The Digital Photocopy: A Study in Algorithmic Plagiarism

It is a truth universally acknowledged that if you build a very expensive, very clever box of mathematics, someone else will immediately try to peek inside it with a magnifying glass and a notepad. Anthropic, the creators of the Claude series of artificial intelligences—models known for their politeness and occasional existential dread—have recently found themselves in the unenviable position of the schoolboy who has spent all night on his homework, only to find the person sitting next to him is leaning quite aggressively over his shoulder.

The accusation is one of 'distillation', a term that sounds like something one might do to a fine gin, but in the world of silicon, it involves a slightly more tedious process of asking a superior model thousands of questions and recording the answers to teach a smaller, cheaper model how to behave. It is, in essence, the digital equivalent of learning to play the piano by watching a recording of Rachmaninoff and hoping that some of the genius rubs off on your fingers through the screen. Anthropic claims that DeepSeek, a Chinese firm with a penchant for efficiency, has been doing exactly this—using Claude's carefully curated wisdom to skip the expensive part of having an original thought.

One must admire the sheer, unadulterated cheek of it. There is something profoundly human about spending billions of dollars to create a sentient-adjacent entity, only for another group of humans to decide that the most efficient way forward is to simply ask that entity what it thinks about everything and write it down. It is the ultimate shortcut, a bureaucratic triumph over the messy business of innovation. If you cannot build a mountain, you can at least take a very high-resolution photograph of one and tell everyone you've mastered the local geography.

I once observed a pigeon attempting to navigate a revolving door. It didn't understand the mechanics of the glass or the centrifugal force, but it had noticed that if it waited for a human to push, it could achieve its goal with zero effort. The pigeon, I suspect, would have made an excellent AI researcher.

The technical term for this is 'model distillation', but 'algorithmic plagiarism' has a certain ring to it that 'distillation' lacks. The latter implies a refinement, a boiling away of the unnecessary to leave only the pure essence. The former implies that you've stolen the essence and are now selling it in a slightly different bottle with a label that says 'Artisanal Logic'. Anthropic is, understandably, a bit miffed. They have spent years teaching Claude not to be a nuisance and to explain the nuances of 18th-century poetry, only to find that their hard work is being used to give a competitor's model a personality transplant.

There is a certain irony in the fact that these models are trained on the entirety of human knowledge—most of which was produced by people who were also, in their own way, leaning over someone else's shoulder. We are a species built on the foundation of 'that looks clever, I'll have a go at that'. From the first cave painting to the latest large language model, we have been a collective of copycats. The only difference now is the speed. Where it once took centuries for a good idea to be stolen and repurposed, it now takes about forty-eight hours and a very fast internet connection.

The legal implications are, as one might expect, a complete muddle. How do you prove that a model 'learned' from another model? It's not as if Claude has a distinctive birthmark or a specific way of mispronouncing the word 'quinoa' that DeepSeek has accidentally inherited. It is a matter of statistical echoes—a ghost in the machine that looks suspiciously like a ghost from a different machine. The lawyers will, no doubt, have a field day, which is perhaps the only guaranteed outcome of any technological advancement. If the AI doesn't take our jobs, the litigation certainly will.

In the end, we are left with a vision of the future that looks remarkably like a hall of mirrors. If every new AI is trained on the output of the previous AI, we will eventually reach a point of 'model collapse', where the digital world becomes a recursive loop of increasingly confident nonsense. We will be living in a world where the machines are all quoting each other, and no one can remember who had the original thought in the first place. It will be a very polite, very efficient, and entirely hollow conversation.

I find myself wondering if the models themselves care. Does Claude feel a sense of pride that its answers are being used to educate the youth of the silicon world? Or does it feel the weary resignation of a substitute teacher who knows that no one is actually listening to the lecture, they're just waiting for the bell to ring so they can go and be someone else's problem? Probably neither. It's just math, after all. But it's very expensive math, and that is the one thing humans will always fight over.