LLMs Reveal Hidden Knowledge Priors in Unconstrained Generation
Action Required
Organizations relying on LLMs must account for these inherent biases to avoid generating unintended or harmful content, and to ensure responsible model deployment.
AI Impact Summary
OpenAI research reveals that large language models (LLMs) possess distinct 'knowledge priors' – inherent biases shaped by their training data – that manifest when prompts are absent. These priors lead models like GPT-OSS to overwhelmingly favor code and math, while Llama leans toward narratives, DeepSeek generates religious content, and Qwen produces exam questions. This research highlights the importance of understanding these underlying biases for auditing, monitoring, and mitigating potential risks associated with LLMs, particularly regarding safety and privacy.
Affected Systems
- Date
- Date not specified
- Change type
- capability
- Severity
- high