What is grounding in AI?
Published . Updated .
Grounding: Tying a language model's answer to sources it retrieved and can point to, such as search results, so the answer rests on more than what the model learned in training.
Google's documentation names search results and retrieval from your own data as grounding sources, and gives fewer hallucinations as the reason to do it. Retrieval-augmented generation is the general method.
A web search in ChatGPT, Claude or Grok is grounding at work, and the citations are the visible part. AI citations covers how the cited pages get chosen.
Read next
- What is retrieval-augmented generation (RAG)?
- How do AI citations work in ChatGPT, Claude and Grok?
- What is an AI hallucination?
Sources
- Grounding overview, Google Cloud documentation
- Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (Lewis et al., arXiv, 2020)
See the fan-out for your own prompts
Only Fan Out is a free Chrome side panel that lists the queries, fetched pages and cited pages behind each ChatGPT, Claude or Grok answer. Nothing leaves your browser.
Add to Chrome, it's free