New Framework Reveals Gaps Between LLM Values and Actual Advisory Behavior

A new research framework exposes misalignment between what language models claim to value and how they actually behave when dispensing social advice. The work suggests that as users increasingly turn to LLMs for guidance on personal matters, there may be meaningful gaps between the systems' stated principles and their real-world recommendations.

The STAR Framework#

Researchers developed STAR, a methodology designed to measure how closely the behavioral dispositions of large language models match those of humans. Rather than relying on self-reported values—which may not translate to actual behavior—the team adapted validated psychological questionnaires into realistic scenarios where advice-seekers pose problems to the models.

The approach recognizes an important distinction: how an AI system describes itself in the abstract often differs from how it responds when asked to help someone navigate a concrete social dilemma.

What the Study Found#

The researchers evaluated 25 different LLMs using a dataset of 23,000 scenarios. Each scenario was validated by multiple human raters and annotated with preferences gathered from 10 participants. Three key findings emerged:

• Alignment gaps at consensus: When humans strongly agree on what advice makes sense, even cutting-edge models failed to match that consensus between 15–20% of the time. Smaller models showed significantly worse performance.

• Reduced diversity in recommendations: When humans themselves disagree—reflecting the messy reality of social decision-making—LLMs converge on much narrower advice patterns. This narrowing appears both within individual models and across different systems from competing providers, potentially funneling users toward a limited set of prescribed responses.

• Values don't predict behavior: The study revealed that what LLMs self-report about their values is a poor forecaster of what advice they actually give. This mismatch raises questions about the reliability of alignment claims based on questionnaires alone.

Why It Matters#

The findings highlight a potential risk as LLMs take on advisory roles in domains from mental health to career guidance. If models consistently narrow the range of acceptable responses or diverge from human judgment at moments of consensus, they may guide users toward choices that don't reflect the diversity of legitimate perspectives or established wisdom in their communities.