RACE: Fine-Grained AI Text Detection That Distinguishes Human-Written, LLM-Polished, and Humanized AI Content

An ACL 2026 accepted paper introduces RACE (Rhetorical Analysis for Creator-Editor Modeling), a method that goes beyond binary "human vs AI" classification to detect fine-grained types of LLM involvement in text. This matters for policy enforcement, where LLM-polished human text and humanized LLM text carry different regulatory implications.

The Problem

Existing AI text detectors only do binary (human vs AI) or ternary classification. This is insufficient because:

• LLM-polished human text — Human writes, AI improves style. Common in professional writing.
• Humanized LLM text — AI generates, human edits. Common in content farms.
• Both trigger different policy consequences but existing detectors can't distinguish them.

The Four Classes

1. Pure human text — Written entirely by humans
2. LLM-generated text — Written entirely by LLMs
3. LLM-polished human text — Human foundation, AI style editing
4. Humanized LLM text — AI generation, human editing

How RACE Works

• Creator modeling — Uses Rhetorical Structure Theory to build a logic graph representing the original writer's foundation
• Editor modeling — Extracts Elementary Discourse Unit-level features to identify editing style
• Outperforms 12 baselines with low false alarm rates

Why It Matters

As AI writing tools become ubiquitous, regulation needs precision. A student using Grammarly-style polishing shouldn't be treated the same as someone submitting fully AI-generated content. RACE provides the granularity needed for fair enforcement.