Realism and anti-realism in an age of artificial intelligence

By Matthew Parish, Associate Editor

Tuesday 4 August 2026

The debate between realism and anti-realism in the philosophy of language has acquired an unexpected urgency in the age of artificial intelligence. For decades it was a dispute that occupied philosophers, linguists and logicians, often appearing detached from practical affairs. Today, however, the emergence of large language models has transformed what once seemed an abstract disagreement into a matter with immediate consequences. When a large language model writes persuasively about history, law, poetry or war, the question is no longer merely what language means. It becomes whether meaning itself requires an underlying reality to which words correspond or whether language is simply an intricate system of relationships between symbols. Large language models stand precisely at the fault line between these competing conceptions.

The realist tradition is deeply rooted in Western philosophy. Although its forms differ, realism broadly maintains that language succeeds because it refers to a world existing independently of those who describe it. Words acquire their significance because they denote objects, properties or states of affairs. The sentence “snow is white” is true because snow really possesses the quality of whiteness, not because speakers happen to agree upon the phrase. Truth, on this account, consists in correspondence between linguistic propositions and an external reality.

This conception stretches from Aristotle’s logic through medieval scholasticism to modern analytic philosophy. Even philosophers who profoundly disagreed over the mechanics of reference frequently retained the conviction that successful language ultimately depends upon a world capable of making statements true or false. Scientific discourse particularly appears to vindicate realism. Physics progresses because nature constrains theory. Engineers cannot negotiate with gravity by changing their vocabulary. Reality eventually has the final word.

Anti-realism approaches language from a markedly different perspective. It questions whether correspondence with an independent reality is either necessary or even intelligible as an account of meaning. Instead, meaning emerges from human practices, conventions, inferential relationships or social activity. Language is not fundamentally a mirror reflecting reality but an instrument through which human beings organise experience.

Ludwig Wittgenstein’s later philosophy represents perhaps the most influential modern expression of this perspective. Words do not possess meaning because they point towards hidden metaphysical entities. Rather, they acquire significance through their participation in “language games” embedded within forms of life. To understand the word “promise” is not to identify some abstract object corresponding to promises. It is to understand the complex network of practices surrounding promising. Similarly, mathematical statements, legal concepts and moral judgments derive their intelligibility from their role within communities rather than from straightforward correspondence with physical objects.

Other anti-realist traditions arrive at comparable conclusions by different routes. Pragmatists evaluate beliefs according to their practical consequences. Verificationists once argued that meaning depends upon methods of verification rather than metaphysical reference. Constructivists emphasise the social creation of categories through which human beings understand the world. While these schools differ substantially, they all diminish the role of an independently existing reality in explaining linguistic meaning.

Large language models occupy a remarkable position between these competing philosophies because they appear almost designed to test their implications. Unlike human beings, they possess no direct sensory experience of the external world. They do not see mountains, smell rain, touch stone or experience fear. Their understanding arises entirely through statistical relationships among linguistic tokens derived from enormous corpora of human writing.

If realism alone fully explained language, this circumstance would appear fatal. An entity without direct contact with reality should be incapable of meaningful linguistic competence. Yet large language models demonstrably exhibit extraordinary facility in generating coherent prose, explaining technical subjects, composing poetry and participating in sophisticated dialogue. Their success appears, at first glance, to strengthen anti-realist intuitions. Perhaps language truly consists primarily in patterns internal to language itself.

Indeed, much contemporary machine learning resembles an enormous empirical demonstration of structuralist theories. A language model predicts each successive word not by consulting reality but by calculating probabilities derived from previous linguistic usage. It discovers relationships among concepts because human texts encode those relationships statistically. “Paris” becomes associated with “France”, “capital” and “Seine” without ever requiring visual acquaintance with the city itself.

This success should not be underestimated. It reveals that an astonishing proportion of ordinary linguistic competence depends upon recognising internal regularities rather than continually consulting external reality. Human beings likewise acquire much of their vocabulary through conversation long before directly encountering everything they discuss. Few children observe electrons, black holes or constitutional law directly. They nevertheless learn to speak coherently about them through participation in linguistic communities.

Yet realism returns with equal force when one considers the limitations of these systems. Language models occasionally produce convincing falsehoods with complete confidence. They may invent legal precedents, fabricate academic references or describe nonexistent events. These failures are not random computational errors. They reflect the absence of direct constraint imposed by reality itself.

Human beings also make mistakes, but our linguistic practices are continually corrected through interaction with the world. A scientist performs experiments. A lawyer consults statutes. A journalist interviews witnesses. Reality interrupts imagination. For language models, by contrast, reality enters only indirectly through the texts upon which they were trained or through external tools integrated into their operation.

Consequently, their impressive linguistic competence simultaneously vindicates and limits anti-realist theories. Language clearly possesses a substantial autonomous structure capable of being mastered independently of direct worldly experience. Nevertheless, reliable knowledge ultimately requires contact with something beyond language itself.

The debate therefore becomes more nuanced than either traditional position anticipated. Large language models suggest that linguistic meaning exists upon multiple levels. At one level, enormous networks of relationships amongst words permit sophisticated reasoning, explanation and creativity without immediate sensory grounding. At another level, successful navigation of the external world eventually demands mechanisms by which language reconnects with empirical reality.

One might therefore distinguish semantic competence from epistemic authority. Language models possess remarkable semantic competence. They understand, in an operational sense, how concepts relate within the immense web of human discourse. What they lack is independent epistemic authority. They cannot settle disputes by appealing directly to experience because they possess none of their own.

Ironically, this observation also illuminates human language. Philosophers have sometimes imagined that every competent speaker constantly grounds every statement in direct acquaintance with reality. In truth, much human communication depends upon trust, testimony and inherited linguistic structures. We know vastly more through language than through personal observation. Civilisation itself depends upon this delegation of knowledge. Artificial intelligence merely magnifies the phenomenon by removing personal experience almost entirely.

The implications extend beyond philosophy into public life. As governments, businesses and universities increasingly rely upon artificial intelligence, the distinction between linguistic fluency and factual reliability becomes ever more significant. Persuasive language should never be mistaken for demonstrated truth. The elegance of a sentence remains no guarantee of its correspondence with reality.

Equally, critics who dismiss language models as mere “stochastic parrots” underestimate what they genuinely accomplish. The statistical architecture underlying these systems captures extraordinarily rich conceptual relationships accumulated across centuries of human culture. Those relationships are not empty simply because they emerge through probability rather than consciousness. They constitute a genuine form of linguistic intelligence, even if not identical to human understanding.

Perhaps the deepest lesson is one of philosophical humility. Neither realism nor anti-realism entirely anticipated what artificial intelligence would reveal. Reality constrains language more profoundly than anti-realists sometimes acknowledged, yet language exhibits a degree of internal autonomy far greater than many realists imagined. Large language models demonstrate that meaning can flourish within complex symbolic systems while simultaneously reminding us that truth ultimately depends upon something outside those systems.

Artificial intelligence has therefore not resolved one of philosophy’s oldest debates. Instead, it has transformed it from an academic controversy into a practical framework for understanding one of the defining technologies of our age. The question is no longer whether realism or anti-realism is entirely correct. Rather, it is how the strengths of each illuminate different aspects of language itself. The machine, paradoxically, has become an instrument through which humanity better understands not only artificial intelligence but also the nature of its own words.