Are We Aligned? A Preliminary Investigation of the Alignment of Responsible AI Values between LLMs and Human Judgment

LLMs show inconsistency between stated values and actual behavior: While LLMs claim to prioritize responsible AI values like fairness, privacy, and transparency in direct questions (Tasks 1-3), they don't consistently apply these same values when making practical decisions about software requirements (Task 4), suggesting a "faithfulness gap" between what they say and what they do.

LLMs align more with AI experts than the general public: The study found that LLMs' value preferences more closely match those of AI practitioners than a representative sample of US citizens, raising concerns about whether AI systems being developed reflect broader societal values or primarily the perspectives of technical specialists.

Human oversight remains essential in software development: The research demonstrates practical risks of using LLMs for requirements engineering—the process of determining what software should do—without human review, since their value judgments don't reliably align with either expert or public perspectives.

Five core responsible AI values emerge as priorities: Across both LLMs and human groups, fairness, privacy, transparency, safety, and accountability were consistently identified as the most important responsible AI values, though their relative importance varied by context and audience.

Current LLMs lack systematic value alignment monitoring: The paper highlights the need for better tools and methods to benchmark, interpret, and continuously monitor whether AI systems actually uphold the responsible AI values they claim to support during real-world deployment.