Scholarlink

<p>The capabilities of Large Language Models (LLMs) to synthesize texts that imitate human language have increased rapidly. While many people adopt this technology, the potential harm caused through texts synthesized by machines is not fully assessed. Factual errors due to model hallucinations are especially impactful in media like news articles, which serve an important function in society. Therefore, users require support in detecting LLM-generation to decrease risks posed by machine text synthesis. For this purpose, we propose the tool unCover based on explainable linguistic analysis.&#xa0;The tool analyzes texts through stylometric writing style analysis for grammatical information and topic modeling for semantic information. Its stylometry is based on character, word and syntactic trigrams. By inspecting the proposed techniques, the differences in LLM-generated texts can be uncovered. These findings are used to explain how text synthesis can be detected. unCover’s result is presented as a classification and visualization of the analysis.&#xa0;The final classification achieved an accuracy of 77.56 This is comparable to state-of-the-art products that detect LLMs while remaining technically explainable. German news articles are classified with 66.45 The visualization supports the decision of the tool and can help users navigate complex texts. These promising results show that unCover addresses challenges posed by AI content with new solutions. This is a step towards safely integrating LLMs into various areas of society.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Detecting and Comparing LLM Capabilities to Human Writers Through Linguistic Analysis

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月1日

查看全文
  • Lucas Liebe,
  • Jannis Baum,
  • Tim Barz-Cech,
  • Willy Scheibel,
  • Jürgen Döllner

摘要

The capabilities of Large Language Models (LLMs) to synthesize texts that imitate human language have increased rapidly. While many people adopt this technology, the potential harm caused through texts synthesized by machines is not fully assessed. Factual errors due to model hallucinations are especially impactful in media like news articles, which serve an important function in society. Therefore, users require support in detecting LLM-generation to decrease risks posed by machine text synthesis. For this purpose, we propose the tool unCover based on explainable linguistic analysis. The tool analyzes texts through stylometric writing style analysis for grammatical information and topic modeling for semantic information. Its stylometry is based on character, word and syntactic trigrams. By inspecting the proposed techniques, the differences in LLM-generated texts can be uncovered. These findings are used to explain how text synthesis can be detected. unCover’s result is presented as a classification and visualization of the analysis. The final classification achieved an accuracy of 77.56 This is comparable to state-of-the-art products that detect LLMs while remaining technically explainable. German news articles are classified with 66.45 The visualization supports the decision of the tool and can help users navigate complex texts. These promising results show that unCover addresses challenges posed by AI content with new solutions. This is a step towards safely integrating LLMs into various areas of society.