Those of us inclined to lob accusations of hubris at new forms of artificial intelligence, on the grounds that we observed (and possibly fomented) the same kind of affliction in the classical forms, should at least clarify the analogy. What are the similarities that advise caution, and what are the differences that might alleviate it?
We address primarily the large language model, and draw attention to its use of text. Consider a traditional linguistics class of some years ago. The first thing students learned was that written language is of lesser interest; the action is in spoken language.1 And “action” means humans’ use of language, its acquisition and development, its sharing and variation—and the implications for the study of human thinking. Written language, which involves planning and editing, is derivative, not spontaneous, not universal, and relies on arbitrary representation learned explicitly. So what, then, does this mean for generative AI?
Each wave of artificial intelligence research has claimed that intelligence via computation is within reach, with a connection between human intelligence, and computed, or artificial, intelligence. The question is how to get from the human form to the artificial form.
The early AI wave focused on Facts and Reasoning as the manifestation of intelligence. Efforts such as Dendral and MYCIN were successful in their limited domains. Computer intelligence was within our grasp—that was the attitude. With Facts and Reasoning implemented as the obvious, Database and Logic, the last step, from there to computer intelligence, would be a piece of cake, meaning “easy” or “suitable for graduate students,” just a matter of scaling up, filling out the details with the computing mechanisms at hand.
But the approach, expert systems, was found to be programs of unfortunate sterility. The piece-of-cake work had to augment it with probabilities, common-sense extensions, planning scripts, proceduralized grammars, and other heuristics,4 yielding systems that offer, yes, significant aspects of intelligence.
We can see that pushing the hard part farther down in the toolchain delayed the revelation of assumptions made along the way. Reliance on Facts and Reasoning to yield intelligence exposed the limitations of that approach (sterility), calling for various ad hoc mechanisms.
Now we are persuaded that intelligence is buried in our language, and can be released by pattern-matching across utterances in extremely voluminous and fine-grained contexts, with amazing results, and even in contexts of the contexts, with even more amazing results. And we have lots of Language—we have huge text corpora! With Language implemented in the obvious way, writing, the last step to computer intelligence is a piece of cake, just a matter of scaling up, filling out the details with the computing mechanisms at hand.
The approach, massive statistical processing via machine learning with transformers, is good with prediction but weak on logic and subject to startling departures from truth. The piece-of-cake work requires tuning, re-modeling, and human feedback, yielding another set of systems that offer, yes, significant aspects of intelligence.
This sounds familiar… from just a few paragraphs back. Analysis of weaknesses in the classical expert-system approach revealed hidden assumptions. Can we drag some out here, too, for generative AI? Yes—We assume that language can be captured in digital data from text, and that language scraped exclusively from writing holds intelligence. Although the reliance on written language seems entirely appropriate for a service intended to produce written text, the processing there requires, again, ad hoc human intervention.
But does the analogy endure? Facts and Reasoning are not the same as Language, which may still hold depths of cognitive material. What about gathering inputs from spoken data? Clever minds in High Tech already are thinking about it,3 and that could obviate one assumption, perhaps replacing it with another. Investigation into the relevant differences in AGI data potential between Facts and Reasoning, and Language; or between written and spoken Language, would be worthwhile.
As an example of assumptions, I question trust in the intellectual completeness of writing (textual input). We should also question trust in other mechanisms, such as tokenization, generation as prediction, reliance on human assessment, and constraints of the digital milieu itself. (What would those constraints be?—Who knows?) A real risk is the human tendency to reverse the assumption, to think that rolling back, from the current state of the art, through the toolchain, will tell us what intelligence IS. Language is important, as McShane and Nirenburg state: “[E]nabling machines to emulate human-level language proficiency is well understood to be an AI-complete problem, one whose full solution requires solving the problem of artificial intelligence in general.”2 But note that, for them, language proficiency would be a manifestion of AGI, rather than the other way around.
An argument from analogy is not deductive, of course. Just as well! These observations are about attitudes and discourse in science, not the science itself. All the commentary that has been devoted to warning modern AI enterprises about past hubris is not proof that current efforts will fail. The point is to raise questions about presumptions.
References
1. Cruz-Ferreira, M., and Abraham, S.A. 2011. The Language of Language. A Linguistics Course for Starters. Section 2.6. OERCommons.org.
2. McShane, M., and Nirenburg, S. 2021. Linguistics for the Age of AI. The MIT Press. Creative Commons license CC-BY-NC-ND.
3. Mehta, S., Jojic, N., and Gamper, H. 2025. Make Some Noise: Towards LLM audio reasoning and generation using sound tokens. 2025 International Conference on Acoustics, Speech, and Signal Processing. April 2025.
4. Schubert, L. “Computational Linguistics.” The Stanford Encyclopedia of Philosophy. Spring 2020 Edition. Edward N. Zalta (Ed.).
Robin K. Hill is a lecturer in the Department of Computer Science and an affiliate of both the Department of Philosophy and Religious Studies and the Wyoming Institute for Humanities Research at the University of Wyoming. She has been a member of ACM since 1978.