With their oft-cited paper on Generative AI (GenAI) as a stochastic parrot, Bender et al. (2021) drew attention to the implications of using anthropomorphic language to describe Large Language Models (LLMs). Vallis et al. (2025) have demonstrated how we—almost instinctively—draw parallels with our own cognitive processes and social relationships to understand our interaction with GenAI. In his 2006 article on the ontological status of the cyborg, H.G. Geertsema noted the error inherent in the analogy of the machine as human organism: “The computer as such does not function in terms of conscious awareness. It has no emotional experience and no sense of responsibility and accountability” (2021, 324). But conscious awareness, emotional experience, sense of responsibility and accountability are precisely what popular culture imagines with its benign humanization of AI machines, such as R2D2and C-3PO. In this respect, Kazuo Ishiguro’s 2021 novel, Klara and the Sun, is a more complex exploration of the affinities of the machine with the human in terms of reason, affections, and will. Despite general anxiety at the prospect of AI turning against its Frankenstein-like creator, when Disney-Pixar’s WALL•E morphs into The Terminator’s Skynet, we continue implicitly to confer autonomy and consciousness upon GenAI with the metaphors that we draw from the human domain to describe the programmatic execution of its tasks.
In this essay, I propose an approach to LLMs that rests in a philosophy of language radically different from the one widely used when thinking about GenAI at present. I will first demonstrate that the theory of language currently underwriting the ascription of language to GenAI is categorically wrong. Then I will consider the affordances, as well as limitations, made possible by the lingual aspect in Herman Dooyeweerd’s theory of modal spheres for evaluating the “language” of LLMs. In the final section, I will turn to the Reformational emphasis on answerability as the quality that distinguishes human language from the parrot-like speech of GenAI. Focussing on language will help clarify and dispel the concerns we may have about the consciousness, sentience, and agency of virtual technologies. My argument is not against the ethical application of AI for social good in computer coding, medical diagnostics, actuarial science, meteorology, financial services, engineering, public safety, and so on. My argument is that we have misunderstood the nature of GenAI’s “language” and, consequently, have given it a misleading modal function and a misguided normative aspect.
GenAI Language: A Category Mistake
In a New York Times opinion piece on the deleterious effect of GenAI on college admission essays, Rebecca Winthrop writes: “For the first time in human history, we have a technology that can generate words separately from the thoughts they represent” (2026). The verb “generate” to denote how LLMs select and arrange words to mimic the human performance of writing seems justified here, given that the phrase “Generative AI” has a technical usage in the field. More interesting than “generate” in this sweeping statement about human history is Winthrop’s assumption that words are the realization of ideas. An idea forms in the mind as a thought, the brain associates the idea with words, and Broca’s Area is activated to voice the words materially in speech. The hearer reverses the process, engaging Wernicke’s Area to translate the word as physically received back into the corresponding idea that originated as a thought by the speaker. Thoughts are abstractions formed in the human mind for which words are the signs that correspond to the ideas they signify. Without thinking, the mechanical sorting system of an LLM searches its vast databases for the words that correspond with the ideas that most effectively meet the parameters set by the user in his or her prompt. GenAI may be remarkably capable at generating words at breath-taking speeds, but it is incapable of generating the original thought. Whereas human beings generate the thought before the word, an LLM is an auto-regressive mathematical next-token generator. Ergo, the classical theory of language as the representation of thought cannot apply to LLMs.
Two further considerations are relevant when applying the thought-to-word model to GenAI. First, by identifying language with thought, one may inadvertently attribute consciousness and interiority to the digital process of the LLM. Those who, following Pope Leo XIV’s emphasis on the distinctiveness of human dignity in “Magnifica Humanitas” (Encyclical Letter, May 15 2026), insist on withholding from GenAI the privilege of consciousness and interiority may practically have already anthropomorphized it by virtue of a thought-to-word theory of language: language is the representation of thought; LLMs use language; therefore, GenAI has consciousness. But the contraposition of the syllogism using the classical theory of language can also obtain here: words are the realization of thought; a system that lacks consciousness cannot think; therefore, the words that GenAI produces is not language.
Second, when the representational theory of language moves from the naming of thoughts to the naming of things, one encounters a similar problem. As Charles Taylor has argued in The Language Animal: The Full Shape of the Human Linguistic Capacity (2016), “human meaning” is “beyond information encoding” because the thing to which languages gesture is itself “the locus of the full corona of liminal meanings” (95). Taylor has in mind the extension of objects in the purview of the subject in a complex, dynamic, and associative experience of its material reality. One does not require a developed phenomenology of objects–such Charles Sanders Peirce’s theory of the dynamic object—to recognize that the thing itself precedes its naming. GenAI has no experience of things; there is literally no thing (res) beyond or outside the lexicon of words (verba) that comprise the LLM upon which GenAI operates. Derrida’s “Il n’y a pas de hors-texte” is genuinely true of GenAI. The classical correspondence of thing and word (res et verba), similar to the collation of thought and word, is categorically not applicable to GenAI. At best, it is an account of language that is inadequate; at worst, it grants GenAI anthropic characteristics that it ontologically cannot possess.
Dooyeweerd’s Lingual Aspect
Herman Dooyeweerd’s cosmonomic philosophy offers another means of demonstrating that anthropomorphizing GenAI’s “language” is problematic. Rather than consciousness being the a priori of being human in the thought-to-word model of language, Dooyeweerd begins with religion as “the absolutely central sphere of human existence” (1953, 2.57). The existential question about AI is not whether it has consciousness but whether it can believe. Not only is GenAI incapable of emotion, responsibility, and accountability (see Geertsema 2021, 324), it is incapable of the religious operations of trust, faith, and obedience. Dooyeweerd’s description of the modal spheres, including the lingual aspect, always already assumes the “radical religious unity” of the person (2.55). Not only are the modal aspects contiguous and interdependent, they are all present in the person in Dooyeweerd’s holistic anthropology. The illusion of humanness that GenAI is able to evoke is the function solely of its use of language, whereas the understanding, empathy, and trustworthiness communicated in human language grow out of the religious root of the whole person. Human speech is not reducible to an idea-to-word equation through an auto-regressive mathematical next-token process; as with other modal aspects, language is the articulation of the religious core of the human person. “For out of the abundance of the heart his mouth speaks” (Luke 6:45 ESV).
When it comes specifically to the lingual aspect, Dooyeweerd’s general theory of the modal spheres can also be helpful in parsing GenAI’s relationship with language. First, the lingual aspect describes any form of “symbolic signification”, not exclusively verbal language (2.137). The lingual follows the analytical and formative aspects in the ordering of the modal spheres, two aspects that are constitutive of computational technologies in which logic and modeling are primary functions. In terms of its retrocipatory quality, then, the lingual aspect of GenAI can be described as the encoding into words of the analysis it executes upon in the databases that comprise it. ChatGPT, Claude, and other word-generating programs match vast arrays of digital symbols with corresponding verbal symbols. Second, by leaning more into the symbolic than the social aspect of language, Dooyeweerd’s model illustrates the machine-like nature of GenAI in a way that further disperses the chimera of its humanity. When seen as an extension of the analytical mode in digital technology, the lingual aspect of GenAI is less likely to be identified as social in the way that human discourse uses language. Third, in light of the anticipatory direction of the lingual aspect, the difference between GenAI language and human language becomes clear:
In the lingual aspect a moral anticipation is implied in the love of a language, a tendency to signify our feelings, volition, thoughts etc. in the linguistically most adequate way inspired by the affection for language in its pure form. (2.152)
The lingual aspect further anticipates the expression of the profoundly human experience of harmony, justice, ethics, and faith in their respective aesthetic (poetry), juridical (forensic rhetoric), ethical (contract), and pistic (prayer and praise) aspects.
While reality is inter-modal, multi-aspectual, and irreducible, Dooyeweerd’s discussion of the aspects implies that each is governed by a particular set of norms. Unfortunately, Dooyeweerd does not describe directly the law-side of the lingual aspect of language (as opposed to other symbolic signification), dealing with it only in relation to the higher-order spheres which it anticipates. Thus, in the juridical sphere, the norm for the lingual aspect is “exactness” and “univocal signification” (2.139-140). In the aesthetic sphere, the lingual aspect is directed by the norm of aesthetics, namely, harmony. The word that recurs most frequently in Dooyeweerd’s references to the lingual aspect is economy. “Economy of speech”, or “linguistic economy”, is the mark of “a scientific and developed stage of thought and language” (2.66):
there is a tendency to an ever increasing degree of “economization” in the aspect of symbolic signification at the higher stages of historical development. This becomes evident if we compare modern and primitive verbal languages. (2.126)
If frugality is the norm of the economic aspect, then succinctness in language is the ideal of the lingual aspect. Analogous to Occam’s razor in logic, elegant proof in mathematics, and parsimony in statistics, simple efficiency is the litmus test in the lingual sphere, strictly speaking.
Language as the Performance of Relationality
While Dooyeweerd’s general theory of modal spheres is useful in mitigating the anthropomorphism of GenAI language, particularly in its emphasis on the religious root of the person, it does not finally distinguish the nature of language in human discourse from that produced by GenAI. It is only in its anticipatory qualities in the social, aesthetic, juridical, ethical, and pistic spheres that language acquires its moral dimension for Dooyeweerd. Ontologically, language as a form of “symbolic signification” is not qualitatively different from non-linguistic sign systems, such as computer code. Replacing the predominantly structuralist and semiotic approach of Dooyeweerd’s lingual aspect, where language is “symbolic signification”, with a guarded phenomenological approach to language as the performance of relationality identifies the gap between GenAI and human discourse. Following Vanhoozer (2001), I have argued that covenant theology is the foundation of a Reformed theory of language as the enactment of relationship (Faber 2020, 2024, 2026). Extending the special covenant events in biblical history to our mundane speech acts, language is the performance in space and time of promise and obligation, call and response, invitation and answerability. The nuclear meaning of language, then, is not “symbolic signification” but the performance of embodied relationality, the principal norm for which is not economy or efficiency or accuracy. These are valid subsidiary, technical norms governing the means but not the end of language. Instead, language aims at mutual regard, faithfulness, commitment, and honesty—in a word, love: “the central command of love” is “the radix of all modal aspects which unfolds the divine law in temporal reality” (Dooyeweerd, Prolegomena, 1.61). GenAI’s production of words does not meet the test of language thus defined.
As long as we continue to think of language as the representation of thought rather than covenantal speech acts, we will continue tacitly to bestow consciousness and sentience upon GenAI. Without a body, GenAI is incapable of relating with the user’s experience of joy, pain, elation, shame, hope, doubt, confidence, anxiety, peace, and guilt. GenAI’s manipulation of the relational affect of language is purely an illusion. An LLM is perfect Derridean écriture and différance in its semiotic sampling of digital databases for signification without transcendence; an identification of the auto-regressive mathematical next-token generator of words with real human language, paradoxically, is an implicit acceptance of a poststructuralist construal of language as a game of shifting signifiers. The temptation is to replace relationship with efficiency as the final cause of language when we “chat” with OpenAI’s ChatGPT, ask Microsoft’s Copilot to “create content”, and appoint Anthropic’s Claude as our “agent”. Language is not a tool, with various degrees of efficiency, but of the substance of our religious identity as human beings in dialogue with our God, our neighbor, and our world.