Does Artificial Intelligence Speak Our Language?: A Gadamerian Assessment of Generative Language Models

• Published date: 01 September 2024
• DOI: http://doi.org/10.1177/10659129241243038
• Author: Phillip Pinell
• Date: 01 September 2024
• Subject Matter: Articles

Article

Political Research Quarterly

2024, Vol. 77(3) 713–728

© The Author(s) 2024

Article reuse guidelines:

sagepub.com/journals-permissions

DOI: 10.1177/10659129241243038

journals.sagepub.com/home/prq

Does Artif‌icial Intelligence Speak Our

Language?: A Gadamerian Assessment of

Generative Language Models

Phillip Pinell

1



Abstract

The language argument is a classic argument for human distinctiveness that, for millenia, has been used to distinguish

humans from non-human animals. Generative language models (GLMs) pose a challenge to traditional language-based

models of human distinctiveness precisely because they can communicate and respond in a manner resembling hu-

manity’s linguistic capabilities. This article asks: have GLMs acquired natural language? Employing Gadamer’s theory of

language, I argue that they have not. While GLMs can reliably generate linguistic content that can be interpreted as

“texts,”they lack the linguistically mediated reality that language provides. Missing from these models are four key

features of a linguistic construction of reality: groundedness to the world, understandi ng, community, and tradition. I

conclude with skepticism that GLMs can ever achieve natural language because they lack the se characteristics in their

linguistic development.

Keywords

generative language models, Gadamer, artif‌icial intelligence, language, democratic theory

Artif‌icial Intelligence and the Question of

Human Distinctiveness

The language argument is a classic argument for human

distinctiveness that, for millennia, has been used to dis-

tinguish humans from non-human animals.

1

Aristotle

famously called human beings the animal with logos,

which, although usually translated as “reason,”is more

accurately understood as “language”(Gadamer 1977).

2

While dogs and dolphins can communicate by sending

limited signals of information to each other, few would

equate these limited capacities with the limitless subject-

discussion characteristic of human language (Gregg 2022,

61). Humans have long fantasized about creating artif‌icial

life that communicates and responds to us in a way that is

virtually indistinguishable from or even equivalent to

human language. An advanced language-speaking arti-

f‌icial intelligence machine—like a shrewd H.A.L.

9000 capable of communicating verbally with human

astronauts or Victor Frankenstein’s well-spoken

Creature—have always been a projection of a possible

future, usually created to warn current generations of the

need to ask ethical questions amidst the rapid growth of

technology, science, and culture (Botting 2021). Only

very recently has the possibility of AI employing

something like natural language become a reality. The

explosion of interest in generative language models since

early 2023 (e.g., OpenAI’s ChatGPT, Google’s Bard, or

Microsoft’s Bing) is evidence of this. The generative

language models are remarkable for being able to con-

struct coherent, grammatically correct, and informative

prose. The most advanced of these models, operating

under the large language model (LLM) architecture—like

OpenAI’s GPT-4—can be tasked with text summariza-

tion, chatbot behavior, live web searches, code generation,

and essay generation. Some AI models have even become

capable of generating content with greater f‌luency and

precision than human authors.

3

The rapid advancement of generative language models

has convinced many that language generation models

have f‌inally acquired the capacity for natural language.

1

University of Wisconsin-Madison, Madison, WI, USA

Corresponding Author:

Phillip Pinell, University of Wisconsin-Madison, 110 North Hall,

1050 Bascom Mall, Madison, WI 53706, USA.

Email: pinell@wisc.edu

Generative models thus pose a challenge to traditional

language-based models of human distinctiveness pre-

cisely because they can communicate and respond in a

manner resembling humanity’s linguistic capabilities.

The question for this paper is this: to what extent, and

in what ways, can the “language”of generative language

models accomplish the tasks that natural language can?

4

In posing this question, I am not asking whether AI have

become human by acquiring linguistic capabilities.

Rather, I am assessing whether this classic feature of

human distinctiveness vis-`

a-vis the functions of language

is no longer unique to humans. Answering this question

requires that we examine not only the linguistic content

that these models produce but also what the models

themselves are doing when generating linguistic content

and “learning”to perform more complex tasks from past

errors. Employing Gadamer’s theory of language, this

paper assesses the linguistic capabilities of AI to deter-

mine whether recent generative language models have

acquired natural language.

5

In this assessment, two cat-

egories are analyzed: (1) the linguistic content of the

sentences that generative language models produce and

(2) the actions that AI perform to generate language. It is

argued that generative language models have not acquired

an artif‌icial equivalent to natural language because they

lack the elements of groundedness to the world—and by

extension, understanding—community, and tradition that

are embedded in natural language acquisition and ex-

pression. While the content produced by generative

language models can be understood as a “text”with

linguistic content and a “world”that can merge with our

own, the models that produce this content lack a lin-

guistically mediated connection to the world and hence

have not acquired the exact functions of natural language.

My argument proceeds as follows. I begin by re-

constructing Gadamer’s theory of language. Next, I an-

alyze the claims about the linguistic capacities of AI

language models and assess whether they can generate

linguistic content—“texts,”in Gadamerian terms. I then

assess four categories that demonstrate why generative

language models lack language according to Gadamer’s

linguistic theory: disclosure to the world, understanding,

linguistic tradition, and linguistic community. I conclude

that generative language models are merely artif‌icial

language, and I ref‌lect on whether these models could ever

acquire the functions of natural language with further

software advancements.

Gadamer’s Theory of Language

The history of political thought is ripe with thinkers who

have discussed theories of language and artif‌icial intel-

ligence, respectively. Among philosophical approaches to

linguistics are Rousseau, Herder, Humboldt, Derrida,

Ricœur, and others.

6

Those interested in artif‌icial intel-

ligence nowadays are even greater in number (e.g.,

Kasirzadeh and Gabriel 2023;Monett and Lewis 2018;

OpenAI et al. 2023;Paris, Swartout, and Mann 2013;

Rakover 2023;Rodman 2023;Wu et al. 2023). Strictly

speaking, Gadamer did not write about artif‌icial intelli-

gence. He did, however, write extensively about language.

I turn to Gadamer for my analysis for the utility of his

philosophical hermeneutics in current linguistic debates

on AI technology. Gadamer’s theory of language offers

insight into the deeper structure of human language

through his phenomenological approach. Gadamer’s

philosophical hermeneutics is more useful for an as-

sessment of the linguistic structures of AI, for example,

than the scientif‌ic methods employed by computer sci-

entists and linguists because it investigates the experience

of what language is and does, in addition to the observable

effects of language. In so doing, Gadamer’s linguistics

goes beyond the merely empirical assessment of an AI’s

textual output, and instead points toward the fundamen-

tally phenomenological question of what relation to the

world enables the production of that output in the f‌irst

place.

7

Hence, Gadamer’s approach will help assess both

whether AI can generate texts with linguistic content, and

whether AI and humans share the linguistic connection to

reality that has long been referenced to distinguish hu-

mans from non-human animals.

Our f‌irst task is to understand Gadamer’s seemingly

abstruse def‌inition of language. For Gadamer, “Being that

can be understood is language”(Gadamer 2003b, 474).

This understanding of language approaches linguistics

through the lens of phenomenology. His conception of

language amounts to far more than words and syntax—

what Darren Walhof has called “the toolbox model of

language”(Walhof 2017,18–19).

8

Gadamer understands

language as the medium that makes possible the disclo-

sure of the world to us. The disclosive character of lan-

guage is possible only because what is said and written is

always connected to and situated within much more that is

unsaid or unwritten. The actual words used in a particular

context always point beyond themselves, to the whole of a

given language in which truth is sedimented. To interpret

something spoken, then, requires that we look both at the

words of a given text and the external direction that the

words point us toward.

The central task of Gadamer’s hermeneutics is getting

at the meaning of “texts”—which, not exclusive to written

text from books, can include events and experiences. Such

texts may include written expressions of language but

could also include the non-spoken features of commu-

nication. For example, language contains a historical

component, since there is a whole linguistic tradition that

outlasts any interpreter’s individual consciousness.

9

There

is also a communal element to language development

714 Political Research Quarterly 77(3)