The Shape of Original: building a toolkit to read machine writing
A research paper found that AI fiction has a measurable fingerprint. So I built the whole loop around it: an interactive essay, an orchestration pipeline, four pocket utilities, and a guessing game. Here is the making-of.
StoryScope (Russell et al., 2026) mirrored ten thousand human stories with five language models and found something uncomfortable: machine fiction has a measurable shape. It over-determines its themes, ties every thread into a neat bow, and reaches for the same handful of words. I wanted to do more than read the paper, so I built the whole loop around it: an interactive essay, the pipeline that generates it, four small utilities, and a game. This is the making-of.
The essay: an argument you can poke
The piece, The Shape of Original, argues that the real story is not detection but construction: a measurement instrument quietly builds the boundary it claims to merely find, and once you publish that boundary it becomes a prescription. It is a self-contained interactive document, embedded so its experiments stay sandboxed. The thread running through it is that originality, legally and statistically, is not novelty. Statistical rarity is corpus-relative; legal originality turns on authorship history. Conflating them is the category error the whole field keeps making.
The toolkit: a twenty-stage pipeline with gates
The essay is generated by an orchestration layer that runs a dependency graph of prompts: seed the spine, run seven analytic lenses in parallel, fuse them, draft section by section, build one HTML file, then run fact-check and editorial gates before anything ships. Those gates earned their keep. On the first run the fact-checker caught the draft inventing model version numbers that contradicted the source paper, and the editor flagged that the build had been silently truncated. Both were real bugs, surfaced before a human read a word.
A fingerprint and a forgery guide are the same document.
The utilities: ideas you can hold
Four pocket tools turn the claims into things you can operate, all running entirely in the browser. A Narrative Scorer places a story on the human-to-AI axis using the paper's thirty core features, weighted by their published rank. An AI-Tell Linter highlights the surface patterns of machine prose. A Tell Stripper rewrites the automatable ones back out. And a Rarity Visualizer lets you drop a point in a cloud and watch its originality drift as the surrounding corpus fills in, because rarity is a property of the pool, not the point.
The game: tells are not proof
Finally, Human or AI? puts you on the other side of the instrument. Read a passage, call it, then see the tells you missed. The trick is in the deck: the human passages are public-domain classics, and some of them trip every detector. Dickens opens with a tricolon on stilts; Melville uses an em dash in his second sentence. The point lands faster in a game than in an argument. The signals are statistical, never proof, and a classifier that forgets that starts failing real writers.
What I would tell you
The uncomfortable part is not that machines have a style. It is that the moment we can measure the difference, we are tempted to enforce it, and the measurement quietly becomes the rule everyone writes toward. Build the boundary deliberately, or it gets built accidentally and you inherit whatever it happened to optimize. Read the essay, play the game, run your own paragraph through the linter, and watch how easily a clean sentence gets flagged.