Image: GitHub
Why it mattersEvery tool that removes signs of AI writing works on word choice. The study says the signs are mostly in the structure, so most of that effort is aimed at the wrong level.
sepia is a portable Agent Skill for removing the signs that a machine wrote something. The repository was created on 28 August 2026 and had 1,197 stars three days later, under an MIT licence. It installs through the Skills CLI, which the project says supports more than 77 agents, with native plugin packaging for Claude Code, Codex, Grok Build and Antigravity.
The interesting part is not the tool. It is the finding it is built on.
The finding
The project cites StoryScope (Russell et al., 2026, arXiv:2604.03136), a study of 61,608 stories written by humans and by five leading language models.
Two numbers from it, as the repository reports them. A classifier using narrative-structure features alone detected AI-written fiction at 93.2 percent macro-F1. And when the surface style was edited away, detection fell from 95.5 percent to 93.9 percent.
Read that second pair again. Rewriting the sentences changed the score by 1.6 points.
The signs that the study says remain after that editing are structural: the narrator explaining the theme, a single plot line where every cause leads neatly to its effect, emotion shown only as physical sensation, no references to the real world, no sense of a reader, strictly linear time, and an ending resolved by the main character growing and accepting something.
What sepia does with it
Three passes, in order. Narrative architecture first, then discourse flow, then surface style. The claim is that fixing the last one first is why other tools do not work.
It also has a separate track for professional writing, with rules matched to the document type: release notes, pull request replies, postmortems, tickets, technical articles. The failures it names there are recognisable to anyone who has read a generated changelog. Filler that carries no information. Hedging where a judgment was needed. Leftover chatbot phrasing. Formatting that looks machine-made.
There is a 30-feature diagnosis rubric and per-model correction sets for Claude, GPT, Gemini, DeepSeek and Kimi. Four operations: write, review, refactor and recreate.
Why this belongs in a developer AI News
Because the same argument applies to code, and it is the more useful version of it.
Teams checking AI-written code mostly check the surface: formatting, naming, a lint pass. The study's point, applied to code, is that the surface is the layer easiest to fix and least diagnostic. What reveals generated work is structure, and structure is what a surface check cannot see.
One caveat. The 93.2 percent figure is the study's, reported here through the repository's summary of it. The paper is linked above and it is short enough to read yourself before quoting the number to anyone.
Source
- Nanako0129/sepia on GitHub, MIT licence. Star and date figures read from the GitHub API on 31 August 2026
- StoryScope, arXiv:2604.03136, Russell et al., 2026
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.