AI Output Grounding Fact-Check

A legal tech company post-checks every AI-generated contract summary. The system extracts 8 verifiable claims per summary (statute references, filing deadlines, fee amounts) and verifies each against Scavio search results. Cost: $0.04/summary. The system caught 3 incorrect statute citations in the first week that would have caused compliance issues.

AI Output Grounding Fact-Check is relevant to Google. Scavio provides a unified API to access data from all of these platforms.

LLMs hallucinate. Grounding fact-checks catch hallucinations after generation by extracting verifiable claims from the output and searching for evidence. The process: 1) Parse LLM output for factual claims (names, dates, prices, statistics, events). 2) For each claim, construct a verification search query. 3) Compare search results against the claim. 4) Flag claims with no corroborating evidence or direct contradictions. This is distinct from RAG (which grounds generation with pre-retrieved context) because it operates post-generation as a verification step. Both approaches are complementary: RAG reduces hallucinations at generation time, and grounding fact-checks catch what RAG misses. The cost is proportional to the number of verifiable claims. A typical 500-word LLM output contains 5-10 verifiable claims. At Scavio's $0.005/query, fact-checking one output costs $0.025-0.05. For high-stakes outputs (financial advice, medical information, legal guidance), this cost is trivial compared to the liability of incorrect information. Tavily's extract endpoint can also support this workflow at $0.03/request on the Researcher tier. The key engineering challenge is claim extraction -- determining which statements are verifiable versus subjective opinions.