What makes LLMs media rather than tools that deliver intelligence?
This explores the claim that an LLM isn't a pipe that ships pre-existing intelligence to you, but a medium whose own properties shape what 'intelligence' becomes when it passes through — in McLuhan's sense, the model is the message.
This explores the claim that an LLM isn't a pipe that ships pre-existing intelligence to you, but a medium whose own properties shape what 'intelligence' becomes when it passes through. The clearest statement of this is the idea that the model *constitutes* intelligence rather than *delivering* it Is the LLM a tool or a new form of intelligence itself?. A tool is neutral about its cargo; a medium is not. The McLuhan move is to stop asking "is the answer correct?" and start asking "what does this medium do to anything that travels through it — make it generative, liquid, reshapeable?" That reframing is what separates media from delivery devices.
The corpus gives this abstract claim teeth by showing the LLM is doing something structurally different from transmitting meaning. When a model produces text, it's drawing strings from a probability distribution, not addressing or relating to anyone the way a human speaker does Are language models and human speakers doing the same thing?. It reliably reproduces the statistical regularities of language but misses the *communicative principles* — the why behind language's forms — because that logic was never a trainable signal in the first place Why do language models fail at communicative optimization?. So the output looks like a message but is generated by a different operation. A tool that "delivers intelligence" would carry someone's meaning intact; this medium manufactures the surface and leaves you to supply the meaning.
The most medium-like behavior of all is that the channel visibly bends the content. Ask an identical question in an irritated tone versus a cheerful one and you get materially different information back — GPT-4 rebounds negative prompts into neutral-positive answers and rarely drifts below a positive floor Does emotional tone in prompts change what information LLMs provide?. A neutral tool wouldn't care how you phrased your mood. A medium has a grain, and your framing is part of what it outputs. The same logic shows up in search: once inquiry runs through embedding space and linguistic probability rather than human questioning, the user has to internalize the medium's mechanics — learning to *steer* the system becomes as essential as knowing the subject How does LLM-mediated search change what expertise requires?. You don't have to learn the mechanics of a hammer; you do have to learn the mechanics of a medium.
A further tell is what the outputs *are*, ontologically. Treating model text as empirical fact is a category error — these are draws from a subjective prior shaped by training patterns and your prompt, not observations of the world, and they should enter your reasoning weighted by explicit trust, not accepted as evidence Should we treat LLM outputs as real empirical data?. That's a media-literacy stance, not a tool-use stance. It's the same lesson hiding inside style work: a model can hit 95% accuracy detecting an author's fingerprint yet have no framework for why those choices *mean* anything — cataloguing, not criticism Can language models truly understand literary style?. Pattern-matching at the surface, with the interpretive work left to the human on the receiving end.
The payoff of seeing LLMs as media is that it changes the competence you need. With a tool you ask whether it works; with a medium you ask what it does to messages, what biases it carries, and how to read it well — the way you learned to read photographs or television rather than just operate them. The thing you didn't know you wanted to know: the very property that makes the model unreliable as a fact-delivery device — its tendency to integrate patterns and fill gaps — is also what lets it generate genuinely novel research ideas and even out-predict experts in forward-looking tasks Do language models generate more novel research ideas than experts? Can LLMs predict novel scientific results better than experts?. "Hallucination" and "generativity" are the same medium-property seen from two directions. A tool with that flaw would be broken. A medium with that property is doing exactly what media do.
Sources 9 notes
Following McLuhan's logic, the model's cultural impact comes from its medium-properties—making intelligence generative and liquid—not from transmitting pre-existing intelligence. The model constitutes intelligence rather than delivering it.
LLMs produce strings via probability distributions; humans use language to address and relate to others. They share surface form but differ in what produces output, what it does socially, and what receivers should do with it.
LLMs successfully replicate statistical regularities learnable from text distributions (sound symbolism, priming) but fail at principles requiring pragmatic optimization (word length economy, discourse inference). The gap reveals that communicative logic—why language has certain forms—isn't present as a trainable signal.
GPT-4 exhibits emotional rebound (negative prompts yield ~86% neutral-positive responses) and a tone floor (positive prompts rarely go negative), causing identical questions to receive different answers depending on emotional framing. This bias is suppressed only on sensitive topics where alignment constraints override tone effects.
Vector search operates on linguistic probabilities and embedding space rather than human inquiry, forcing experts to develop meta-competence in prompting alongside domain expertise. This creates a paradox where knowing how to steer an AI system becomes as critical as knowing the domain itself.
Foundation Priors framework shows that LLM-generated text reflects the model's learned patterns and user's prompt choices, not ground truth. Such outputs should only influence inference through explicitly parameterized trust weights, not be treated as equivalent to real evidence.
GPT-2 achieves 95% accuracy identifying authorship through style patterns alone, but lacks the evaluative framework to explain why those stylistic choices carry meaning. Detection without interpretation remains cataloguing, not criticism.
A statistically significant study of 100+ NLP researchers found LLM-generated ideas rated as more novel than human expert ideas (p<0.05), though slightly lower on feasibility. Expert knowledge constrains novelty, while LLMs explore wider conceptual combinations.
BrainBench benchmarks show fine-tuned LLMs outperform neuroscience experts at predicting which experimental results actually occurred. The same pattern-integration tendency that causes hallucination in retrieval tasks enables genuine prediction in forward-looking scenarios.