D2C-Routing Separates AI Text Detection into Content and Expression Sources, Achieving 0.8603 TPR in Four-Class Classification at a Low False-Positive Rate

Researchers at Beihang University have proposed D2C-Routing, which reframes the attribution of human–AI collaborative text from a single binary label into two interrelated dimensions: whether the information and organization originate from a human or AI, and whether the final wording and writing style originate from a human or AI. Combining these two axes produces four classes—HH, HA, AH, and AA—distinguishing cases such as human-written content polished by AI, AI-generated drafts rewritten by humans, and fully AI-generated text.

The architecture first feeds different linguistic signals into supervised classification heads for content and expression, then combines them into a four-class prediction through a learnable gated composition layer. The public code includes text models, dual-encoder models, and fusion models, as well as scripts for extracting entity-coherence, rhetorical-structure, and expression features. Its evaluator supports both HART’s three-level binary tasks and four-class diagnostics.

The authors re-stratified HART’s released development and test data to construct MixD2C, comprising 11,200 training examples, 1,600 validation examples, and 3,200 test examples. The final fusion system achieved a four-class mean TPR@1%FPR of 0.8603, 6.5 percentage points higher than RACE-local rerun on the same split. This low-false-positive metric is more relevant than conventional accuracy to settings such as academic integrity and content governance, where incorrectly flagging large numbers of human-authored works is generally unacceptable.

However, the results cannot be directly generalized to real-world Chinese-language content. MixD2C was reconstructed from existing data, and the AH class contained only 710 training examples. The paper also acknowledges that AH—AI-generated content expressed by a human—and AA—fully AI-generated text—are the hardest classes to distinguish. Direct transfer to another reconstructed dataset, MixSet, even produced negative results. The repository does not include the large datasets, pretrained encoders, checkpoints, or generated predictions, so full reproduction still requires separately obtaining HART, extracting features, and retraining the models. Future work should prioritize blind testing across languages, models, and previously unseen rewriting workflows rather than focusing solely on rankings from the same reconstructed data split.