Methodology
How our research is produced
We publish a research stream because the patterns we see across thousands of drafts contain useful information. This page explains exactly what kind of evidence that is, how we frame it, and where it does and does not apply.
What our research is
Our research stream is editorial observation, framed as such. It draws on the patterns our team encounters in the rewrites our platform processes, the editorial work we and our partners perform, and the structural analysis we conduct on samples of AI-generated and human-edited writing. We use it to surface recurring patterns, name them clearly, and offer practical guidance grounded in what we actually see — not in speculation about what models might do.
What our research is not
It is not peer-reviewed academic research. It does not involve controlled experiments, statistically validated samples, or formal hypothesis testing. We do not publish precision figures we cannot defend, and we avoid the format conventions (numbered findings, abstract, conclusion) that imply a level of rigor we are not claiming. When we describe a pattern as common or consistent, we mean it appears reliably across the work we see — not that it has been measured to a specific confidence level.
Privacy and data handling
We do not analyse, store, or surface the substance of any individual user's writing for research purposes. Patterns we describe are derived from anonymised structural signals (sentence length distribution, opening conventions, hedge density) aggregated across very large numbers of drafts, or from publicly available samples of AI-generated and human-edited text. No user-identifiable content is used in research outputs. See our privacy policy for full data-handling terms.
How we frame uncertainty
We use language that signals the strength of a claim. "We observe," "in our work," and "consistently appears" describe pattern recognition from editorial experience. "Often" and "tends to" describe trends with notable exceptions. We avoid bare assertions of universal fact unless they can be verified independently. Where a claim depends on a particular model, era, or use case, we say so.
Updates and corrections
AI tools and the writing they produce evolve quickly. Pieces in our research stream are reviewed on a rolling basis; observations that no longer hold are revised or retired. When we update a piece materially, the updated date changes and significant changes are noted. We welcome corrections from readers who have evidence that contradicts what we have published.
Have research-quality data we should consider?
We are interested in collaborating with academic researchers, editorial teams, and product organisations who study AI writing rigorously. Get in touch.