Vincent Carchidi

Discussions about the encroachment of Large Language Models (LLMs) on the domain of human creativity often conflate productivity with creativity; they conflate the ability to generate new structures according to rules or other systems (productivity) and the ability to use one’s productive capacity without fixed reliance on identifiable stimuli while nevertheless producing structures appropriate to the situation (creativity). For a behavior to be “creative” is to use the productivity of a capacity like language in ways that are stimulus-free, frequently novel, yet appropriate to situations. This paper argues that, because of this conflation, improvements in the productive capacities of LLMs have been mistaken for improvements in this richer notion of creativity. Although LLMs – underpinning both chatbots and agentic systems – exhibit remarkable productive capacities, their relation to their environment does not exhibit this “creative” aspect. Where the appropriateness of human linguistic behavior does not appear governed by a causal relationship between the individual and their environment, LLMs exhibit a functional appropriateness that sees their productive capacities tethered to identifiable stimuli in the local environment. They are, in the Cartesian sense, impelled to act, but not inclined, contrasting with the human who is ‘incited and inclined but not impelled.’

Ordinary human language use appears to be uncaused, often novel, yet appropriate to the situations in which it is used. This is operationalized by generative linguists like Chomsky as stimulus-freedom, unboundedness, and appropriateness. Taken together, these characteristics locate human language use outside the bounds of determinacy and randomness. This “creative aspect of language use,” or CALU, has downstream implications for explanatory theory that are deeply embedded in the generative linguistic enterprise. These include the invocation of a competence-performance distinction that segments linguistic behavior off from abstract knowledge of language, the internalist orientation of linguistic inquiry, and the postulation of a modular and domain-specific generative language faculty. CALU and its implications have been sorely neglected by recent enthusiasm around the use transformer-based Large Language Models (LLMs) as models or theories of human language, failing to account for the depth of the problem it poses for a science of human language and failing to impose the requisite theoretical constraints. Such approaches are fundamentally deficient and limit the role of LLMs in linguistic inquiry.

[IN PRESS] Matthias Mahlmann’s Mind and Rights (M&R) argues that the mentalist theory of moral cognition - premised on an approach to the mind most closely associated with generative linguistics - is the appropriate lens through which to understand moral judgment’s roots in the mind. Specifically, he argues that individuals possess an inborn moral faculty responsible for the principled generation of moral intuitions. These moral intuitions, once sufficiently abstracted, generalized, and universalized by individuals, gave rise to the idea of human rights embodied in such conventions as the Universal Declaration of Human Rights. I am sympathetic to this argument. I thus take my comments on M&R as an opportunity to preempt topical challenges to the mentalist theory arising in machine learning. Through this challenge, I make clear the foundations of mentalism (generativism). On these foundations, I argue against the use of computational models as theories of human cognition and for limiting their scope as proxies for theories. Counter-arguments are developed, centering on the notion that computational models, like Large Language Models (LLMs), circumvent the leap from non-linguistic to linguistic during (pre-) training. Additionally, LLMs’ learning objectives are constrained by closed-sets of language-specific data over which they operate, circumventing a test of domain-specific vs. domain-general cognitive mechanisms for language acquisition. Focuses of this argument include the format of the input, the constraints on learning, and the means of data collection. These arguments are tied into M&R’s core themes, specifically concerning the necessity of rich descriptive accounts and the theory-dependence of data-interpretation.

Descartes famously constructed a language test to determine the existence of other minds. The test made critical observations about how humans use language that purportedly distinguishes them from animals and machines. These observations were carried into the generative (and later biolinguistic) enterprise under what Chomsky in his Cartesian Linguistics, terms the “creative aspect of language use” (CALU). CALU refers to the stimulus-free, unbounded, yet appropriate use of language—a tripartite depiction whose function in biolinguistics is to highlight a species-specific form of intellectual freedom. This paper argues that CALU provides a set of facts that have significant downstream effects on explanatory theory-construction. These include the internalist orientation of linguistics, the invocation of a competence-performance distinction, and the postulation of a generative language faculty that makes possible—but does not explain—CALU. It contrasts the biolinguistic approach to CALU with the recent wave of enthusiasm for the use of Transformer-based Large Language Models (LLMs) as tools, models, or theories of human language, arguing that such uses neglect these fundamental insights to their detriment. It argues that, in the absence of replication, identification, or accounting of CALU, LLMs do not match the explanatory depth of the biolinguistic framework, thereby limiting their theoretical usefulness.

The victories of the Go-playing artificial intelligence “AlphaGo” against professional player Lee Sedol in 2016 had a profound impact on public and academic perceptions of AI. This event shocked observers, as the ability of a machine to defeat a world champion human in a highly complex game seemed to indicate that a machine had achieved human-like—or more than human—intelligence. But why was AlphaGo so readily anthropomorphized by academic and non-academic audiences alike? Drawing from existing analyses of reactions to and arguments concerning AlphaGo and AI generally, this paper argues that “generative” cognitive science—a school of thought exemplified by the linguistic work of Noam Chomsky—offers two novel contributions to this subject. First, generativism sheds light on an irrational double standard in the study of the human mind in contrast to the study of non-cognitive systems—“methodological dualism”—which, I argue, has been transferred to evaluations of AlphaGo and other AI. Second, by exposing this irrational double standard in perceptions of AI, I employ generativism’s more well-known arguments concerning the nature of human intelligence and its scientific study to the evaluation of AI, exposing deficient interpretations widely used in the case of AlphaGo and AI generally.

I argue that theoretical developments in the study of the Universal Declaration of Human Rights (UDHR) should occur alongside progress in moral psychology, particularly moral cognition. More specifically, I argue that Universal Moral Grammar (UMG), a model positing an innate, regulative, and universal moral faculty characterizable in terms of rules and principles, fulfills the role of the foundational model needed to usefully conceptualize the UDHR. As such, I provide a detailed account of UMG against competing models in moral psychology. Furthermore, I combine UMG with Talbott’s Historical-Social Process of Moral Discovery and Rawls’s reflective equilibrium to show how the UDHR represents a major development in moral exploration, one indicating a more penetrative look into the inner moral nature of humans that attempts to reach, but does not attain, one interpretation of reflective equilibrium tied to cognitive moral psychology.