Cory Doctorow's Pluralistic Essay Argues LLMs Are Real but 'AI' Is a Marketing Fiction

Cory Doctorow published a new essay on his Pluralistic blog titled "God in the Box," drawing a hard line between large language models as functioning software and "AI" as a category he calls largely a marketing construct. The piece, dated September 12, 2026, argues that the gap between what these systems actually do and what the label implies is where much of the industry's commercial work happens. His distinction is narrow and deliberate: the model is real, the branding is not.

In Doctorow's framing, an LLM is a statistical system trained on text that produces useful output for summarization, drafting, and code completion. Those are genuine capabilities, and he does not dispute them. What he disputes is the umbrella term that bundles those tools together with claims about general reasoning, autonomy, and near-human cognition. That bundle, he argues, is where investors, executives, and vendors operate, because it supports valuations and narratives the underlying technology cannot carry on its own.

The essay lands amid a running argument over how much of the current boom rests on technical capability versus positioning. Doctorow has spent years writing about concentrated control of digital infrastructure, and his recent work has focused on ownership of platforms and models. Pluralistic, his long-running blog, is where he publishes those arguments alongside links and commentary, and it holds a steady readership among technologists and policy specialists. "God in the Box" extends that project to the vocabulary of the field itself.

Reaction split along lines the essay seems designed to expose. Some readers agreed that collapsing narrow language tools into a claim of general intelligence inflates valuations and muddies public understanding. Others countered that dismissing the "AI" label understates how fast model capabilities have improved and how widely they are already deployed in production systems. Both positions appeared in the same discussion, which suggests the disagreement is less about facts than about what the words should cover.

The stakes are not semantic. Naming shapes regulation, procurement, and where investment dollars flow. Treat a language model as a class of software and oversight looks like existing product-safety and consumer-protection law, applied through agencies that already exist. Treat it as a new kind of intelligence and the case for dedicated AI agencies, licensing regimes, and new liability frameworks gets much stronger. Doctorow's essay pushes toward the first reading, and that push has consequences for anyone building on these systems.

What to watch is whether this framing travels beyond blog threads and comment sections. The argument is likely to resurface as regulators draft rules for model developers and as companies market increasingly capable systems to buyers who cannot easily verify the claims. The open question is whether policymakers and the public keep treating "AI" as one thing, or start separating the working software from the story told about it.