Here is a video of Geoffrey Hinton on the subject

Machine Learning: Machine Learning is the process of optimizing a function's parameters for given inputs and outputs so that it can calculate new outputs for new inputs. Even Linear Regression is a type of machine learning and a Deep Neural Network is actually a function. It is used interchangeably with AI but they don't mean the same. AI answer WHAT while Machine Learning answers HOW. (Not exactly, but close)

Let me give you some examples in order to clarify AI and ML difference.

  • Deep Learning is not AI. It is ML.
  • Amazon's Alexa is an AI.
  • Optimizing a Genetic Algorithm (GA) is ML. A bot playing Snake game using that GA parameters is an AI.

NOTE: However, currently all the methods and structures we use to build AI falls under the term Machine Learning. Thus, it is correct to say that we use Machine Learning to build Artificial Intelligence.

  • $\begingroup$ Let me summarize: AI is the blackbox between input and output, similar to the “process” box in the IPO model. And the computation in the process box is done by by machine learning. On the first look, this explanation is short and exact, but it describes not what AI is, it defines only classical programming. The IPO model is used to determine what programmers are doing. They define the rules for transforming input into output. If some kind of blackbox in the process model is equal to AI, why are thousands of papers written each years about the subject? $\endgroup$ – Manuel Rodriguez Nov 5 '18 at 15:38
  • $\begingroup$ @ManuelRodriguez "why are thousands of papers written each year about the subject?" I don't know how to answer this. Can you ask your question differently? $\endgroup$ – ozgur Nov 5 '18 at 17:39
  • $\begingroup$ Suppose, AI is equal to a linear regression function between input and output values. Solving AI can be done with machine learning, that means, the algorithm will find a mapping. I believe this assumption is too easy, because many academic papers are written about non-machine learning topics like biped walking, human vision and semantic understanding. It seems, that AI is located outside of machine learning and has to do with knowledge itself. $\endgroup$ – Manuel Rodriguez Nov 5 '18 at 18:04
  • $\begingroup$ @ManuelRodriguez I agree that AI is a more abstract concept. AI to ML is like Turing Machine to Real Computer. Implementation and methodology cannot contain the concept it lives in. BTW, I never said AI is a function. I said Machine Learning is optimization of a function. Meaning that a Deep Neural Network is actually a function. And it is extremely difficult to train a DNN, let alone to find the global minimum. Moreover, checking if we found the global minimum is NP-Hard, almost impossible. $\endgroup$ – ozgur Nov 5 '18 at 18:20
  • $\begingroup$ @ManuelRodriguez I edited my answer for worse=) you may want to read it. $\endgroup$ – ozgur Nov 7 '18 at 6:56

It's an intelligence at the machine level rather than shown by human beings which are driven by the algorithms.


Not the answer you're looking for? Browse other questions tagged or ask your own question.