The Hall of Mirrors
AI Is Not Channeling, Translation, or Copying
The moment someone mentions AI and authorship in the same breath, the legal frame snaps into place. Is it copyrightable? Who owns the output? Courts and copyright offices have had to answer those questions faster than anyone prepared for, and the answers have leaned on a metaphor that I find understandable — but mechanistically wrong.
The metaphor goes roughly like this: the AI is an oracle, an external intelligence delivering text to a human vessel. The human types a prompt and receives output, the way a medium receives a transmission. The law is suspicious of this arrangement because there’s no human author doing the originating work, just a passive recipient copying down what arrives.
But the right image isn’t an oracle. It’s closer to a calculator. You supply the inputs, the machine performs a mechanical operation on them, you get an output. No entity on the other end. No transmission arriving from somewhere else. Just input, process, output. The operation is vastly more complex than arithmetic — and that complexity matters — but the basic shape is the same. What changes the picture is not the complexity. What changes the picture is understanding that complexity well enough to work with it.
This framing has real legal history behind it — spiritual dictation cases were litigated decades before any chatbot existed — and it shapes how people argue about AI on both sides of the debate today. But it’s the wrong picture. And as long as we’re working from the wrong picture, we’ll keep drawing the wrong conclusions.
Let me show you three models people reach for when they talk about AI output, and why each one misses.
Model One: Channeling
The channeling frame says the AI is generating something from beyond the human — that an independent mind is speaking through the machine, and the human is just the pipeline. This is, in some ways, the most dramatic version of the argument, and it threads straight through certain copyright cases about divine or spiritual dictation.
Here’s what breaks the model: a channel has a source. When we talk about channeling, there’s something on the other end — an entity with intention, awareness, a point of view. AI has none of those things. The model running inside ChatGPT or any other large language model doesn’t have desires, doesn’t have opinions, doesn’t have a perspective on what it’s doing. It’s not waiting on the other side of a metaphysical barrier, choosing what to deliver. There is no “other side.” The system takes an input, processes it through a very large and complicated set of learned weights, and produces the statistically likeliest continuation. That’s it. Treating that as channeling is not just metaphorically strained — it attributes agency to a process that has none.
Model Two: Translation
Translation feels more grounded. A translator takes an existing text — ancient Greek, legal German, a C program ported to Rust — and renders it into a new form. The ideas aren’t theirs, but the structural choices are, and translators hold valid copyright over those choices. The analogy to AI seems to go: the AI was trained on source material, and now it renders that material into new outputs.
The problem is that translation requires a source text. Translation is always of something. The translator has a document in front of them; their job is to carry meaning from one form to another while preserving as much of the original as possible.
Training is not that. When a large language model trains on text, it doesn’t store those documents in a filing cabinet somewhere inside the model. What remains after training is not the source text — it’s a set of statistical weights that encode something like patterns of relationship between words and ideas. As best I understand it, the training data is gone from the model after training is complete. You can’t open up a language model and pull out the novel it read. You can ask it questions that show it learned from that novel, the way I can show you I learned from a book I read years ago — but that is not translation. The source text is not in there to be translated.
Model Three: Copying
The copying model is closest to what happens in certain legal arguments about AI training data — that the system ingested copyrighted work and is now reproducing it. This is a real concern in some narrow cases, and the legal battles around it are ongoing. But as a general description of how AI works, it’s still wrong.
When I was in art school, copying was a deliberate exercise. You took a master’s painting and you copied it — same composition, same palette, same approach to light. You did this over and over. At first you were just trying to match it. Over time, the copy became yours in a different way; your hand was in it, your particular misreadings and emphases, the specific distance between what the master did and what you did with the same intention. Even an honest copy carries the copyist’s signature.
AI output isn’t a copy in that sense because there’s no single original being reproduced. A language model generating a paragraph about loss hasn’t retrieved a specific paragraph from its training data and reproduced it. It has generated text that statistically coheres with what the context demands, drawing on patterns absorbed from an enormous range of human writing. The result may rhyme with something that exists somewhere, but it wasn’t copied from a definite source. Graphic design, which routinely appropriates fonts designed by others, photographs taken by others, and color systems developed by others, involves more direct copying than most AI outputs do — and we’ve mostly made peace with the idea that graphic design can still be original work. (I know that particular argument annoyed a few people in my typography class, for what it’s worth.)
The Third Category: An Interactive Mirror
So if it’s not channeling, not translation, not copying — what is it?
The best word I’ve found is mirroring. Specifically, a hall of mirrors.
There’s a line of chatbot history worth knowing here, going back much further than ChatGPT, that I’ll trace in a later post. For now, notice this: the app Replika got its name right. The early chatbots were explicitly designed to reflect the user’s input back — to take what you gave them and shape a response around it. That’s not a coincidence of design; it’s the core of what these systems do. Modern large language models are more sophisticated — much more — but the mirror dynamic is still the operative one.
Here’s what I mean by that. When you interact with a language model, the context of the conversation builds from what you’ve brought to it. The model has been trained on enormous amounts of human text, which is why its outputs sound human. But in any given session, the direction the output takes is shaped primarily by you — your question, your framing, your tone, the specifics of what you’ve described. The model reflects that back, filtered and refracted through its training. Two people working with the same system on the same general topic will get something recognizably different, because the inputs are different.
That’s the mirror.
The hall part matters too. A flat mirror gives you back what you put in, relatively faithfully. A hall of mirrors distorts. The reflections are bent, repeated, layered on each other, and sometimes they show you things you didn’t know were in the room. Language models do this because they carry strong pattern preferences — certain linguistic structures, certain ways of organizing an argument, certain reflexive moves that show up over and over because they showed up over and over in the training data. If you walk into a hall of mirrors unprepared and mistake a reflection for the exit, you’ll stay lost. If you understand what you’re working with, you can work it.
