The Shape of Original: building a toolkit to read machine writing

Here's the claim in The Shape of Original: we think we're detecting a boundary. We're actually building one. A measurement instrument draws the line it pretends to merely find, and once that line gets published, it quietly turns into a rule everyone has to follow. The whole thing lives as a self-contained interactive document, embedded so the experiments stay sandboxed. And the thread holding it together? Originality is not novelty, not in law and not in statistics. Statistical rarity only means anything relative to a corpus. Legal originality depends on authorship history. Fold one into the other and you've committed the exact mistake the field commits over and over.

The essay comes out of an orchestration layer running a dependency graph of prompts. Seed the spine. Fire seven analytic lenses in parallel, fuse them, draft section by section, assemble one HTML file, then push it through fact-check and editorial gates before anything ships. Those gates earned their keep. On the very first run, the fact-checker caught the draft inventing model version numbers that flat-out contradicted the source paper, and the editor flagged that the build had quietly truncated itself. Real bugs, both of them. Caught before a human read a word.

Four pocket tools take the claims and make them something you can actually poke at, all of it running in the browser. The Narrative Scorer drops a story onto the human-to-AI axis using the paper's thirty core features, each weighted by its published rank. The AI-Tell Linter flags the surface patterns of machine prose. The Tell Stripper writes the automatable ones back out. And the Rarity Visualizer? You drop a point into a cloud and watch its originality drift as the surrounding corpus fills in. Rarity lives in the pool, not the point.

Finally, there's Human or AI?, which sticks you on the other side of the instrument. You read a passage, you call it, and then the game shows you the tells you missed. Here's where it gets sneaky: the human passages are all public-domain classics, and a good chunk of them trip every detector out there. Dickens opens with a tricolon on stilts. Melville uses an em dash in his second sentence. Turns out you learn this faster playing a game than arguing about it. The signals are statistical and nothing more. They're never proof, and any classifier that forgets that will start failing the people who actually write.

Here's what actually bugs me. It's not that machines have a style of their own. It's that once we can measure how they differ from us, we can't help enforcing it, and the measurement sneaks in as the rule everybody starts writing toward. Draw that line yourself, or somebody draws it for you and you're living inside whatever it optimized. Try it. Read the essay, play the game, drop your own paragraph into the linter. See how quickly a good sentence gets flagged as fake.