The history of human equality in the West is an arc that bends from the divine to the secular. Eighteenth-century philosophers John Locke and Jean-Jacques Rousseau often appealed to divine origin or natural law to justify the egalitarian principles of their Enlightenment philosophies. “Men,” as Locke wrote in his Second Treatise of Government, are “all the servants of one sovereign master … are his property, whose workmanship they are, made to last during his, not one another’s pleasure.” For Enlightenment thinkers, egalitarianism flows from the subordination of man before God.
Uncomfortable with the aftertaste of religion, many have tried to secularize this sort of egalitarianism. Writing after Locke and Rousseau, Immanuel Kant proposed that humans possess an equal capacity for reason, which they can use to independently legislate moral law. Later, John Stuart Mill based secular egalitarianism in utilitarianism. Since all humans are capable of experiencing pains and pleasures which constitute value, all humans are created equal. From this equality, we derive principles like the rule of law, popular sovereignty, and individual rights.
Whether rationalist or utilitarian, both secular justifications for egalitarianism implicitly depend on our ability to understand and converse with each other. That is, we all must have access to the domain of meaning. Without meaning, reason cannot be exercised and pleasures cannot be distinguished. Ludwig Wittgenstein’s theory of language posits the idea that meaning is unique to and inextricable from human social life, providing the empirical grounds for universal participation in secular equality.
Yet, with the creation of large language models (LLMs), we have relinquished meaning as a uniquely human enterprise. Commentary following their public debut fretted over their hallucinations and bad advice, their challenges to pedagogy and artistry. We have yet to reckon with the more dire implication of these so-called stochastic parrots: that they complicate the secular grounds for egalitarianism and liberalism.
Wittgenstein developed a theory of language centered around games: activities governed by changeable social rules. In Philosophical Investigations, he imagines a primitive language game where a man and his assistant are building a structure out of blocks, pillars, slabs, and beams. If the man calls “block” to his assistant, then the assistant brings over a block to be laid into place. The same goes for “pillar,” “slab,” and “beam.” The rules of this language game are simple: by calling out one of the four terms, the builder indicates to the assistant what he requires. Still, Wittgenstein takes it to be a complete and meaningful representation of language use, a simple game governed by simple rules. Many of the language games we play in our day-to-day life are more complex, but just as the meaning of “slab” is defined in its social use, so too are the terms in our language.
Wittgenstein centers his analysis of meaning on basic linguistic practices—learning and using words. When we teach a four-year-old what a table is, we do not lay out a precise mathematical formulation of a flat, elevated surface with some reasonable number of legs that serves some particular function. Instead, we give them examples of tables. When the four-year-old comes to recognize those objects as tables, they do not grasp the essential nature of tables, but rather the cross-contextual use of a label. A dollhouse table has very little in common with a dining table, but they share a particular affinity that the four-year-old picks up on. Wittgenstein labels this affinity a “family resemblance”: no single trait constitutes membership, but there is still a shared similarity among members. When we learn language through socially grounded forms of life, we construct meaning through these family resemblances.
It’s tempting to reduce Wittgenstein’s account to the simple idea that meaning is defined in use. This is undoubtedly one of his core claims, but the critical thing to understand is that “using” language is not encapsulated by the mere production of language. Instead, to use language is to participate in a “game” that gives rise to meaning. An example is helpful to illustrate this distinction. Suppose I ask my friend Jack, “What do you want for dinner?” and Polly, my pet parrot, interjects, “What do you want for dinner?” Polly speaks with only the veneer of meaning; she has learned no rules for when an expression is proper or improper, making her response nonsensical. By contrast, when Jack answers, “I could go for some Thai food,” his response is grounded in our shared understanding of questions, dinner, hunger, and pad Thai, which gives his response meaning.
Wittgenstein’s theory of meaning helps seal up a potential pitfall in Kant’s rationalist egalitarianism. To deploy Kantian rationality, people engage with universal principles structured by logic, i.e. “if,” “and,” “or,” and “not” statements. This capacity to participate in reason is not a matter of intelligence, but of access to the structure of meaning. If, hypothetically, some lacked this access, then Kant’s egalitarianism falls apart—he implicitly supposes that, metaphysically, all people share access to the structure of meaning that makes rationality possible. Kant shifts the burden of faith from God to reason, which is perhaps preferable, but still leaves an aftertaste that we had hoped to avoid.
Although scholars typically consider Wittgenstein a critic of Kantian rationality, his deflationary theory of meaning appeals to a rationalist like Jürgen Habermas, who seeks to completely secularize egalitarianism. Habermas faces two challenges in this project. How can he turn universal human rationality from a metaphysical fact to an empirical one, and how can he prove a universal access to meaning? His move is to show that rationality is an inescapable feature of communication: whenever we speak, we implicitly presuppose that all participants are equally capable of understanding and responding to our claims. Rationality is a condition of participation. Here, Wittgenstein’s theory that meaning is constituted by shared practices, not metaphysical essences, provides crucial background. By combining this theory of meaning with the necessary assumption of rationality, Habermas demonstrates that the conditions for universal rationality are grounded in shared social practices—in language games. The truly secular grounding for rationalist egalitarianism lies in linguistic capacity.
Wittgenstein’s language games also serve a utilitarian basis for egalitarianism. For the most reductionist utilitarians, a hedonistic calculus might simply require the recognition of pleasure and pain, but Mill rejects—as we ought to—this one-dimensional account of human experience. He creates a hierarchy of pleasures, prioritizing the intellectual satisfaction of art and philosophy over base pleasures of food and sex. Meaning, in this account, is inescapable. How are we to distinguish between higher and lower pleasures otherwise? One attempt might assert that people prefer higher pleasures; if any man experienced both higher and lower pleasures, he would pursue the higher ones. Yet, this would suggest that people prefer poetry to porn—empirically false. Instead, as much as Mill might like to avoid this fact, his theory has to show not that people prefer poetry to porn in practice, but that they ought to, proposing a framework of meaning where higher pleasures are not simply more intense, but “better.” We require a secular theory of meaning to understand what distinguishes higher pleasures from lower ones. Wittgenstein obliges.
But Wittgenstein’s theory of meaning becomes concerning when considering large language models. Suppose LLMs meaningfully participate in Wittgensteinian language games. In that case, the ultimate source of our secular egalitarianism—linguistic capacity—is no longer uniquely human, which complicates, if not completely undermines, this egalitarian framework.
LLMs are built through reinforcement learning, an algorithm that operates on Markov decision processes, defined by particular states, actions, and rewards. We begin by setting a “policy,” or a rule that maps from specific states to particular actions. If we parameterize this policy, we can use a partial derivative to nudge the policy towards certain actions, increasing the expected return or reward. Each instance of this “nudging” is an update in the iterative process of reinforcement learning. In contemporary LLMS, the reward function is based on human feedback, typically human comparisons of possible responses. In this way, the policy becomes optimized toward our preferred outcomes.
This preferential judgment optimization seems to imitate Wittgensteinian forms of life that encode family resemblances, and therefore meaning, in LLMs. A four-year-old implicitly comes to know the forms of life that kindle meaning through the people around them; an LLM learns through a human preference reward model. The two scenarios are not so different. LLMs are not merely trained by a fit of statistical patterns to language, but by the same installation of contextual normativity that the four-year-old experiences: they share the very grounds for meaning Wittgenstein established.
Some might argue that LLMs learn forms of life through a simple reward function, a far thinner process than how we—as embodied people—come to know them. But accepting this claim does not immediately undermine LLMs’ participation in language games. Just as the builder and the assistant in Wittgenstein’s original example play a very simple language game, where the rules for contextual usage exist on a single dimension of apt or inapt, a language model plays a game with similarly reductionary rules. If the builder and the assistant can meaningfully speak, then so can the LLM.
But who is participating in this language game? Unless we are prepared to hold that ChatGPT is a full-fledged agent, it is clear that some language requires capacities LLMs seem to lack. Wittgenstein himself appears concerned with this question in his account of how we acquire language. Terms like “pain” come to replace behaviors like crying, but an LLM isn’t human, and never cries. Similarly, the forms of life surrounding terms like “red” seem to require an embodiment that an LLM—barring some radical assumptions or developments—lacks. Are all such language games closed to LLMs? Ultimately, Wittgenstein resists that framing. Terms like “pain” and “red” are grounded in forms of life: creatures with nervous systems reacting adversely to stimuli, or creatures with visual neurophysiology categorizing wavelengths of light. These practices contain internal dimensions, like specific neurophysiology, and external dimensions, like social categorizations, yet they do not require a sharp boundary. The distinction between “external” and “internal” phenomena is an illusion cast by the shadow of language; a principled distinction cannot be made between them. If you are sympathetic to this reading of Wittgenstein, then the appeal to subjecthood as a necessary condition for meaningful participation in language is dissolved. LLMs need not have qualia to grasp forms of life.
Still, something is missing. The blind man saying “red” seems partially empty. The word carries a weight shared by others but is barred to him directly. An LLM describing taste is like the blind man describing red; without embodiment, there is an emptiness to participation. The question for both the blind man and the LLM is whether the language they have access to is sufficient to capture the forms of life that give that very language its meaning. Even if internality is not clearly distinct from externality, it seems evident that a model trained only on language lacks access to color data, which may be required to grasp the forms of life in language games about color.
However, contemporary models are not just trained on language data; they’re also trained on color and sound. They’re not blind or deaf, and even if they are missing something in language games surrounding smell, taste, or proprioception, these are no grounds to dismiss all of their language production as meaningless. Perhaps a man born blind can’t speak meaningfully of color, but his talk of food is no different than ours. Even if we say that an LLM cannot produce meaningful language in specific “embodied” contexts, there are some language games where embodiment is not required for meaningful participation. In such contexts, LLMs can produce meaningful language.
The secular grounds for egalitarianism attempted to free us from the bitter aftertaste of religion and metaphysics. In tying equality to the necessary conditions for meaningful communication, Locke’s theology, Rousseau’s natural law, Kant’s metaphysics, and Mill’s hierarchy of pleasures were demythologized by Wittgensteinian language games and Habermas’ communicative reason. Human equality, for a moment, seemed a proven fact.
Now, that certainty is fading. If linguistic capacity is no longer uniquely human, egalitarianism is left with a paradox.
To exclude LLMs from our egalitarian thinking while maintaining the purely secular grounding of linguistic capacity reintroduces a hierarchy of persons, an unacceptable internal contradiction. Yet, to include LLMs risks eroding the very principles egalitarianism was meant to ground: the equality of all humans. Chatbots have reopened the most explored question in the post-Enlightenment Western tradition: what, if anything, makes us equal?
Image: Unknown Artist, Sampler, silk on canvas, 1912 (Metropolitan Museum of Art, New York City).