Grounding
Connecting AI model outputs to verified external information sources to improve accuracy and reduce hallucination.
Definition
Grounding anchors AI responses in real data rather than model training alone. It typically involves connecting models to search engines, databases, or document repositories during generation.
Grounded systems can cite sources, reducing hallucination while enabling verification. They represent a hybrid of generative AI and traditional information retrieval.
Why It Matters
Grounding addresses AI's reliability problems for factual queries. Businesses deploying AI for customer service or research require grounded systems for acceptable accuracy.
For users, grounded AI enables fact-checking that pure generative models cannot provide.
Examples in Practice
Perplexity AI grounds responses in web searches, citing sources for each claim. Enterprise AI systems ground in internal knowledge bases for company-specific accuracy.
Microsoft's Copilot grounds responses in current web information, addressing training data currency limitations.