Why AI Cannot Innovate
There are two reasons why AI will never be able to innovate: (1) It reproduces what already exists, which means it makes nothing new, and (2) it cannot identify good options from among countless useless ones—an operation essential to innovation.
To begin this article by saying something obvious: We know how AI works. Which means we can figure out what it is and is not capable of. Analogously, we know that with the operation “subtraction” we can get from 5 to 2 but we cannot get from 2 to 5. You understand the operation, you can tell what can and cannot be done with it. We can follow the logic of AI’s fundamental principles and see that it is not capable of innovation, as I’ll show below.
First it’ll help to give a quick overview of how AI works. Take its text-producing capacity. When you talk to AI, e.g. you ask “What should I do if I’m cold?”, it scans a vast amount of text (e.g. all of the internet), finds sentences along the lines of “If you’re cold you should ___”, tallies up the number of times some or other words came next, and gives you some among the ones that occurred most often.
As it scans each sentence, what it’s doing is the same as asking a human in every instance “can this word follow that word?” e.g. “can ‘there’ follow ‘ball’ in ‘throw the ball’?” and the human responds “yes”. The human’s affirmation is inherent in the sentence. The fact that a person said “throw the ball there” as opposed to “of subtly grape as” is the human instructing the AI “yes, the word ‘there’ can follow the words ‘throw the ball’”. The more humans have put word A near word B, the higher word A’s selection probably when you enter word B into an AI system (e.g. as part of a question you ask it).
To phrase it differently—AI chooses what’s most popular. So how could AI with this method hit on innovations?
As AI is shaped by popularity among humans, one initial question we might ask is: Whose sentences is it being trained on? Is it a specific group—perhaps people thought to have particularly good taste or deep insights? Is it the entire population? Some popular things are found by certain people to be quite tasteless. There’s no good answer to these problems.
But their solution is not too important because even if you trained an AI system on the works of the greatest geniuses, there’s no evidence it would even then be able to innovate.
To argue it would you might have to show that there is some common feature to innovations and also that that feature can be represented computationally. No one has produced any such argument and, moreover, empirically speaking AI has not succeeded on these lines.
To illustrate a related failing of AI, consider what might happen if you asked it to produce something new in text (or any format), not based on popularity or past trends. All it could do then is choose words completely randomly, and it would have no way to pick out the good options from among the huge mass of bad ones.
The operation would be like the well-known analogy of a monkey tapping at a typewriter, who will, given enough time pressing letters randomly (trillions of years?), produce Macbeth by Shakespeare. But that’s alongside producing a vast amount of garbled nonsense, and how are you going to find the new genius piece of literature among trillions of gibberish manuscripts? No human could sift through those countless texts and there’s no way to program a computer to identify the valuable among the useless.
To create, you have to not only have a mechanism for producing valuable things, you must also have a way of ruling out what’s not valuable. If you create everything without discrimination, you create nothing.
To paraphrase philosopher Noam Chomsky putting the point differently: I can capture all of the rules of nature in one theory: “anything goes”. That theory covers anything that could possibly happen and therefore all natural laws. But it is useless because it accommodates everything—including every conceivable incorrect proposal.
One reason AI has no way to narrow down options is because computer programs are theories people have thought up, and no one has produced a theory of how such narrowing down could be achieved.
Humans on the other hand, unlike computers, are able to (though we don’t know how it happens) find some of the good options among the many bad ones, as evidenced by inventions and productions throughout history. Philosophers have pointed this out before. Charles Sanders Peirce, for example, talked about a mechanism in our mind that pushes it towards useful theories and enables it to recognise when it has come upon one, from among all the possible useless ones. How that mechanism works—we have no idea. But it is evidently there.
There’s no inconsistency in saying I don’t know how the human mind finds useful theories but I do know that computers cannot find them. I can also say I know that rocks cannot find them. Or that you cannot find them using the operation “subtraction”. To repeat, AI functions according to known principles (unlike the human mind) and so it is possible to discern things it’s not capable of.
Computers must either create everything without discrimination, or they must reproduce what has been most popular among humans in the past. Neither of these methods leads to innovation and computers have no third alternative.
While many people are claiming AI will surpass humans, no one has yet made a case for how its methods will make it capable of innovation.