Trustworthiness in Large Language Models - Key Factors Influencing Software Company Adoption and Industry Barriers

Sammanfattning

As large language models (LLMs) become increasinglyintegrated into software engineering workflows, concernsaround their trustworthiness pose significant challenges tobroader organizational adoption. While LLMs offer substantialgains in productivity and innovation, issues such as bias, lackof transparency, security vulnerabilities, and unpredictabilityhinder their reliable implementation. Existing ethical guidelines,such as the EU’s principles for trustworthy AI, provide a foundationalframework but often fall short when it comes to operationalization,offering high-level ideals without clear, actionablesteps for integrating those principles into daily engineering practicesor system design choices. This study investigates how companiesapproach the adoption of trustworthy LLMs by combining asurvey of literature with interviews of industry practitioners. Ourfindings highlight three central organizational considerations:transparency, accountability, and human oversight. We find thattransparency must be tailored to the risk level of specific tasks, accountabilityrelies on clearly defined internal roles and processes,and human oversight remains essential across all deploymentcontexts. Furthermore, the study identifies four key barriersto implementation: technical limitations (e.g., hallucinations andprompt injection), organizational gaps in governance structures,social and ethical concerns such as bias and over reliance, andregulatory ambiguity. In response, organizations adopt layeredstrategies that integrate human-centered oversight, model-levelsafeguards, and governance practices adapted to specific usecases. By bridging the gap between abstract ethical principlesand operational needs, this study offers a practical framework tosupport responsible and trustworthy LLM adoption in softwareengineering.