Artificial Intelligence will probably be remembered as the least well-defined technological advancement of humankind.
Ever since the two words were reportedly put together in a sentence - and we know when that happened: at a workshop in Dartsmouth in 1957 - the term "Artificial Intelligence" has been morphing its intended or perceived meaning, such that even today, when arguably artificial intelligence is so close and present in our daily life that we can literally touch it, confusion reigns supreme. When one has to deal with the word intelligence one may expect that some confusion will arise, because intelligence in itself is a very hard to define concept; however, the combination with the apparently innocuous adjective artificial has blown the uncertainty out of scale. So let me first comment on what we mean by artificial.
Artificial, in Webster's dictionary, is defined as "something made or contrived by human skill and labor rather than occurring naturally". That source also reminds us that the adjective is commonly used to indicate something that is feigned, fictitious, or insincere; but these additional meanings have little to do with our confusion in defining AI . If we instead turn to Oxford's dictionary, we get it defined as "anything made or produced to copy something natural, and explicitly not real". Alas, I would have been happy to stay with Webster's definition, as Oxford throws in quite a bit of dust by arising the doubt that when we speak of AI we hint at something fake: that possible meaning was also mentioned above, but it was a sideline; here instead we get it in the package. In summary, already the adjective in AI is bringing with it a degree of fuzziness - are we saying some intelligent behavior is artificial to praise the cunning of creating a system exhibiting it, or to blame its not genuine nature?
Shaken a little bit by the above preamble, we now turn to the noun. Intelligence is in and of itself one of the most debated concepts around, and adding a drop to that sea of debates is not something I wish to entertain myself with in this column. Suffices for me to say here that one may find intelligence in the behavior of bacteria, all the way up to identifying intelligence in the design of life on Earth, and beyond; and at the same time, one may deny that any of those systems are intelligent, by a slight tweak of the definition of the word, well within the spread of meanings accepted in practical use of our language.
So, when you put together A and I into AI you get a concept that is disturbingly flexible. Is there artificial intelligence in a windshield wiper or in a thermostat? Those systems are undeniably artificial - they do not grow on trees - and arguably intelligent; they perfectly well fit in a definition of artificial intelligence that at first sight looks quite sound: "Intelligence demonstrated by machines - instruments that can perceive the surrounding environment and act to fulfil a purpose."
Other definitions I have seen floating around include Tom Mitchell's practical rule: "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." This explains what learning is, but ignores the concept of intelligence.
At the other edge of the spectrum, we are sometimes still debating whether our most powerful language models are intelligent: naysayers will wear your patience down by pointing out that these are just large neural networks that put together one token after another: they do not "know" what they are talking about.
I find that argument utterly dumb, as it completely neglects to consider that human intelligence may not be doing anything different. To me, an operational definition of AI is all we can afford in this Babel of meanings and counterarguments.
And today I read another definition that I found objectionable. Dahlike in 2024 proposed this: "Artificial intelligence refers to digital machine-based systems that process inputs, infer concepts, and generate outputs for explicit or implicit objectives. It can be expected to perceive and influence its environment to varying degrees of autonomy. The purpose of fulfilling these objectives is to automate human labor and/or augment human intelligence."
The above definition is a good bid, but it raises a number of issues in my opinion. First of all, it rearranges the task of defining AI by leveraging other concepts that need a definition of their own, like "infer concepts" - how do we infer concepts, and when and how can we claim a machine can do the same? Also, probably the last part of the sentence is ancillary, as it unnecessarily restricts the application by explaining what AI is supposed to be used for, in a way that is itself objectionable or liable to fail to describe applications of interest. Finally, I object to the very incipit: it refers to "digital machine-based systems". Why should we constrain AI to be based on digital systems? I work in neuromorphic computing and I have seen and developed intelligent systems that employ analogic signal processing. So the word "digital" is already hurting. But AI might one day be developed on biological substrates, too, properly repurposed and engineered by humans - I would not cut bridges by referring to machines.
As you can see, there is lots of room for further debates on what AI is and isn't. While I called the situation disturbing above, I believe there is something positive about it, too - we are developing AI systems and technologies and we should not let definitions constrain our imagination! What is your favourite way to define artificial intelligence, by the way? You can use the comments thread to contribute!
