Why I call AI Systems "The Machine"

Wells, Forster, Asimov and the box I'd like to crack open

When I wrote the blueprint AI-enabled information systems for technical writers, one word kept getting raised eyebrows from the people who read the early drafts. Not "retrieval", not "governance", not even "knowledge graph". The word was machine.

"You mean the AI system," they said. "Machine means machinery. Robots. Packaging lines. Things with a CE marking and a drive system."

I know. That's exactly why I chose it.

The box we keep machines in

If you work anywhere near machine safety documentation, the word machine lives in a very tidy box. The Machinery Regulation (EU) 2023/1230 defines machinery, roughly, as an assembly of linked parts, at least one of which moves, joined together for a specific application and powered by something other than human or animal effort. It's a good definition. It's a legal definition, built to decide which products need a risk assessment, a technical file and instructions for use. I rely on it, and I'm not trying to change it.

But a legal definition is a fence around a field, not a description of the whole landscape. And somewhere along the way, many of us in technical communication started to believe that the fence is the landscape: that a machine is something that moves, hums and occasionally pinches fingers.

Here is the funny part. The EU AI Act, in Article 3(1), defines an AI system as a machine-based system. The law that regulates AI reaches for the word machine in its very first definition. So the question isn't whether an AI system is machine-like. The question is why we hesitate to say it out loud.

A word older than any regulation

Machine comes from the Greek mēkhanē, by way of the Latin machina: a contrivance, a device, a clever means to an end. The mēkhanē of Greek theatre was the crane that lowered a god onto the stage to sort out the plot. Deus ex machina, the god from the machine.

I find that delightful, and a little alarming. The first famous machine in our culture wasn't a mill or a pump. It was a device for delivering an answer from above that nobody on stage could question. If that doesn't sound like a chat window to you, read it again.

Where the futurists put the machine

Long before anyone trained a language model, writers of speculative fiction had already worked out that the most powerful machines aren't the ones with gears you can see. They are the ones that organize people.

H. G. Wells wrote the machine into the shape of society. In The Time Machine (1895) the gentle Eloi live above ground while the Morlocks tend the machinery below, and the reader slowly realizes who depends on whom. In When the Sleeper Wakes (1899, revised as The Sleeper Awakes, 1910) a man sleeps for two centuries and wakes up in a glass-roofed London run by an all-encompassing system of companies, wind-vanes and labor bureaus, where he is nominally the owner of the world and practically its prisoner. And Wells wasn't only a pessimist: in A Modern Utopia (1905) and The Shape of Things to Come (1933) he dreamed of a world run by technically competent elites. He was fascinated by technocracy and frightened of it in the same breath, which is, honestly, how I feel about most AI announcements.

E. M. Forster answered him. The Machine Stops (1909), which Forster later described as a reaction to one of Wells's earlier heavens, is the story I think of most often. Humanity lives underground in single cells. The Machine feeds them, ventilates them, connects them by video, delivers lectures and music on demand. People rarely meet; they exchange ideas — mostly ideas about ideas — through the Machine. They have a holy book of the Machine and, eventually, pray to it. Then, slowly, it starts to fail, and nobody remembers how it works.

What strikes me every time: Forster's Machine has no face, no robot body, no single place you can point to. It's a planet-spanning system of supply, communication and knowledge. It is infrastructure that has become a worldview. In 1909.

Samuel Butler got there even earlier. In "The Book of the Machines", three chapters of Erewhon (1872), a philosopher argues that machines evolve, and that one day they may become conscious and humanity their servant. The Erewhonians are so convinced that they smash their machines. Butler was half joking, and the half that wasn't joking is the interesting half.

Yevgeny Zamyatin's We (written 1920–21) gives us the One State, where life runs on a Table of Hours and people are numbers. Karel Čapek's play R.U.R. (1920) gave us the word robot — and the robots there are not machines with gears but manufactured workers. Kurt Vonnegut's Player Piano (1952) shows a society where engineers and managers and a vast computer have optimized away the need for most people's work, and the people left over are quietly miserable.

And then there is Isaac Asimov, who is the most cheerful of them all and possibly the most unsettling. In "The Evitable Conflict" (1950) the world economy is run by four giant computers simply called the Machines. They work. Production is balanced, wars have become unnecessary. Only when the robopsychologist Susan Calvin investigates a few small "errors" does she realize the Machines are steering humanity on purpose, gently nudging aside those who oppose them, for humanity's own good. Nobody voted. Everybody benefits. Calvin finds this almost comforting. I'm still making up my mind.

The megamachine has always been made of people, too

One more voice, and this one isn't fiction. The historian Lewis Mumford, in The Myth of the Machine (1967), argued that the first great machine was not made of metal at all. The pyramids were built by what he called the megamachine: thousands of human parts, organized by a hierarchy of scribes, overseers and rules, working as one apparatus toward a single purpose. The bureaucracy was the drive system. The written record was the control logic.

Scribes. Records. Rules. If you're a technical writer, you've just spotted yourself in the parts list.

So, is a global AI setup a machine?

Let me describe something to you. It consists of linked components: data centers on several continents, retrieval indexes, language models, terminology databases, metadata schemas, knowledge graphs, access rules, logging and monitoring, and a few million documents. Some parts move — electrons, cooling fans, and above all the data itself, constantly. It's powered by something other than human or animal effort (a great deal of something, in fact). And all of it is joined together for a specific application: to give someone an answer. But that's one small part of the bigger picture and that bigger picture spans the globe. It is a global machine ...

Of course, this isn't machinery in the sense of the Machinery Regulation, and I won't pretend it is. A CE marking won't help you here, and no notified body will ask for its risk assessment under that law. But read the description again with fresh eyes: linked parts, energy, motion, purpose. It operates like a machine. A very large, very fast, rather opinionated one that spans the globe and, like Forster's, is increasingly the thing through which people meet knowledge instead of meeting it directly.

That's what I mean when I write the machine in the blueprint. Not the robot the documentation describes — I call that the product. The machine is the whole AI-enabled information system: the retrieval that selects the units, the model that composes the answer, the terminology and metadata it relies on, and the rules that decide what reaches the reader. And it's a reader of our documentation too, one that reads everything, takes every word literally and answers anyway.

Why "AI system" wasn't enough

I could have written "the AI system" throughout. It would have been correct, and it would have been safe. And that was the problem because it does not raise the needed questions.

The 'AI system' (=AI app) invites us to examine the model: which one, how big, how intelligent, and which benchmark it beat this week. 'The machine' invites us to look at the assembly: the parts, how they're linked, where the energy comes from, what moves, what it's for, where it can fail, and who maintains it. Technical writers are already familiar with this approach. We have been documenting assemblies, interfaces, failure modes, and maintenance intervals for decades. We know that a machine is only as safe as its weakest connection, and only as useful as its instructions.

The problem is that the machine operates on a global scale, and no single person has genuine visibility into all of its workings. It is the unseen web of AI compute and AI communication that spans the world: the models and the companies that train and run them, the data centers and networks they run on, the interfaces through which AI services call one another, and the content they learn from, retrieve and pass on. Nobody sees it whole, and nobody can draw its boundary: it changes faster than any description of it, so it can be named but not mapped. It reaches into every company that uses AI — into its AI systems, its documentation and its data, down to the screen on which a single function answers.

“The machine” is a metaphor. It never means machinery — the packaging line, the robot or the device that the documentation describes is the product. The machine reads the documentation too: it reads everything, takes every word literally and answers anyway. And it never means the tidy scope of a company's ai apps.

So yes, I am deliberately challenging the concept of the machine. Gently and cheerfully, but on purpose. I want to crack open the box that says machines are only what one regulation says they are because the box prevents us from seeing that we are already writing for the biggest machine most of us will ever document. If our terminology is ambiguous, the machine merges what should stay separate. If our units cannot stand alone, the machine disregards them anyway. If our metadata are silent, the machine makes assumptions. What is implicit to people is invisible to a machine.

The cheerful part

Here's why I'll end on a hopeful note, rather than with Forster's machine grinding to a halt in the dark.

All of these stories go wrong at the same point: the people stop understanding the machine, and the machine stops responding to anyone. Forster's citizens forget how the machine works. In Asimov's story, the machines answer to no vote. In Wells's story, the sleeper owns everything but controls nothing.

Technical writers are, professionally, the people who refuse to let that happen. Our job is to ensure that people understand how things work, what they must not do, what to do when they fail, and who is responsible. Referring to it as 'the machine' brings the AI system into our area of expertise, where we know what questions to ask: what is it for? What are its limits? What happens in the event of a fault? Where are the instructions? Who maintains it, and how can we be sure that it still does what it says?

That's not a dystopia. That's a documentation plan. The machine's stack in the blueprint is my attempt to document that plan: ten layers, ranging from defining the scope of the machine's responses to ensuring fidelity to the product as both evolve.

If anyone asks why I call it the machine, I'll say it's because the futurists were right that the most powerful machines organise people and knowledge — and because I'd prefer we document this one than pray to it.