Writing voice replication - qualitative benchmark

Benchmark range

Lower bound: 0. Median value: 0. Upper bound: 0.

Measurement period: per voice profile built and calibrated.

Methodology

The research supports three claims, none of which is a productivity figure for this package. First, that individual writing style is measurable: authorship attribution research shows function-word adjacency networks capture author-specific fingerprints and outperform word-frequency methods, and ACL work shows purely syntactic features improve cross-genre attribution while proper nouns carry topic rather than style - which is why the skill measures structure and treats topic vocabulary with suspicion. Second, that the problem is real: the Binoculars method detected over 90% of LLM outputs at a 0.01% false-positive rate, so default model prose is statistically distinguishable from human writing, and both Doshi and Hauser in Science Advances and a 36-participant arXiv study found AI assistance raises individual quality while compressing distinctiveness across a set. Third, that style match has commercial value: Journal of Service Management research found matching language style to the audience increased brand trust through perceived interaction effort, benevolence and interaction quality. CaseLedger publishes no numeric ROI for this listing because the honest answer is that voice match is judged by the author, which is exactly why the package ships a scoring rubric rather than a promised percentage.

Cited sources: 6. Measurement period: per voice profile built and calibrated.

Original sources

• The impact of language style accommodation during social media interactions on brand trust

  Journal of Service Management

  Matching a brand's language style to customers in social media interactions increased brand trust through higher perceived interaction effort, benevolence, and interaction quality.

• Generative AI enhances individual creativity but reduces the collective diversity of novel content

  Science Advances

  Doshi and Hauser found access to generative AI ideas made individual stories better rated and more enjoyable, especially for less creative writers, but made the set of stories more similar to one another.

• Comparing creativity support tools and ChatGPT for idea generation

  arXiv

  In a 36-participant study, ChatGPT users produced less semantically distinct ideas than users of an alternative creativity-support tool, even as they generated more ideas in greater detail.

• Syntactic features in cross-genre authorship attribution

  Association for Computational Linguistics

  Purely syntactic models improved cross-genre authorship attribution, while proper nouns were shown to carry topic interference rather than stable author style.

• Authorship attribution using function word adjacency networks

  arXiv

  Function-word adjacency networks captured author-specific stylometric fingerprints and outperformed authorship methods based on word frequencies alone.

• Spotting LLMs with Binoculars - zero-shot detection of machine-generated text

  Proceedings of Machine Learning Research

  The Binoculars method detected over 90% of ChatGPT and other LLM outputs at a 0.01% false-positive rate without any training data, indicating default model prose is statistically distinguishable from human writing.