Major companies, such as Microsoft and Google, use technology that already exists, including their own search engines, which is pretty good at identifying falsehood from truth, to check what large language models are checking.

The claim: major AI companies such as Microsoft and Google use existing technology, such as search engines, to check or ground what their LLMs produce, and that technology is fairly good at telling truth from falsehood. True: as of March 2025, there is documentation (product docs, technical papers, announcements) that Microsoft and/or Google use search-based grounding or retrieval to check or ground LLM outputs (e.g. Bing grounding in Copilot, Google Search grounding in Gemini), and there is no strong evidence that such grounding is ineffective. Mostly true: the grounding practice is documented, but evaluations show only modest gains in factual accuracy, or the evidence covers only one of the named companies. Mixed: grounding is used, but evidence on how well it separates truth from falsehood is substantially mixed. Mostly false: grounding exists only in limited or experimental form and is not used to check outputs as described, or evaluations show it largely fails to catch falsehoods. False: these companies do not use search or other existing technology to check LLM outputs.

Set before the research began.

1. E1 What information does Copilot use to answer my prompt? Microsoft Support · 15 September 2026

   “Connecting a response to real information like this is called grounding. Grounding is what lets Copilot give you answers anchored in your actual work and the real world, complete with citations you can check.”

2. E2 What information does Copilot use to answer my prompt? Microsoft Support · 15 September 2026

   “Grounding improves accuracy and relevance and gives you citations to verify, but you should still review the sources and confirm critical details before you share or act on them.”

3. E3 Hallucinating Detail In Simple Summaries: Why LLM "Grounding" Doesn't Work To Combat Hallucination GDELT Project blog · 28 September 2023

   “As the results below demonstrate, neither grounding nor setting temperature to 0.0, both widely-touted strategies that are commonly promoted by poorly-trained AI consultants as quick methods of eliminating hallucination, fail.”

The Microsoft documentation describes grounding Copilot responses in "real information" with citations [E1, E2], but it is dated 2026, does not say search engines are used, and no item covers Google. On effectiveness, E2 says grounding improves accuracy and E3 says it fails to eliminate hallucination, so the evidence does not settle the claim as of March 2025 under the criteria.