Factuality refers to an AI model's ability to generate content that is accurate and consistent with verified, real-world knowledge.
The concept is closely related to hallucination but is much broader. It includes hallucinations but also covers outdated information and gaps in domain knowledge.
Factuality is especially important in high-stakes domains like healthcare and law, where a confidently stated inaccuracy can have real consequences. A model that cites a nonexistent legal precedent or gives an outdated drug interaction is potentially harmful.
Several factors contribute to factual errors in LLMs.
- Knowledge is frozen at the model's training cutoff, so anything that changes after that point won't be reflected in its outputs.
- Specialist domains that weren't well-represented in training data can leave significant gaps.
- Even when a model has the relevant knowledge, it can fail to connect the right pieces when reasoning across multiple steps.
Common approaches to improving factuality include retrieval-augmented generation (RAG), which grounds the model's responses in external sources at inference time, and domain-specific fine-tuning for applications where accuracy in a particular field is vital.