Mapping Trustworthiness in Large Language Models: A Bibliometric Analysis Bridging Theory to Practice

Trustworthiness remains poorly defined: The research identified 18 different definitions of trust/trustworthiness across the literature, with no clear consensus on what the term actually means in practice. This fragmentation creates confusion about what companies and developers are actually trying to achieve when they claim their LLMs are "trustworthy."

Three dimensions dominate the conversation: Transparency, explainability, and reliability emerged as the most commonly cited aspects of trustworthiness across the 2,006 publications analyzed. However, the paper suggests these definitions remain largely theoretical without consistent practical implementation standards.

Most solutions are controlled by developers, not users: The analysis identified 20 strategies to enhance trustworthiness, with fine-tuning and retrieval-augmented generation (RAG) being most prominent. Critically, most of these strategies are "developer-driven and applied during the post-training phase," meaning users and regulators have limited ability to verify or enforce trustworthiness claims.

"Ethics washing" is a real risk: The paper warns that fragmented terminologies and lack of unified frameworks enable "ethics washing"—where companies adopt ethical language without genuine regulatory commitment or accountability. This allows organizations to appear trustworthy without substantive changes.

Theory and practice remain disconnected: The research highlights "persistent gaps between theoretical taxonomies and practical implementation," suggesting that academic discussions about trustworthiness often don't translate into meaningful safeguards in deployed systems.