DSIPA: Training-Free Framework Detects LLM-Generated Text via Sentiment Analysis
Researchers have proposed DSIPA, a novel training-free framework for detecting LLM-generated text by analyzing sentiment distributional stability under controlled stylistic variation. The framework addresses security challenges posed by machine-generated content used for misinformation, impersonation, and content forgery. Unlike existing detection methods that struggle with adversarial perturbation, paraphrasing attacks, and domain shifts, DSIPA operates in a zero-shot, black-box manner without requiring access to model parameters or large labeled datasets. It leverages two unsupervised metrics—sentiment distribution consistency and sentiment distribution preservation—based on the observation that LLMs produce emotionally consistent outputs while human-written texts exhibit greater affective variation. The paper is available on arXiv under identifier 2604.26328.
Key facts
- DSIPA is a training-free framework for detecting LLM-generated text.
- It uses sentiment distributional stability under stylistic variation.
- The framework is zero-shot and black-box.
- It employs two unsupervised metrics: sentiment distribution consistency and sentiment distribution preservation.
- LLMs produce emotionally consistent outputs; humans show greater affective variation.
- Existing detection methods are vulnerable to adversarial perturbation and paraphrasing attacks.
- The paper is published on arXiv with ID 2604.26328.
- The approach addresses misinformation, impersonation, and content forgery.
Entities
Institutions
- arXiv