Scholarlink

<p>In engineering education, Generative Artificial Intelligence (GenAI) might be misused to complete assessments with limited understanding. On courses that allow the use of GenAI, students might also forget to acknowledge its assistance. There is a need to identify such assistance. We present an automated detector with over-uniqueness and sample matching. GenAI-assisted submissions are identified based on their uniqueness and their similarity to a GenAI sample. Unique to our GenAI detector, it requires no training data and/or dedicated rules for each programming/scripting language. Further, the method can be integrated into any existing similarity detectors to identify plagiarism. The detector covers five similarity measurements, two similarity modes, and eight programming/scripting languages. Our evaluation of four data sets with thousands of submissions shows that our detector is effective (71% MAP). However, many factors can affect its effectiveness, including submission length and student attempts to align the code. Combining both mechanisms does not result in higher effectiveness, yet it takes longer to process.</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 GenAI assistance in programming assessments with over-uniqueness and sample matching

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月5日

查看全文
  • Oscar Karnalim,
  • Hapnes Toba,
  • Maresha Caroline Wijanto,
  • Yehezkiel David Setiawan,
  • Nico Surantha,
  • Michael Liut

摘要

In engineering education, Generative Artificial Intelligence (GenAI) might be misused to complete assessments with limited understanding. On courses that allow the use of GenAI, students might also forget to acknowledge its assistance. There is a need to identify such assistance. We present an automated detector with over-uniqueness and sample matching. GenAI-assisted submissions are identified based on their uniqueness and their similarity to a GenAI sample. Unique to our GenAI detector, it requires no training data and/or dedicated rules for each programming/scripting language. Further, the method can be integrated into any existing similarity detectors to identify plagiarism. The detector covers five similarity measurements, two similarity modes, and eight programming/scripting languages. Our evaluation of four data sets with thousands of submissions shows that our detector is effective (71% MAP). However, many factors can affect its effectiveness, including submission length and student attempts to align the code. Combining both mechanisms does not result in higher effectiveness, yet it takes longer to process.