Pastiche on Bitsy Wiki

Generative Exaggeration

Mon, 01 Jan 0001 00:00:00 +0000

Asked to write as someone, a language model finds the features that most distinguish that someone and turns them up. Everything else about the person flattens out. Nudo et al. named this generative exaggeration — “a systematic amplification of salient traits beyond empirical baselines” — and measured it moving the wrong way with better information: a fuller picture of the person produces a more stereotyped portrait, not a less stereotyped one.

What Transfers and What Does Not

Mon, 01 Jan 0001 00:00:00 +0000

A style is not one thing, and the parts of it come across a pastiche prompt at very different rates. The measurable summary is that the perceptible layer — how long the sentences run, how long the words are, what the punctuation does — moves in the right direction, and the layer underneath it does not move at all. George Mikros put numbers on both halves by having GPT-4o imitate Ernest Hemingway and Mary Shelley and then running the output through four independent stylometric feature sets.

Reversion to House Style

Mon, 01 Jan 0001 00:00:00 +0000

A prompt asking for an author’s voice pins down some of the output and leaves the rest open. The open part does not drift toward some other author, and it does not drift randomly. It goes to the register the model writes in when nobody asks it to write in anyone’s — the same default that LLM overused words documents from the vocabulary side. Reversion is that register winning the parts of the page the instruction never reached.

Authorship Survives Imitation

Mon, 01 Jan 0001 00:00:00 +0000

An imitation can satisfy a reader and still be separated from the author’s real writing by a statistical test that takes a second to run. The reason is structural rather than a matter of the imitation being poor: the features a reader judges style by and the features that identify an author are almost disjoint sets, and a prompt can only address the first.

The features that identify are not the features that are noticed#

Stylometry established the split before computers were involved in writing: attribution runs on the rates of high-frequency function words, because those rates are stable within a writer and independent of what the writing is about. The Federalist Papers settle the comparison — Hamilton and Madison differ by four hundredths of a word in average sentence length and are separated outright by their function-word rates.

Rules Versus Examples

Mon, 01 Jan 0001 00:00:00 +0000

There are two ways to tell a model what a style is. Show it prose and let it infer, or describe the prose and let it follow the description. Both are bounded, and they are bounded differently: examples carry everything about the sample including the parts you did not mean, and descriptions carry only what someone managed to notice. The measurements favour supplying both, and then constraining the description so the model cannot invent its own criteria for what the style is — with the caveat this page keeps returning to, that they were taken on code generation and on style transfer rather than on pastiche.

Law and Ethics

Mon, 01 Jan 0001 00:00:00 +0000

Copyright protects expression and not ideas, and style has always been filed on the idea side. That is the settled part, it long predates generative models, and it is also routinely overstated: courts treat stylistic similarity as evidence bearing on whether protected expression was copied, so “style is not copyrightable” is the beginning of the analysis rather than the end of it. The live questions are elsewhere — in the right of publicity, in what training on a corpus requires, and in a market-harm argument that has been formulated but not yet tested.