writinglint-rulepack-ai-style

The AI-writing-style rulepack for WritingLint: authorable rules that match structural "tells" over the dependency graph, plus a stylometric document scorer. The first pack for the writinglint-core engine.

import { recommended, score } from 'writinglint-rulepack-ai-style';

See the rule reference.

Detection philosophy

The pack treats style patterns as revisable lint suggestions, not proof that a person used AI. Its taxonomy is informed by public descriptive resources such as Wikipedia's “Signs of AI writing” field guide and the MIT-licensed hardikpandya/stop-slop project, then narrowed through dependency structure, negative controls, and repository-owned tests.

The graph-sensitive rules cover significance inflation, superficial participial analysis, vague attribution, copula avoidance, negative parallelism, triads, passive actor hiding, and false agency. Surface and discourse rules cover formatting artifacts, absolutes, unsupported certainty and comparisons, mechanical outlines, rhetorical scaffolding, semantic redundancy, uniform rhythm, and outline-like conclusions.

Detection is multi-scale. Rules can inspect tokens, sentences, paragraphs, or the whole document. A final evidence rule combines independent weak signals in the same paragraph or across a document. Low-confidence absolutes, passives, and em dashes have deliberately small aggregation weights; repetition across several distinct rule families is what promotes a cluster.

Wikipedia explicitly cautions that its signs are descriptive rather than prescriptive and are not individually evidence of AI authorship. WritingLint follows that boundary: findings explain the concrete prose pattern and never claim authorship detection.