New Benchmark Shows Context Matters More Than Previously Thought in AI Safety Testing

A new research benchmark reveals that testing large language models for safety has been missing a critical dimension: context. A team of researchers introduces CASE-Bench (Context-Aware SafEty Benchmark), which factors in the circumstances surrounding a user query when assessing whether a model responds appropriately.

The Problem with Current Approaches#

Safety benchmarks for large language models typically focus on whether an AI will refuse to answer potentially harmful requests. However, this narrow framing creates a blind spot. Many current tests overlook that the same question can be entirely appropriate in one setting yet problematic in another. This leads to an unintended consequence: AI systems often refuse legitimate requests that happen to contain words or topics flagged as risky, frustrating users even when no actual danger exists.

How CASE-Bench Works#

Rather than treating queries in isolation, CASE-Bench assigns formal contexts to questions based on Contextual Integrity theory—a framework that examines how information flows appropriately within different social relationships and situations. The researchers recruited a statistically sufficient pool of human annotators (not just a handful, as many prior studies do) to evaluate how context shapes safety judgments.

What the Research Found#

The results were striking. Context had a statistically significant and substantial impact on how humans judge whether LLM responses are safe (p<0.0001 from a z-test). The analysis also uncovered notable gaps between human expectations and actual model behavior, especially among commercial LLMs in contexts that should be considered safe. These mismatches suggest that today's most widely deployed language models may be either over-refusing legitimate requests or failing to recognize genuinely problematic ones depending on the surrounding context.

Why It Matters#

The work underscores a fundamental oversight in how AI safety is currently evaluated and developed. As language models become more integrated into professional, medical, educational, and other sensitive domains, understanding whether refusals make sense in their actual use contexts becomes essential for both user trust and model reliability.