Three years ago, in a bland office at Google’s headquarters, an engineer became convinced that the company’s chatbot had developed consciousness. Blake Lemoine’s widely publicized claims were swiftly dismissed, yet they exemplify a persistent muddle in society’s thinking about artificial intelligence. As language models like ChatGPT churn out increasingly convincing prose, a crucial question emerges: Have machines finally cracked the code of human language?
The evidence seems compelling at first glance. Modern AI systems engage in witty banter, write passable poetry, and help craft legal briefs. This has led some tech evangelists to revive a bold claim first made by Chris Anderson, former editor of Wired magazine, in 2008: that with enough data, theory becomes unnecessary. “Correlation supersedes causation,” he declared, suggesting that patterns alone could reveal all there is to know about the world. Applied to language, this thinking suggests that by ingesting enough text, machines could master human communication.
Yet this view stumbles upon a philosophical obstacle that no amount of data can overcome. Even if one could capture every utterance ever spoken—from Mesopotamian marketplace haggling to yesterday’s Twitter posts—multiple competing theories could explain the same linguistic patterns. This “underdetermination problem,” as philosophers call it, reveals why data alone cannot unlock the essence of language.
Consider a construction site where builders communicate with just four words: “block,” “pillar,” “slab” and “beam.” As philosopher Ludwig Wittgenstein observed, these words’ meanings emerge not from dictionary definitions but from their practical use—one builder shouts “slab!” and another brings the correct item. This simple example illuminates how language is grounded in what Wittgenstein termed “forms of life”—the shared activities and contexts that breathe meaning into words.
Modern language models, sophisticated as they are, operate more like extremely advanced autocomplete systems than true language users. They process language as Ferdinand de Saussure’s parole (instances of language use) but lack access to langue (the underlying system that generates meaning). The difference is akin to a cook who can only reproduce memorized recipes versus one who understands culinary principles deeply enough to create novel dishes.
The dynamic nature of language further exposes the limitations of AI systems. Consider “algospeak,” the evolving dialect of social media users trying to outsmart content moderation algorithms. When people write “unalive” instead of “dead” or “le dollar bean” instead of “lesbian,” they are not merely substituting words—they are actively participating in language evolution. Such innovation demonstrates how human language use is inherently diachronic (evolving over time) rather than merely synchronic (existing at a single point in time).
This is not to diminish the remarkable achievements of modern AI. Indeed, viewing these systems through the lens of the “extended mind hypothesis”—proposed by philosophers Andy Clark and David Chalmers—suggests a more nuanced perspective. Rather than threatening to replace human linguistic capacity, language models might better be understood as powerful cognitive extensions, much as writing systems dramatically expanded humanity’s ability to preserve and share thoughts.
The rise of sophisticated language models has not reduced human uniqueness so much as helped clarify it. What makes human language unique is the ability to manipulate symbols or recognize patterns and our embodied, participatory engagement in meaning-making. As these AI systems become increasingly integrated into daily life, they serve not as replacements for human linguistic capability but as mirrors that reflect what makes human language genuinely remarkable.
This has profound implications for how society develops and deploys AI systems. Rather than chasing the mirage of machines that truly understand language, a more productive approach would focus on developing tools that enhance human linguistic capabilities while acknowledging their fundamental limitations. The future of AI lies not in replicating human language but in augmenting it—a distinction that makes all the difference.
Leave a Reply