By Mohammad Mazhari

Exclusive: Linguist says ChatGPT has invalidated Chomsky's ‘innate principles of language’

March 28, 2023 - 3:8

TEHRAN - An American linguist, Daniel Everett, has criticized Noam Chomsky's argument about "innate principles of language," emphasizing that ChatGPT has demonstrated how a language can be learned without any hard-wired principles of grammar.

Artificial intelligence (AI) has been a subject of fascination and concern for many years. In previous interviews with renowned linguists Noam Chomsky and Ray Jackendoff, Tehran Times explored their views on the relationship between language and the mind, and how it relates to the study of AI. Today, Tehran Times continues this conversation with Daniel Everett, a prominent linguist known for his work on the Piraha language and the implications it has for our understanding of language and cognition. We will dive into his insights on the relationship between language, culture, and cognition, and how they could inform the development of AI technology.

In the interview, Everett challenges the nativist theories of Noam Chomsky, arguing that Chomsky only focuses on grammar, which is only a small part of language and does not consider social or cultural origins. Everett discusses his findings on Piraha, one of nearly 8,000 spoken languages in the world, arguing that it is no different from any other language in terms of language acquisition.

 Everett states that current theories of language learning do not consider semiotics and inferential reasoning and need revision. The interview also touches on the relationship between language and thought, where Everett argues that language and culture exist in a symbiotic relationship, with each one necessarily shaping and affecting the other.

According to Everett, ChatGPT has shown that language acquisition is possible without any innate grammar or language rules. This is achieved through the use of massive amounts of data, as demonstrated by Large Language Models. He argues that as ChatGPT continues to evolve and more advanced models are developed, it will mature from its current "baby" stage to adolescence and eventually adulthood.

Following is the text of interview with Daniel Everett:

Q: Let's start the discussion with the origins of language. While some anthropologists claim that language is a product of cultural evolution and interaction between humans and their environment, scholars like Noam Chomsky argue that language is an innate capability of the human brain. What is your view on this debate, and how does it relate to your research on the Piraha people and their language? Given that Piraha is just one case study, can we claim that your findings about this case are enough to challenge Chomsky's arguments, whereas most languages are similar in terms of structure?

A:   The first thing to do in discussing Chomsky's ideas about "language" is to realize that when he uses this word, he is not referring to what most people think of as language. He means in fact "grammar" and only those parts of grammar that have no social or cultural origins (which, a priori, one cannot know).  It is misleading, therefore, to say that Chomsky believes this or that about language because his work does not consider language only grammar, a small and (in my new book on the philosophy of linguistics) fairly insignificant part of language. There are no studies of discourse, to take one example, in a Chomskyan theory. And stories are likely the most important component of any theory of how language works. Chomsky only looks at sentences, phrases, and words.

There are two justifications that have been used for applying the concept "innate" to Chomskyan grammars: (i) everyone speaks a language regardless of their mental abilities or social context; (ii) those grammars are remarkably similar around the world; (iii) children all learn their language before puberty, much earlier for most.

Chomsky indirectly predicts in fact that not all children should be able to learn all languages. This is because if language is carried on the genes, then we expect natural selection to eventually adapt the innate genetic language to the local environment and grammar and thus to make it easier to learn the local language and harder to learn languages with certain properties that diverge from the local language. Since there have been biological adaptations to the local environment in short time spans (e.g., lactase persistence, which is only about 6,000 years old, enabling adults to drink milk without adverse reaction in societies which have dairy cultures of about this time depth), one would expect language to show similar local adaptations. The property of omitting a subject (what Chomsky calls "pro-drop") is about 6,000 years old in Indo-European. If language were on the genes, one would expect speakers of pro-drop languages, e.g., Spanish and Portuguese, among other Romance languages, to struggle with learning English, say. But this is not correct. With regard to point (ii) the grammar will look similar depending on your theory. With regard to (iii) children learn lots of things very rapidly due to the general nature of the brain when they are born. Over time it becomes more difficult for all of us to learn new things, relative to our ease before puberty.

Q:   How do you think the Piraha language challenges or supports other theories of language acquisition, such as social interactionist or usage-based approaches? In what ways might your findings contribute to ongoing debates about the role of culture, socialization, and individual cognitive processes in language learning and use?

A: The Piraha language is one of nearly 8,000 spoken in the world. In recent work, Geoffrey K. Pullum's comparative research on recursion across languages has shown several other languages that, like Piraha, seem to lack sentential recursion. In my new work on the philosophy and practice of linguistics (just submitted to Oxford University Press), I argue that nativism is not needed. There is no cross-theoretically convincing evidence that children use any more than understanding signs (icons, indexes, symbols, and the like) plus inference (deduction, induction, and abduction) to learn their languages. Piraha in this way of looking at language is no different from any other language. While Piraha represents a serious problem for nativist theories of language acquisition, it is no problem for other theories of the type you mention. However, all current theories fail to consider any role for semiotics and inferential reasoning in the task of language-learning (first or second language learning) and so I consider them all in need of revision.

Q:    Some philosophers and linguists have argued that language and thought are intimately related, and that language plays a key role in shaping cognitive processes and mental representation. How does your research on the Piraha language support or challenge this view? What kinds of evidence do you think we need to collect in order to better understand the relationship between language and thought, and to disentangle the complex interplay between biological, social, and cultural factors?

A:  All humans think with language. Charles Sanders Peirce pointed this out in the 19th century and William James discussed examples that showed language affecting thought around the same thing. Culture and language exist in a symbiotic relationship. Each one necessarily affects the other. My son, Caleb Everett, has published a great deal on how language affects the way we think and in my work I have written on culture affecting language. I do not believe that speaking a different language means that you are incapable of learning to speak another nor that it prevents the ability to think about new things. But the languages we speak do seem to affect our initial performance in different experiments or environments, as has been shown in many studies over many years.

Q:  How do you see the future of languages in light of recent advancements in artificial intelligence (AI) and machine learning (ML)? What will be the impact of AI and ML on the diversity of languages and cultures, and how might this affect our understanding of linguistic diversity as a valuable resource for scientific and societal purposes? As we move towards more automation and digitization, do you predict that AI will eventually impose a universal language and culture on the world, or do you see potential for new forms of linguistic and cultural creativity to emerge in response to these developments?

A:  Computational research has no logical connection to the diversity of languages and cultures. But I believe that more and more researchers will begin to recognize that culture needs to be considered by AI modeling just as language does.  I have written extensively on culture's importance for the cognitive sciences, especially my U of Chicago book on "Dark Matter of the Mind."

Q:  Linguists like Chomsky have argued that machines like ChatGPT are far from achieving true artificial intelligence. What is your perspective on the relationship between AI and human intelligence, and how do you think AI can be used to advance our understanding of human cognition and language acquisition? Do you believe that there are specific areas of language acquisition or language processing that are particularly difficult for machines to replicate, and if so, what are they?

A:   So here is what ChatGPT has done: it has falsified in the starkest terms Chomsky's claim that innate principles of language are necessary to learn a language. ChatGPT has shown that without any hard-wired principles of grammar or language this program, coupled with massive data (Large Language Models), can learn a language. Chomsky has recently discussed examples that ChatGPT does not yet understand. What is interesting about that very small number of examples is that they are also misunderstood by many children and second language learners (e.g., Bill was too stubborn to give the book to). ChatGPT is a kind of baby, just beginning. As improved models come out it will reach adolescence and then adulthood (and beyond). This paper is to me one of the most important papers to emerge in decades and is on this exact topic: (this paper was written for a festschrift in my honor to be held at MIT's Brain and Cognitive Sciences Department in June). I believe that LLMs and programs like ChatGPT will be able to master any area of language.

Q:  Historians like Yuval Harari have warned in the New York Times that "If We Don’t Master AI, It Will Master Us." What is your perspective on this warning, and what steps do you think we need to take to ensure that AI is developed and used in a responsible and ethical manner? How can we balance the potential benefits of AI and ML with the potential risks and challenges, and what role can linguists and other researchers play in this effort?

A: I do not fear AI any more than I fear my refrigerator. AI's only potential problem is how we use it as a society. I distrust human intelligence more than artificial intelligence (and it is very artificial).

Leave a Comment