A progressive verification method for false information detection using LLMs in human-AI hybrid information environments

Purpose The rapid advancement of large language models (LLMs) has substantially lowered the cost of generating false information while increasing its persuasiveness. Consequently, AI-generated false information is becoming increasingly prevalent online, posing significant challenges to conventional detection approaches. This study aims to investigate how the technological capabilities of LLMs can be leveraged to enhance false information detection in human-AI hybrid information environments, which is critical for contemporary network information governance. Design/methodology/approach A progressive verification method for false information detection using LLMs is proposed. The method consists of a four-stage detection pipeline executed by the LLM: (1) independently reasoning to form an initial judgment of a text's veracity; (2) extracting of key 5W1H elements; (3) matching the text against false information features with external knowledge support; and (4) deriving a final decision by integrating the results from the preceding stages. The study also compares the detection performance of LLMs across human-generated and AI-generated false information, and examines feature differences between human-generated and AI-generated content. Findings Experimental results indicate that the progressive verification method significantly enhances detection performance across three widely used LLMs: DeepSeek-V3.2, GPT-5.4-mini, and Qwen3-Plus. Among them, DeepSeek-V3.2 integrated with our method achieves superior overall detection performance compared with all other baseline approaches, exhibiting remarkable advantages in detecting both human-generated and AI-generated false information. Further analysis reveals that AI-generated false information is often more deceptive, relying on ambiguous sources or enriched background details, which complicates detection. Originality/value This study contributes to the literature by developing a progressive verification detection method that systematically improves LLMs' performance in detecting false information. By guiding LLMs to form an initial independent judgment and progressively verify it through structured information extraction and feature matching with external knowledge support, the approach provides reliable methodological support for improving false information detection in human-AI hybrid information environments.