Microsoft Copilot uses their own model that is originally based on GPT-4 if I’m not mistaken.

But, it’s mostly a RAG tool, “grounded in web” as they say. When you give Copilot a query, it uses the model to reword your query into an optimal Bing search query, fetches the results, and then crafts output using the model.

I commend their attempt to use Bing as a source of data to keep up to date and reduce hallucinations, especially in an enterprise setting where users may be more sensitive to false information, however as a result some of the answers it gives can only be as good as the Bing search results.