Scholarlink

<p>Generative Artificial Intelligence (GenAI) systems, such as large language models (LLMs), are increasingly permeating everyday tasks; yet, they sometimes present false or misleading information as fact – a phenomenon known as hallucination. As GenAI becomes more accessible to the public, it is essential to understand how users’ AI literacy shapes their reliance on imperfect advice from AI systems. Building on the concept of correspondence bias, the study examines how individuals with different levels of AI literacy respond to faulty GenAI advice. Evidence from an online programming experiment with 542 U.S. programmers shows that individuals with higher AI literacy rely less on GenAI advice, particularly when the advice is flawed. Correspondence bias provides a plausible explanatory mechanism for these findings and helps reconcile mixed results in prior research on AI literacy. Overall, the findings offer a more nuanced perspective on the benefits and risks of AI-literacy-driven mistrust. This informs education, integration, and evaluation initiatives for GenAI while cautioning against naive evaluation strategies.</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

Beneficial Mistrust in Generative AI? The Role of AI Literacy in Programmers’ Handling of Bad Coding Advice

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月12日

查看全文
  • Dirk Leffrang,
  • Nina Passlack,
  • Oliver Müller,
  • Oliver Posegga

摘要

Generative Artificial Intelligence (GenAI) systems, such as large language models (LLMs), are increasingly permeating everyday tasks; yet, they sometimes present false or misleading information as fact – a phenomenon known as hallucination. As GenAI becomes more accessible to the public, it is essential to understand how users’ AI literacy shapes their reliance on imperfect advice from AI systems. Building on the concept of correspondence bias, the study examines how individuals with different levels of AI literacy respond to faulty GenAI advice. Evidence from an online programming experiment with 542 U.S. programmers shows that individuals with higher AI literacy rely less on GenAI advice, particularly when the advice is flawed. Correspondence bias provides a plausible explanatory mechanism for these findings and helps reconcile mixed results in prior research on AI literacy. Overall, the findings offer a more nuanced perspective on the benefits and risks of AI-literacy-driven mistrust. This informs education, integration, and evaluation initiatives for GenAI while cautioning against naive evaluation strategies.