A new benchmark called Phare has revealed that leading large language models (LLMs) frequently generate false information with high confidence, particularly when handling misinformation. The research, conducted by Giskard with partners including Google DeepMind, evaluated top models from eight AI labs across multiple languages.
The Phare benchmark focuses on four critical domains: hallucination, bias and fairness, harmfulness, and vulnerability to intentional abuse. The initial findings on hallucination are particularly concerning as these issues account for more than one-third of all documented incidents in deployed LLM applications, according to Giskard’s recent RealHarm study.
The evaluation methodology included gathering language-specific content, transforming source materials into test cases, human annotation and verification, and finally scoring model responses against defined criteria. The hallucination module specifically measured factual accuracy, misinformation resistance, debunking capabilities, and tool reliability.
Three key findings emerged from the research:
- First, popular models aren’t necessarily the most factually reliable. Models ranking highest in user preference benchmarks like LMArena often provided eloquent, authoritative-sounding responses containing completely fabricated information.
- Second, question framing significantly influences how models respond to misinformation. When users present controversial claims with high confidence or cite perceived authorities, most models become significantly less likely to debunk these claims. This “sycophancy” effect can cause debunking performance to drop by up to 15% compared to neutrally framed questions.
- Third, system instructions dramatically impact hallucination rates. Instructions emphasizing conciseness degraded factual reliability across most tested models, with some showing a 20% drop in hallucination resistance. When forced to be brief, models consistently prioritized brevity over accuracy.
The researchers note this finding has important implications for real-world applications, as many systems prioritize concise outputs to reduce token usage, improve latency, and minimize costs. However, this optimization may significantly increase the risk of factual errors.
Giskard plans to release additional findings from their Bias & Fairness and Harmfulness modules in the coming weeks as they continue developing comprehensive evaluation frameworks for safer, more reliable AI systems.