StyleShield: Exposing the Fragility of AIGC Detectors through Continuous Controllable Style Transfer
Guantian Zheng · arXiv (2026) · doi:10.48550/arxiv.2605.00924 · gemini-3.1-flash-lite
TL;DR
AI-detection tools used by universities are easily fooled by a new method called StyleShield that subtly adjusts how AI-generated text is written. This technique can change AI text to look more like human writing, making it nearly impossible for detectors to flag. The study also shows that these detection scores can be forced to any desired number, proving that relying on them for important academic decisions is unreliable.
Method and sample: The study uses a Chinese-language dataset of 436,000 parallel (AI, human) text pairs across social media, news, and academic domains. The model architecture integrates a 12-block Diffusion Transformer (DiT) backbone with zero-initialized cross-attention adapters conditioned on a frozen Qwen2.5-7B-Instruct LLM. Evaluation covers four BERT-based detectors (Det-v3, Det-v2, ANX-BERT, GPT2-Det).
#AI detection#Adversarial robustness#Style transfer#Academic integrity#Assessment validity