What Current AI Benchmarks Leave Unmeasured: Modality, Search, Citations, and Implications (for Safety Evaluations)

Standard benchmarks may not reflect real-world deployment conditions: The study found that enabling web search—a common feature in deployed systems—reduced accuracy by up to 8 percentage points on safety benchmarks, and in some cases completely reversed which interface (chat UI vs. API) performed better, suggesting that lab evaluations without search may give misleading safety assessments.

How you access a model matters significantly: ChatGPT's chat interface produced less accurate responses than the same model accessed via API on both benchmarks tested, indicating that safety claims based on one access method may not generalize to how users actually interact with the system.

LLM responses are inconsistent, even on repeated identical prompts: Up to 21% of prompts produced different responses across three repeated runs, meaning a single test run per prompt—standard practice in current benchmarks—may miss important reliability and safety issues that only appear intermittently.

Citation practices vary unpredictably across modalities: The two interfaces grounded their answers in different sources and citations, raising concerns about whether safety evaluations that don't measure citation accuracy are missing a key dimension of trustworthiness and potential for spreading misinformation.

Accuracy alone is an incomplete safety metric: The paper demonstrates that measuring only accuracy overlooks other critical dimensions like response consistency, abstention behavior (when models refuse to answer), and citation grounding—all of which matter for responsible deployment but go unmeasured in typical benchmarks.