AI Content Detection Systems Mislabel Human Work in China
AI content detection systems deployed by Chinese schools and online platforms are systematically mislabeling original human work as AI-generated, according to reports from students and journalists. The systems reject well-written papers and authentic human videos due to overly sensitive algorithms, forcing users to degrade the quality of their work or pay for remediation services to pass detection.
What happened
Multiple AIGC detection platforms used by Chinese schools and students have consistently mislabeled original human-written academic papers as AI-generated content. A reporter tested the system by submitting a pre-AI-era article and received an AI suspicion rating of 18.12%, with sections marked as 60-70% AI-generated. Students report results varying drastically across platforms, with free websites showing 30-40% AIGC rates for papers that register 1-3% on school systems using the same content.
Video verification systems on content platforms rejected authentic human-filmed videos as not being real-person submissions. A woman attempted to upload her own real video to create a digital avatar but received repeated "failed real-person verification" rejections despite the video containing only her genuine footage with no AI synthesis or face-swapping.
Reported impact
Students have paid hundreds of yuan for platform services to artificially lower AIGC detection scores on their original work. Some students distort their writing quality by adding filler words, restructuring sentences, and removing professional language to pass detection, resulting in lower-quality academic work.
The mislabeling has created a situation where people must prove they are human to AI systems rather than the reverse, with particular disadvantage to those with high-quality, well-structured original work or aesthetically refined video content.
Index analysis
The detection systems are based on statistical analysis of features rather than actual understanding of content. AI training on human-generated examples makes AI-written content highly stable, logically consistent, and free of grammatical errors, causing algorithms to classify well-written human work as AI-generated due to its lack of the "confusion" typical of human writing.
Detection algorithms lack transparency. Platforms provide no explanation of their detection methodology, creating a situation where users cannot understand why their content was flagged or how to appeal. The systems operate as a 'black box' unlike plagiarism detection, which specifies matching sources.
Governance controls
Schools have set AIGC rate thresholds for acceptable thesis submissions, typically capping at 30%, but these controls depend on unreliable detection systems with no standardized methodology across platforms. Detection platforms offer paid services to lower AIGC scores, creating a perverse incentive where users must pay to bypass flawed detection rather than rely on legitimate detection improvements.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.