The Limits of AI (Artificial Intelligence): "ChatGTP" vs. Chomsky?

Some reflections that arose while reading the following article that appeared in the “Theran Times of 3.28.2023): TEHRAN - An American linguist, Daniel Everett, has criticized Noam Chomsky's argument about the "innate principles of language," noting that ChatGPT demonstrated how a language can be learned without any grammatical principles. 

https://www.tehrantimes.com/news/483187/Exclusive-Linguist-says-ChatGPT-has-invalidated-Chomsky-s-innate

 Noam Chomsky definitely doesn't need my defense!

In fact, here I just want to clarify this dispute which is widespread and which emerges from a fundamental misunderstanding of one of the most important theories of the last century on language learning and on the very nature of human language.

Since its first appearance in 1957's Syntactic Structures, Chomsky's theory has undergone many changes but one point has remained the same despite all subsequent revisions of his theory: Simply put, the human (and to some extent animals I might add) must have an innate brain structure focused on language learning. This means that language learning differs from all other human learning as it cannot be explained only by the linguistic input that the child receives, therefore the child can produce more and different items than those received. This is intuitive if we consider that all languages are complex systems of rules with specific structures, even if some can be very general (universal grammar) however each language has its own specific system. In this case, knowing these rules it is possible to check whether e.g. a sentence is compatible with the system, another said has been "generated" (in the mathematical sense) by the rules, or it is not, that is, it is incorrect.

 If we take this fundamental point of Chomsky's theory into consideration, one might mistakenly assume that a machine can do the same. But here's the thing: humans don't just learn a language: they recreate it. In learning to speak, children of each generation recreate a language that is somewhat different from that passed on to them by the previous generation. This is a well known and documented fact, language change occurs in phonetics, syntax, lexicon and even intonation. This is understandable since language is not like mathematics a fixed system but it is a communication tool and has to be continuously adapted to the changing situation of human life throughout history.

 We could even think of all languages as some sort of "very beneficial virus" that replicates using humans as tools. This in fact explains the continuous language mutations.

 No machine could do such a thing: computers can only process input but they are sterile and could never output anything new. A combination of linguistic material received is only apparently something new: in fact, one can reconstruct the entire process performed by the computer to achieve the result and discover that it is simply a program written by a human being.

 I am therefore strongly against this thesis. The truth is exactly the opposite: ChatGTP confirms that learning a language is based on a specific structure of the human brain (and perhaps to some extent this also applies to animal languages).

 ChatGTP is therefore definitive proof that a machine cannot learn a language: such computer programs can only process huge amounts of linguistic material and produce texts that are apparently new but are actually just a combination of the inputs received. Thus Chomsky's theory of the need for an innate structure in humans for language learning is confirmed. Obviously, however, although necessary, this is not enough.

 There is further proof of this fact: the first machine translation programs basically used something very similar to Chomsky's theory: they wrote programs which contained the rules of the given languages, added the lexicon and tried to make the system work: they did not there was no way to get a valid result. Then someone involved in machine translation realized that the best method was to feed computers with a huge amount of texts with corresponding translation: something that modern computers can easily compare in the shortest time possible and then select what seems statistically more appropriate as a translation.

 We interpreters do practically the same thing and in fact we learn a lot from experience: the more we translate the better we become. And we can only be sure if we focus on the context of the communication, which no machine can do: so google or ChatGTP are a  valuable help, but unreliable if used without human check.

What appears to be flawed or deficient in Chomsky's theory is mainly due to its scholastic acceptance by non-linguists. Here the prof. Everett is absolutely right: language is more than "grammar" (syntax, morphology). But Chomsky's theory in my understanding was never intended to reduce language to grammar: When Chomsky wrote about an innate LAD, the language acquisition device, he was simply emphasizing the importance of a specific brain structure focused on language learning .

But this is only as ONE of the conditions.

Not without irony, the psychologist Jerome Bruner (I recall in one of his conferences at the University of Constance) had proposed extending the concept by adding a LASS (Language Acquisition Sustaining System), i.e. remembering that something else is required beyond learning grammar. And for mnemonic purposes he had added the theory to be completed as follows: "A Lad needs a Lass" (a boy must have a girlfriend) since "lad" is precisely synonymous with boy and "lass" with girl. Perhaps today he would no longer have said such a thing, but that's another story.