Noam Chomsky, Ian Roberts, and Jeffrey Watumull, in “ The False Promise of ChatGPT ,” (New York Times, March 8, 2023), lament the sudden popularity of large language models (LLMs) like OpenAI’s ChapGPT, Google’s Bard, and Microsoft’s Sydney. What they do not consider is what these AIs may be able to teach us about humanity.
Chomsky, et al., state, “we know from the science of linguistics and the philosophy of knowledge that they differ profoundly from how humans reason and use language.” Do we know that? They seem much more confident about the state of the “science of linguistics and the philosophy of knowledge” than I am. One of the principles of science is that when an experiment yields a surprising result, we should be reluctant to dismiss the experiment and stubbornly cling to our preconceptions. I have yet to encounter any scientist, even experts in machine learning, who are not surprised by the astonishing linguistic capabilities of these LLMs. Could they teach us something about how humans reason and use language?
The authors continue, “These differences place significant limitations on what these programs can do, encoding them with ineradicable defects.” But the defects they cite, to me, strikingly resemble defects in humans. We make stuff up. We parrot lies. We take morally inconsistent positions or weasel our way out of taking a position at all.
The authors assert that, “the human mind is a surprisingly efficient and even elegant system that operates with small amounts of information.” I have studied (and taught) information theory, and any measure I can imagine for the information provided to human brain during its 20 or so years of development into an educated, rational being is not small. They speak of the “miniscule data” and “minimal exposure to information” that lead a child to be able to distinguish between a grammatically well-formed sentence and one that is not. They then cite the “consciously and laboriously … explicit version of the grammar” constructed by (adult, highly-educated) linguists as evidence that “child’s operating system is completely different from that of a machine learning program.” To me, it could be evidence to the contrary. The child learns from examples, like the large language models, albeit with fewer examples. The child is not able to synthesize the explanations that the adult linguists have laboriously constructed. Interestingly, the LLMs can synthesize these explanations, but only because they have “read” all the works of these adult linguists. Leave those texts out of the training data, and their sentences would be no less grammatical, but they would lose the ability to explain the grammar.
The authors speculate, “[b]ecause these programs cannot explain the rules of English syntax, for example, they may well predict, incorrectly, that `John is too stubborn to talk to’ means that John is so stubborn that he will not talk to someone or other (rather than that he is too stubborn to be reasoned with).” I asked ChatGPT what their sentence means (on March 9, 2023), and it responded:
The sentence "John is too stubborn to talk to" means that John is so unyielding and inflexible in his opinions, ideas, or behavior that it is difficult or impossible to have a productive conversation with him. His stubbornness makes it challenging to reason with him or persuade him to consider alternative viewpoints or options. The use of the word "too" implies that his stubbornness is excessive and beyond what is reasonable or tolerable in a conversation or interaction.
Another principle in science is that, to test a theory, one should make a prediction (as Chomsky et al. did), and then perform an experiment to see whether the outcome is consistent with the prediction. It seems that they did not perform the experiment, but rather clung to their prior paradigm. This, unfortunately, is the all-too-human way that science is often conducted, as exposed in the 1960s by Thomas Kuhn.
The authors observe that the programmers of AIs have struggled to ensure that they steer clear of morally objectionable content to be acceptable to most of their users. What they fail to observe is that humans also struggle to learn to apply appropriate filters to their own thoughts and feelings in order to be acceptable in society, to avoid being “cancelled.” Perhaps the LLMs can teach us something about how morally objectionable thoughts form in humans and how cultural pressures teach us to suppress them.
In a reference to Jorge Luis Borges, the authors conclude, “[g]iven the amorality, faux science and linguistic incompetence of these systems, we can only laugh or cry at their popularity.” When Borges talks about experiencing both tragedy and comedy, he reflects on the complex superposition of human foibles and rationality. Rather than reject these machines, and rather than replacing ourselves with them, we should reflect on what they can teach us about ourselves. They are, after all, images of humanity as reflected through the internet.