Marshall McLuhan argued in the 1960s that the content carried by a medium matters less than what the medium itself does to us — how it reshapes cultural structures, sensory environments, and social behaviour. Half a century later, that framework maps onto generative AI with unsettling precision.

A media theory analysis published by the data and AI consultancy Statworx applies McLuhan's lens directly to large language models (LLMs), arguing that the dominant public debate about generative AI — is it intelligent? will it hallucinate? will it surpass humans? — systematically misses the more consequential question: how is AI, as a medium, transforming what it means to be human?

The Medium Is the Message — Applied to LLMs

McLuhan defined a medium as an extension or reduction of human senses and the body. Its "message" is not the content it delivers but "the change of scale or pace or pattern that it introduces into human affairs." By that definition, generative AI is unmistakably a medium: it does not merely transmit information, it restructures how information is sought, processed, and valued.

The analogy to earlier media is instructive. The printing press diminished oral tradition and communal storytelling while enabling independent reading and the private accumulation of knowledge — conditions that produced the Enlightenment. Television introduced multisensory, emotionally driven communication and eroded the linear, rational processing associated with text, while simultaneously creating what McLuhan called the Global Village. Each medium reorganised cognition and society, not through any single piece of content it carried, but through the new cognitive and social architecture it imposed.

Generative AI follows the same pattern — but operates at a different layer. Where the printing press extended access to recorded knowledge and television extended audiovisual perception, LLMs extend something closer to reasoning itself. They retrieve, synthesise, and generate language at a scale no human can match, which means the human skills that retain value are precisely those AI cannot replicate: strategic and procedural knowledge (knowing how to instruct and direct AI systems effectively), and social and contextual knowledge (understanding the norms, relationships, and situational conditions AI cannot perceive, lacking both consciousness and a physical body).

Emergence, the Zettelkasten, and Neural Networks

The article draws a compelling parallel between LLMs and Niklas Luhmann's Zettelkasten — the 90,000-note slip-box the sociologist maintained between 1951 and 1997. Luhmann's system worked through an intricate cross-referencing method that allowed ideas from disparate domains to collide unexpectedly, producing insights he described as arising from a "dialogic partner." Neural networks operate analogously: by forming complex, non-linear interconnections between data points, they can surface patterns and relationships that were not explicitly designed into them.

Whether this constitutes genuine emergence — phenomena arising from component interactions that produce properties unpredictable from those components alone — remains contested. Researchers at Google DeepMind have argued mathematically that AI systems must learn implicit causal models of data relationships to adapt robustly across changing conditions. Yet critics note that current LLMs appear to rely heavily on approximate retrieval and recombination of training data, rather than true generalisation to novel, out-of-distribution problems. The evidence for genuine causal reasoning remains inconclusive.

What is less contested is that skilled prompting — structuring questions to encourage LLMs to form complex conceptual connections across domains — can generate ideas that would not have emerged through unaided human cognition alone. In that sense, the thinking-machine framing holds, whether or not the machine is "thinking" in any philosophically rigorous sense.

Data as the Deeper Medium

The analysis pushes the McLuhan framework one level further: if media reshape how we perceive information, what happens when data itself is understood as a medium? The argument is that data is not merely the input to AI systems — it is the substrate on which AI learns to see the world. The quality, diversity, and representativeness of training data determine what an AI system can perceive, and therefore what it will systematically miss.

From this vantage point, AI becomes its own message: it symbolises and enacts the transition to a data-driven world in which human and machine intelligence increasingly merge. The mere existence and functionality of AI already reshapes what skills are considered valuable. A Microsoft study cited in the analysis found that 66% of executives say they would not hire someone without AI skills, and 71% would prefer a less experienced candidate with AI proficiency over a more experienced one without it — evidence that AI is restructuring labour markets not through any specific output, but through the structures and expectations its presence creates.

Why the Framing Matters

The practical stakes of the medium-versus-tool distinction are significant. If AI is primarily understood as a tool — a sophisticated instrument for generating text, images, or code — then the evaluative questions are instrumental ones: Is the output accurate? Is it creative? Is it safe? Those questions are not unimportant, but they bracket the deeper structural effects.

If AI is understood as a medium in McLuhan's sense, the questions change. They become: What cognitive habits is AI making obsolete, and which is it making indispensable? How is it reorganising the boundary between human and machine agency? What social and cultural structures is it quietly reshaping, independently of any particular content it produces?

Those are the questions that historically have proven most consequential — and most consistently underestimated — each time a new medium has emerged.

Source: Statworx Content Hub, June 2024.