lynote-ai/ai-detector-skill

This repository is a tool for estimating whether a piece of text was likely written by an AI. It is designed to be cautious and transparent rather than making confident claims, because the problem of detecting AI-written text is genuinely hard and overconfident tools can do more harm than good. The tool works by measuring what it calls "AI-like signals": patterns in writing that tend to appear more often in AI-generated text than in human writing. Each signal is weighted and combined into a score from 0 to 100, along with a verdict label such as "high_ai_likelihood" or "mixed_or_uncertain." If the text is too short for a meaningful estimate, the tool says so instead of guessing. The output also includes a confidence level, the strongest signals that influenced the score, warnings attached to the result, and suggested follow-up steps. No hidden model is involved, the signals are visible and the scoring logic is inspectable. The tool can be used from the command line or called from Python code. It is also packaged as a skill that can be loaded by coding agents such as those that read a SKILL.md file at the root of a repository. The README gives specific use case guidance: a teacher reviewing a student submission, an editor checking for formulaic guest posts, or a moderation team routing suspicious content to a human review queue. It explicitly lists cases where the tool should not be used, including disciplinary decisions, fraud determinations, and very short samples. The README includes results from a reproducible evaluation against a public dataset of human and AI answers across finance, medicine, and general question-answering topics. The tool showed meaningful separation between human and AI scores on average but performed only weakly as a classifier at the tested thresholds. The authors describe it as better suited for triage and explanation than for standalone automated judgment. The project runs on Python 3.9 or newer, includes a test suite and benchmark scripts, and is licensed under MIT.