The Voice Was Never the Ceiling: Constraint Allocation and Machine Style Across Six Frontier Models

Engineering notes

Engineering notes will be added by the aipentium editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为大语言模型、生成式AI、ChatGPT相关技术、计算机视觉、深度学习等高价值论文补充中文说明。

Original abstract

"AI writing sounds like AI" is one of the most repeated complaints about large language models, and it is usually offered as a statement about capability that the machines write in that flat, hedged, faintly corporate register because that is the best they can do. This paper argues the opposite. We gave six frontier models (ChatGPT, Claude, Copilot, DeepSeek, Gemini, and Grok) a single fixed source text and asked each to rewrite it in more than a dozen human voices, from Shakespeare to Snoop Dogg, from a Pentecostal sermon to a used-car commercial. The generic voice did not survive contact with a real constraint. Every model, pushed into a specific register, abandoned the default and produced something with genuine stylistic identity — and the ways they did it were legible, distinct, and characteristic of each model. We read the results through three lenses. The Ferrari Test asks whether a model, handed a metaphor that does not belong in the target voice, reasons about whether to keep it or replace it — and reveals that the interesting signal is not swapping but register sensitivity, keeping the metaphor where it fits and replacing it where it does not. Form versus Voice separates getting the words right from getting the shape right, and finds models that nail one while missing the other. And four allocation policies — substitution, embodiment, performance, and source-text archaeology — describe four fundamentally different routes to the same assignment. The through-line is a single claim: the generic AI voice is not a ceiling the models hit. It is a resource they spend elsewhere by default, and when the prompt makes them spend it on voice, the voice appears.

Links and sources

Comments